diff --git "a/openai_whisper-large-v3-v20240930/AudioEncoder.mlmodelc/model.mil" "b/openai_whisper-large-v3-v20240930/AudioEncoder.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/openai_whisper-large-v3-v20240930/AudioEncoder.mlmodelc/model.mil" @@ -0,0 +1,2693 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}, {"coremltools-component-torch", "2.4.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.0"}})] +{ + func main(tensor melspectrogram_features) { + tensor var_106_pad_type_0 = const()[name = tensor("op_106_pad_type_0"), val = tensor("custom")]; + tensor var_106_pad_0 = const()[name = tensor("op_106_pad_0"), val = tensor([0, 0, 1, 1])]; + tensor var_106_strides_0 = const()[name = tensor("op_106_strides_0"), val = tensor([1, 1])]; + tensor var_106_dilations_0 = const()[name = tensor("op_106_dilations_0"), val = tensor([1, 1])]; + tensor var_106_groups_0 = const()[name = tensor("op_106_groups_0"), val = tensor(1)]; + tensor var_81_to_fp16 = const()[name = tensor("op_81_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor var_87_to_fp16 = const()[name = tensor("op_87_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(983168)))]; + tensor var_106_cast_fp16 = conv(bias = var_87_to_fp16, dilations = var_106_dilations_0, groups = var_106_groups_0, pad = var_106_pad_0, pad_type = var_106_pad_type_0, strides = var_106_strides_0, weight = var_81_to_fp16, x = melspectrogram_features)[name = tensor("op_106_cast_fp16")]; + tensor hidden_states_1_mode_0 = const()[name = tensor("hidden_states_1_mode_0"), val = tensor("EXACT")]; + tensor hidden_states_1_cast_fp16 = gelu(mode = hidden_states_1_mode_0, x = var_106_cast_fp16)[name = tensor("hidden_states_1_cast_fp16")]; + tensor var_146_pad_type_0 = const()[name = tensor("op_146_pad_type_0"), val = tensor("custom")]; + tensor var_146_pad_0 = const()[name = tensor("op_146_pad_0"), val = tensor([0, 0, 1, 1])]; + tensor var_146_strides_0 = const()[name = tensor("op_146_strides_0"), val = tensor([2, 2])]; + tensor var_146_dilations_0 = const()[name = tensor("op_146_dilations_0"), val = tensor([1, 1])]; + tensor var_146_groups_0 = const()[name = tensor("op_146_groups_0"), val = tensor(1)]; + tensor var_121_to_fp16 = const()[name = tensor("op_121_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(985792)))]; + tensor var_127_to_fp16 = const()[name = tensor("op_127_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10816256)))]; + tensor var_146_cast_fp16 = conv(bias = var_127_to_fp16, dilations = var_146_dilations_0, groups = var_146_groups_0, pad = var_146_pad_0, pad_type = var_146_pad_type_0, strides = var_146_strides_0, weight = var_121_to_fp16, x = hidden_states_1_cast_fp16)[name = tensor("op_146_cast_fp16")]; + tensor hidden_states_3_mode_0 = const()[name = tensor("hidden_states_3_mode_0"), val = tensor("EXACT")]; + tensor hidden_states_3_cast_fp16 = gelu(mode = hidden_states_3_mode_0, x = var_146_cast_fp16)[name = tensor("hidden_states_3_cast_fp16")]; + tensor var_164_to_fp16 = const()[name = tensor("op_164_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10818880)))]; + tensor inputs_1_cast_fp16 = add(x = hidden_states_3_cast_fp16, y = var_164_to_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor var_178 = const()[name = tensor("op_178"), val = tensor(3)]; + tensor out_1_axes_0 = const()[name = tensor("out_1_axes_0"), val = tensor([1])]; + tensor var_197_to_fp16 = const()[name = tensor("op_197_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_197_to_fp16, x = inputs_1_cast_fp16)[name = tensor("out_1_cast_fp16")]; + tensor obj_1_mean_0_to_fp16 = const()[name = tensor("obj_1_mean_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14658944)))]; + tensor obj_1_variance_0_to_fp16 = const()[name = tensor("obj_1_variance_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14661568)))]; + tensor obj_1_gamma_0_to_fp16 = const()[name = tensor("obj_1_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14664192)))]; + tensor obj_1_beta_0_to_fp16 = const()[name = tensor("obj_1_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14666816)))]; + tensor obj_1_epsilon_0_to_fp16 = const()[name = tensor("obj_1_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_1_cast_fp16 = batch_norm(beta = obj_1_beta_0_to_fp16, epsilon = obj_1_epsilon_0_to_fp16, gamma = obj_1_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor("obj_1_cast_fp16")]; + tensor query_1_pad_type_0 = const()[name = tensor("query_1_pad_type_0"), val = tensor("valid")]; + tensor query_1_strides_0 = const()[name = tensor("query_1_strides_0"), val = tensor([1, 1])]; + tensor query_1_pad_0 = const()[name = tensor("query_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_1_dilations_0 = const()[name = tensor("query_1_dilations_0"), val = tensor([1, 1])]; + tensor query_1_groups_0 = const()[name = tensor("query_1_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14669440)))]; + tensor layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17946304)))]; + tensor query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = query_1_dilations_0, groups = query_1_groups_0, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = query_1_strides_0, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("query_1_cast_fp16")]; + tensor key_1_pad_type_0 = const()[name = tensor("key_1_pad_type_0"), val = tensor("valid")]; + tensor key_1_strides_0 = const()[name = tensor("key_1_strides_0"), val = tensor([1, 1])]; + tensor key_1_pad_0 = const()[name = tensor("key_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_1_dilations_0 = const()[name = tensor("key_1_dilations_0"), val = tensor([1, 1])]; + tensor key_1_groups_0 = const()[name = tensor("key_1_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17948928)))]; + tensor key_1_cast_fp16 = conv(dilations = key_1_dilations_0, groups = key_1_groups_0, pad = key_1_pad_0, pad_type = key_1_pad_type_0, strides = key_1_strides_0, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("key_1_cast_fp16")]; + tensor value_1_pad_type_0 = const()[name = tensor("value_1_pad_type_0"), val = tensor("valid")]; + tensor value_1_strides_0 = const()[name = tensor("value_1_strides_0"), val = tensor([1, 1])]; + tensor value_1_pad_0 = const()[name = tensor("value_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_1_dilations_0 = const()[name = tensor("value_1_dilations_0"), val = tensor([1, 1])]; + tensor value_1_groups_0 = const()[name = tensor("value_1_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21225792)))]; + tensor layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24502656)))]; + tensor value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = value_1_dilations_0, groups = value_1_groups_0, pad = value_1_pad_0, pad_type = value_1_pad_type_0, strides = value_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("value_1_cast_fp16")]; + tensor var_232 = const()[name = tensor("op_232"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_1_cast_fp16 = reshape(shape = var_232, x = query_1_cast_fp16)[name = tensor("mh_q_1_cast_fp16")]; + tensor var_234_to_fp16 = const()[name = tensor("op_234_to_fp16"), val = tensor(0x1p-3)]; + tensor var_235_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_234_to_fp16)[name = tensor("op_235_cast_fp16")]; + tensor var_236 = const()[name = tensor("op_236"), val = tensor([1, 20, 64, -1])]; + tensor var_237_cast_fp16 = reshape(shape = var_236, x = key_1_cast_fp16)[name = tensor("op_237_cast_fp16")]; + tensor mh_w_1_transpose_x_0 = const()[name = tensor("mh_w_1_transpose_x_0"), val = tensor(true)]; + tensor mh_w_1_transpose_y_0 = const()[name = tensor("mh_w_1_transpose_y_0"), val = tensor(false)]; + tensor mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_235_cast_fp16, y = var_237_cast_fp16)[name = tensor("mh_w_1_cast_fp16")]; + tensor var_240_cast_fp16 = softmax(axis = var_178, x = mh_w_1_cast_fp16)[name = tensor("op_240_cast_fp16")]; + tensor var_241 = const()[name = tensor("op_241"), val = tensor([1, 20, 64, -1])]; + tensor var_242_cast_fp16 = reshape(shape = var_241, x = value_1_cast_fp16)[name = tensor("op_242_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(true)]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_242_cast_fp16, y = var_240_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_245 = const()[name = tensor("op_245"), val = tensor([1, 1280, 1, -1])]; + tensor input_1_cast_fp16 = reshape(shape = var_245, x = attn_1_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor obj_3_pad_type_0 = const()[name = tensor("obj_3_pad_type_0"), val = tensor("valid")]; + tensor obj_3_strides_0 = const()[name = tensor("obj_3_strides_0"), val = tensor([1, 1])]; + tensor obj_3_pad_0 = const()[name = tensor("obj_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_3_dilations_0 = const()[name = tensor("obj_3_dilations_0"), val = tensor([1, 1])]; + tensor obj_3_groups_0 = const()[name = tensor("obj_3_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24505280)))]; + tensor layers_0_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27782144)))]; + tensor obj_3_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = obj_3_dilations_0, groups = obj_3_groups_0, pad = obj_3_pad_0, pad_type = obj_3_pad_type_0, strides = obj_3_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("obj_3_cast_fp16")]; + tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_3_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([1])]; + tensor var_263_to_fp16 = const()[name = tensor("op_263_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_263_to_fp16, x = inputs_3_cast_fp16)[name = tensor("out_3_cast_fp16")]; + tensor input_3_gamma_0_to_fp16 = const()[name = tensor("input_3_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27784768)))]; + tensor input_3_beta_0_to_fp16 = const()[name = tensor("input_3_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27787392)))]; + tensor input_3_epsilon_0_to_fp16 = const()[name = tensor("input_3_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_3_cast_fp16 = batch_norm(beta = input_3_beta_0_to_fp16, epsilon = input_3_epsilon_0_to_fp16, gamma = input_3_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor input_5_pad_type_0 = const()[name = tensor("input_5_pad_type_0"), val = tensor("valid")]; + tensor input_5_strides_0 = const()[name = tensor("input_5_strides_0"), val = tensor([1, 1])]; + tensor input_5_pad_0 = const()[name = tensor("input_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_5_dilations_0 = const()[name = tensor("input_5_dilations_0"), val = tensor([1, 1])]; + tensor input_5_groups_0 = const()[name = tensor("input_5_groups_0"), val = tensor(1)]; + tensor layers_0_fc1_weight_to_fp16 = const()[name = tensor("layers_0_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790016)))]; + tensor layers_0_fc1_bias_to_fp16 = const()[name = tensor("layers_0_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40897280)))]; + tensor input_5_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = layers_0_fc1_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor input_7_mode_0 = const()[name = tensor("input_7_mode_0"), val = tensor("EXACT")]; + tensor input_7_cast_fp16 = gelu(mode = input_7_mode_0, x = input_5_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor hidden_states_5_pad_type_0 = const()[name = tensor("hidden_states_5_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_5_strides_0 = const()[name = tensor("hidden_states_5_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_5_pad_0 = const()[name = tensor("hidden_states_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_5_dilations_0 = const()[name = tensor("hidden_states_5_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_5_groups_0 = const()[name = tensor("hidden_states_5_groups_0"), val = tensor(1)]; + tensor layers_0_fc2_weight_to_fp16 = const()[name = tensor("layers_0_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40907584)))]; + tensor layers_0_fc2_bias_to_fp16 = const()[name = tensor("layers_0_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54014848)))]; + tensor hidden_states_5_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = hidden_states_5_dilations_0, groups = hidden_states_5_groups_0, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = hidden_states_5_strides_0, weight = layers_0_fc2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("hidden_states_5_cast_fp16")]; + tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; + tensor var_296 = const()[name = tensor("op_296"), val = tensor(3)]; + tensor out_5_axes_0 = const()[name = tensor("out_5_axes_0"), val = tensor([1])]; + tensor var_315_to_fp16 = const()[name = tensor("op_315_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_315_to_fp16, x = inputs_5_cast_fp16)[name = tensor("out_5_cast_fp16")]; + tensor obj_5_gamma_0_to_fp16 = const()[name = tensor("obj_5_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54017472)))]; + tensor obj_5_beta_0_to_fp16 = const()[name = tensor("obj_5_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54020096)))]; + tensor obj_5_epsilon_0_to_fp16 = const()[name = tensor("obj_5_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_5_cast_fp16 = batch_norm(beta = obj_5_beta_0_to_fp16, epsilon = obj_5_epsilon_0_to_fp16, gamma = obj_5_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor("obj_5_cast_fp16")]; + tensor query_3_pad_type_0 = const()[name = tensor("query_3_pad_type_0"), val = tensor("valid")]; + tensor query_3_strides_0 = const()[name = tensor("query_3_strides_0"), val = tensor([1, 1])]; + tensor query_3_pad_0 = const()[name = tensor("query_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_3_dilations_0 = const()[name = tensor("query_3_dilations_0"), val = tensor([1, 1])]; + tensor query_3_groups_0 = const()[name = tensor("query_3_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54022720)))]; + tensor layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57299584)))]; + tensor query_3_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = query_3_dilations_0, groups = query_3_groups_0, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = query_3_strides_0, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = tensor("query_3_cast_fp16")]; + tensor key_3_pad_type_0 = const()[name = tensor("key_3_pad_type_0"), val = tensor("valid")]; + tensor key_3_strides_0 = const()[name = tensor("key_3_strides_0"), val = tensor([1, 1])]; + tensor key_3_pad_0 = const()[name = tensor("key_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_3_dilations_0 = const()[name = tensor("key_3_dilations_0"), val = tensor([1, 1])]; + tensor key_3_groups_0 = const()[name = tensor("key_3_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57302208)))]; + tensor key_3_cast_fp16 = conv(dilations = key_3_dilations_0, groups = key_3_groups_0, pad = key_3_pad_0, pad_type = key_3_pad_type_0, strides = key_3_strides_0, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = tensor("key_3_cast_fp16")]; + tensor value_3_pad_type_0 = const()[name = tensor("value_3_pad_type_0"), val = tensor("valid")]; + tensor value_3_strides_0 = const()[name = tensor("value_3_strides_0"), val = tensor([1, 1])]; + tensor value_3_pad_0 = const()[name = tensor("value_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_3_dilations_0 = const()[name = tensor("value_3_dilations_0"), val = tensor([1, 1])]; + tensor value_3_groups_0 = const()[name = tensor("value_3_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60579072)))]; + tensor layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63855936)))]; + tensor value_3_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = value_3_dilations_0, groups = value_3_groups_0, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = value_3_strides_0, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = tensor("value_3_cast_fp16")]; + tensor var_350 = const()[name = tensor("op_350"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_3_cast_fp16 = reshape(shape = var_350, x = query_3_cast_fp16)[name = tensor("mh_q_3_cast_fp16")]; + tensor var_352_to_fp16 = const()[name = tensor("op_352_to_fp16"), val = tensor(0x1p-3)]; + tensor var_353_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_352_to_fp16)[name = tensor("op_353_cast_fp16")]; + tensor var_354 = const()[name = tensor("op_354"), val = tensor([1, 20, 64, -1])]; + tensor var_355_cast_fp16 = reshape(shape = var_354, x = key_3_cast_fp16)[name = tensor("op_355_cast_fp16")]; + tensor mh_w_3_transpose_x_0 = const()[name = tensor("mh_w_3_transpose_x_0"), val = tensor(true)]; + tensor mh_w_3_transpose_y_0 = const()[name = tensor("mh_w_3_transpose_y_0"), val = tensor(false)]; + tensor mh_w_3_cast_fp16 = matmul(transpose_x = mh_w_3_transpose_x_0, transpose_y = mh_w_3_transpose_y_0, x = var_353_cast_fp16, y = var_355_cast_fp16)[name = tensor("mh_w_3_cast_fp16")]; + tensor var_358_cast_fp16 = softmax(axis = var_296, x = mh_w_3_cast_fp16)[name = tensor("op_358_cast_fp16")]; + tensor var_359 = const()[name = tensor("op_359"), val = tensor([1, 20, 64, -1])]; + tensor var_360_cast_fp16 = reshape(shape = var_359, x = value_3_cast_fp16)[name = tensor("op_360_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(true)]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_360_cast_fp16, y = var_358_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_363 = const()[name = tensor("op_363"), val = tensor([1, 1280, 1, -1])]; + tensor input_9_cast_fp16 = reshape(shape = var_363, x = attn_3_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor obj_7_pad_type_0 = const()[name = tensor("obj_7_pad_type_0"), val = tensor("valid")]; + tensor obj_7_strides_0 = const()[name = tensor("obj_7_strides_0"), val = tensor([1, 1])]; + tensor obj_7_pad_0 = const()[name = tensor("obj_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_7_dilations_0 = const()[name = tensor("obj_7_dilations_0"), val = tensor([1, 1])]; + tensor obj_7_groups_0 = const()[name = tensor("obj_7_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63858560)))]; + tensor layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67135424)))]; + tensor obj_7_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = obj_7_dilations_0, groups = obj_7_groups_0, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = obj_7_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("obj_7_cast_fp16")]; + tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = obj_7_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; + tensor out_7_axes_0 = const()[name = tensor("out_7_axes_0"), val = tensor([1])]; + tensor var_381_to_fp16 = const()[name = tensor("op_381_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_381_to_fp16, x = inputs_7_cast_fp16)[name = tensor("out_7_cast_fp16")]; + tensor input_11_gamma_0_to_fp16 = const()[name = tensor("input_11_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67138048)))]; + tensor input_11_beta_0_to_fp16 = const()[name = tensor("input_11_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67140672)))]; + tensor input_11_epsilon_0_to_fp16 = const()[name = tensor("input_11_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_11_cast_fp16 = batch_norm(beta = input_11_beta_0_to_fp16, epsilon = input_11_epsilon_0_to_fp16, gamma = input_11_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor input_13_pad_type_0 = const()[name = tensor("input_13_pad_type_0"), val = tensor("valid")]; + tensor input_13_strides_0 = const()[name = tensor("input_13_strides_0"), val = tensor([1, 1])]; + tensor input_13_pad_0 = const()[name = tensor("input_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_13_dilations_0 = const()[name = tensor("input_13_dilations_0"), val = tensor([1, 1])]; + tensor input_13_groups_0 = const()[name = tensor("input_13_groups_0"), val = tensor(1)]; + tensor layers_1_fc1_weight_to_fp16 = const()[name = tensor("layers_1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67143296)))]; + tensor layers_1_fc1_bias_to_fp16 = const()[name = tensor("layers_1_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80250560)))]; + tensor input_13_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = layers_1_fc1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_mode_0 = const()[name = tensor("input_15_mode_0"), val = tensor("EXACT")]; + tensor input_15_cast_fp16 = gelu(mode = input_15_mode_0, x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor hidden_states_7_pad_type_0 = const()[name = tensor("hidden_states_7_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_7_strides_0 = const()[name = tensor("hidden_states_7_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_7_pad_0 = const()[name = tensor("hidden_states_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_7_dilations_0 = const()[name = tensor("hidden_states_7_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_7_groups_0 = const()[name = tensor("hidden_states_7_groups_0"), val = tensor(1)]; + tensor layers_1_fc2_weight_to_fp16 = const()[name = tensor("layers_1_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80260864)))]; + tensor layers_1_fc2_bias_to_fp16 = const()[name = tensor("layers_1_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93368128)))]; + tensor hidden_states_7_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_1_fc2_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("hidden_states_7_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; + tensor var_414 = const()[name = tensor("op_414"), val = tensor(3)]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([1])]; + tensor var_433_to_fp16 = const()[name = tensor("op_433_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_433_to_fp16, x = inputs_9_cast_fp16)[name = tensor("out_9_cast_fp16")]; + tensor obj_9_gamma_0_to_fp16 = const()[name = tensor("obj_9_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93370752)))]; + tensor obj_9_beta_0_to_fp16 = const()[name = tensor("obj_9_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93373376)))]; + tensor obj_9_epsilon_0_to_fp16 = const()[name = tensor("obj_9_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_9_cast_fp16 = batch_norm(beta = obj_9_beta_0_to_fp16, epsilon = obj_9_epsilon_0_to_fp16, gamma = obj_9_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor("obj_9_cast_fp16")]; + tensor query_5_pad_type_0 = const()[name = tensor("query_5_pad_type_0"), val = tensor("valid")]; + tensor query_5_strides_0 = const()[name = tensor("query_5_strides_0"), val = tensor([1, 1])]; + tensor query_5_pad_0 = const()[name = tensor("query_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_5_dilations_0 = const()[name = tensor("query_5_dilations_0"), val = tensor([1, 1])]; + tensor query_5_groups_0 = const()[name = tensor("query_5_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93376000)))]; + tensor layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96652864)))]; + tensor query_5_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_to_fp16, dilations = query_5_dilations_0, groups = query_5_groups_0, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = query_5_strides_0, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor("query_5_cast_fp16")]; + tensor key_5_pad_type_0 = const()[name = tensor("key_5_pad_type_0"), val = tensor("valid")]; + tensor key_5_strides_0 = const()[name = tensor("key_5_strides_0"), val = tensor([1, 1])]; + tensor key_5_pad_0 = const()[name = tensor("key_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_5_dilations_0 = const()[name = tensor("key_5_dilations_0"), val = tensor([1, 1])]; + tensor key_5_groups_0 = const()[name = tensor("key_5_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96655488)))]; + tensor key_5_cast_fp16 = conv(dilations = key_5_dilations_0, groups = key_5_groups_0, pad = key_5_pad_0, pad_type = key_5_pad_type_0, strides = key_5_strides_0, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor("key_5_cast_fp16")]; + tensor value_5_pad_type_0 = const()[name = tensor("value_5_pad_type_0"), val = tensor("valid")]; + tensor value_5_strides_0 = const()[name = tensor("value_5_strides_0"), val = tensor([1, 1])]; + tensor value_5_pad_0 = const()[name = tensor("value_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_5_dilations_0 = const()[name = tensor("value_5_dilations_0"), val = tensor([1, 1])]; + tensor value_5_groups_0 = const()[name = tensor("value_5_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99932352)))]; + tensor layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103209216)))]; + tensor value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_to_fp16, dilations = value_5_dilations_0, groups = value_5_groups_0, pad = value_5_pad_0, pad_type = value_5_pad_type_0, strides = value_5_strides_0, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor("value_5_cast_fp16")]; + tensor var_468 = const()[name = tensor("op_468"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_5_cast_fp16 = reshape(shape = var_468, x = query_5_cast_fp16)[name = tensor("mh_q_5_cast_fp16")]; + tensor var_470_to_fp16 = const()[name = tensor("op_470_to_fp16"), val = tensor(0x1p-3)]; + tensor var_471_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_470_to_fp16)[name = tensor("op_471_cast_fp16")]; + tensor var_472 = const()[name = tensor("op_472"), val = tensor([1, 20, 64, -1])]; + tensor var_473_cast_fp16 = reshape(shape = var_472, x = key_5_cast_fp16)[name = tensor("op_473_cast_fp16")]; + tensor mh_w_5_transpose_x_0 = const()[name = tensor("mh_w_5_transpose_x_0"), val = tensor(true)]; + tensor mh_w_5_transpose_y_0 = const()[name = tensor("mh_w_5_transpose_y_0"), val = tensor(false)]; + tensor mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_471_cast_fp16, y = var_473_cast_fp16)[name = tensor("mh_w_5_cast_fp16")]; + tensor var_476_cast_fp16 = softmax(axis = var_414, x = mh_w_5_cast_fp16)[name = tensor("op_476_cast_fp16")]; + tensor var_477 = const()[name = tensor("op_477"), val = tensor([1, 20, 64, -1])]; + tensor var_478_cast_fp16 = reshape(shape = var_477, x = value_5_cast_fp16)[name = tensor("op_478_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(true)]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_478_cast_fp16, y = var_476_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_481 = const()[name = tensor("op_481"), val = tensor([1, 1280, 1, -1])]; + tensor input_17_cast_fp16 = reshape(shape = var_481, x = attn_5_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor obj_11_pad_type_0 = const()[name = tensor("obj_11_pad_type_0"), val = tensor("valid")]; + tensor obj_11_strides_0 = const()[name = tensor("obj_11_strides_0"), val = tensor([1, 1])]; + tensor obj_11_pad_0 = const()[name = tensor("obj_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_11_dilations_0 = const()[name = tensor("obj_11_dilations_0"), val = tensor([1, 1])]; + tensor obj_11_groups_0 = const()[name = tensor("obj_11_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103211840)))]; + tensor layers_2_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106488704)))]; + tensor obj_11_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = obj_11_dilations_0, groups = obj_11_groups_0, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = obj_11_strides_0, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("obj_11_cast_fp16")]; + tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_11_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; + tensor out_11_axes_0 = const()[name = tensor("out_11_axes_0"), val = tensor([1])]; + tensor var_499_to_fp16 = const()[name = tensor("op_499_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_499_to_fp16, x = inputs_11_cast_fp16)[name = tensor("out_11_cast_fp16")]; + tensor input_19_gamma_0_to_fp16 = const()[name = tensor("input_19_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106491328)))]; + tensor input_19_beta_0_to_fp16 = const()[name = tensor("input_19_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106493952)))]; + tensor input_19_epsilon_0_to_fp16 = const()[name = tensor("input_19_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_19_cast_fp16 = batch_norm(beta = input_19_beta_0_to_fp16, epsilon = input_19_epsilon_0_to_fp16, gamma = input_19_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_pad_type_0 = const()[name = tensor("input_21_pad_type_0"), val = tensor("valid")]; + tensor input_21_strides_0 = const()[name = tensor("input_21_strides_0"), val = tensor([1, 1])]; + tensor input_21_pad_0 = const()[name = tensor("input_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_21_dilations_0 = const()[name = tensor("input_21_dilations_0"), val = tensor([1, 1])]; + tensor input_21_groups_0 = const()[name = tensor("input_21_groups_0"), val = tensor(1)]; + tensor layers_2_fc1_weight_to_fp16 = const()[name = tensor("layers_2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106496576)))]; + tensor layers_2_fc1_bias_to_fp16 = const()[name = tensor("layers_2_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119603840)))]; + tensor input_21_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = layers_2_fc1_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor input_23_mode_0 = const()[name = tensor("input_23_mode_0"), val = tensor("EXACT")]; + tensor input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = input_21_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor hidden_states_9_pad_type_0 = const()[name = tensor("hidden_states_9_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_9_strides_0 = const()[name = tensor("hidden_states_9_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_9_pad_0 = const()[name = tensor("hidden_states_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_9_dilations_0 = const()[name = tensor("hidden_states_9_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_9_groups_0 = const()[name = tensor("hidden_states_9_groups_0"), val = tensor(1)]; + tensor layers_2_fc2_weight_to_fp16 = const()[name = tensor("layers_2_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119614144)))]; + tensor layers_2_fc2_bias_to_fp16 = const()[name = tensor("layers_2_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132721408)))]; + tensor hidden_states_9_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = hidden_states_9_dilations_0, groups = hidden_states_9_groups_0, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = hidden_states_9_strides_0, weight = layers_2_fc2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("hidden_states_9_cast_fp16")]; + tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; + tensor var_532 = const()[name = tensor("op_532"), val = tensor(3)]; + tensor out_13_axes_0 = const()[name = tensor("out_13_axes_0"), val = tensor([1])]; + tensor var_551_to_fp16 = const()[name = tensor("op_551_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_551_to_fp16, x = inputs_13_cast_fp16)[name = tensor("out_13_cast_fp16")]; + tensor obj_13_gamma_0_to_fp16 = const()[name = tensor("obj_13_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132724032)))]; + tensor obj_13_beta_0_to_fp16 = const()[name = tensor("obj_13_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132726656)))]; + tensor obj_13_epsilon_0_to_fp16 = const()[name = tensor("obj_13_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_13_cast_fp16 = batch_norm(beta = obj_13_beta_0_to_fp16, epsilon = obj_13_epsilon_0_to_fp16, gamma = obj_13_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor("obj_13_cast_fp16")]; + tensor query_7_pad_type_0 = const()[name = tensor("query_7_pad_type_0"), val = tensor("valid")]; + tensor query_7_strides_0 = const()[name = tensor("query_7_strides_0"), val = tensor([1, 1])]; + tensor query_7_pad_0 = const()[name = tensor("query_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_7_dilations_0 = const()[name = tensor("query_7_dilations_0"), val = tensor([1, 1])]; + tensor query_7_groups_0 = const()[name = tensor("query_7_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132729280)))]; + tensor layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136006144)))]; + tensor query_7_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = query_7_dilations_0, groups = query_7_groups_0, pad = query_7_pad_0, pad_type = query_7_pad_type_0, strides = query_7_strides_0, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor("query_7_cast_fp16")]; + tensor key_7_pad_type_0 = const()[name = tensor("key_7_pad_type_0"), val = tensor("valid")]; + tensor key_7_strides_0 = const()[name = tensor("key_7_strides_0"), val = tensor([1, 1])]; + tensor key_7_pad_0 = const()[name = tensor("key_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_7_dilations_0 = const()[name = tensor("key_7_dilations_0"), val = tensor([1, 1])]; + tensor key_7_groups_0 = const()[name = tensor("key_7_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136008768)))]; + tensor key_7_cast_fp16 = conv(dilations = key_7_dilations_0, groups = key_7_groups_0, pad = key_7_pad_0, pad_type = key_7_pad_type_0, strides = key_7_strides_0, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor("key_7_cast_fp16")]; + tensor value_7_pad_type_0 = const()[name = tensor("value_7_pad_type_0"), val = tensor("valid")]; + tensor value_7_strides_0 = const()[name = tensor("value_7_strides_0"), val = tensor([1, 1])]; + tensor value_7_pad_0 = const()[name = tensor("value_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_7_dilations_0 = const()[name = tensor("value_7_dilations_0"), val = tensor([1, 1])]; + tensor value_7_groups_0 = const()[name = tensor("value_7_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139285632)))]; + tensor layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142562496)))]; + tensor value_7_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_to_fp16, dilations = value_7_dilations_0, groups = value_7_groups_0, pad = value_7_pad_0, pad_type = value_7_pad_type_0, strides = value_7_strides_0, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor("value_7_cast_fp16")]; + tensor var_586 = const()[name = tensor("op_586"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_7_cast_fp16 = reshape(shape = var_586, x = query_7_cast_fp16)[name = tensor("mh_q_7_cast_fp16")]; + tensor var_588_to_fp16 = const()[name = tensor("op_588_to_fp16"), val = tensor(0x1p-3)]; + tensor var_589_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_588_to_fp16)[name = tensor("op_589_cast_fp16")]; + tensor var_590 = const()[name = tensor("op_590"), val = tensor([1, 20, 64, -1])]; + tensor var_591_cast_fp16 = reshape(shape = var_590, x = key_7_cast_fp16)[name = tensor("op_591_cast_fp16")]; + tensor mh_w_7_transpose_x_0 = const()[name = tensor("mh_w_7_transpose_x_0"), val = tensor(true)]; + tensor mh_w_7_transpose_y_0 = const()[name = tensor("mh_w_7_transpose_y_0"), val = tensor(false)]; + tensor mh_w_7_cast_fp16 = matmul(transpose_x = mh_w_7_transpose_x_0, transpose_y = mh_w_7_transpose_y_0, x = var_589_cast_fp16, y = var_591_cast_fp16)[name = tensor("mh_w_7_cast_fp16")]; + tensor var_594_cast_fp16 = softmax(axis = var_532, x = mh_w_7_cast_fp16)[name = tensor("op_594_cast_fp16")]; + tensor var_595 = const()[name = tensor("op_595"), val = tensor([1, 20, 64, -1])]; + tensor var_596_cast_fp16 = reshape(shape = var_595, x = value_7_cast_fp16)[name = tensor("op_596_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(true)]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_596_cast_fp16, y = var_594_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_599 = const()[name = tensor("op_599"), val = tensor([1, 1280, 1, -1])]; + tensor input_25_cast_fp16 = reshape(shape = var_599, x = attn_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor obj_15_pad_type_0 = const()[name = tensor("obj_15_pad_type_0"), val = tensor("valid")]; + tensor obj_15_strides_0 = const()[name = tensor("obj_15_strides_0"), val = tensor([1, 1])]; + tensor obj_15_pad_0 = const()[name = tensor("obj_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_15_dilations_0 = const()[name = tensor("obj_15_dilations_0"), val = tensor([1, 1])]; + tensor obj_15_groups_0 = const()[name = tensor("obj_15_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142565120)))]; + tensor layers_3_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145841984)))]; + tensor obj_15_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = obj_15_dilations_0, groups = obj_15_groups_0, pad = obj_15_pad_0, pad_type = obj_15_pad_type_0, strides = obj_15_strides_0, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("obj_15_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_15_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([1])]; + tensor var_617_to_fp16 = const()[name = tensor("op_617_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_617_to_fp16, x = inputs_15_cast_fp16)[name = tensor("out_15_cast_fp16")]; + tensor input_27_gamma_0_to_fp16 = const()[name = tensor("input_27_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145844608)))]; + tensor input_27_beta_0_to_fp16 = const()[name = tensor("input_27_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145847232)))]; + tensor input_27_epsilon_0_to_fp16 = const()[name = tensor("input_27_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = batch_norm(beta = input_27_beta_0_to_fp16, epsilon = input_27_epsilon_0_to_fp16, gamma = input_27_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1, 1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1, 1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor layers_3_fc1_weight_to_fp16 = const()[name = tensor("layers_3_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145849856)))]; + tensor layers_3_fc1_bias_to_fp16 = const()[name = tensor("layers_3_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158957120)))]; + tensor input_29_cast_fp16 = conv(bias = layers_3_fc1_bias_to_fp16, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = layers_3_fc1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor input_31_mode_0 = const()[name = tensor("input_31_mode_0"), val = tensor("EXACT")]; + tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = input_29_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor hidden_states_11_pad_type_0 = const()[name = tensor("hidden_states_11_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_11_strides_0 = const()[name = tensor("hidden_states_11_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_11_pad_0 = const()[name = tensor("hidden_states_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_11_dilations_0 = const()[name = tensor("hidden_states_11_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_11_groups_0 = const()[name = tensor("hidden_states_11_groups_0"), val = tensor(1)]; + tensor layers_3_fc2_weight_to_fp16 = const()[name = tensor("layers_3_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158967424)))]; + tensor layers_3_fc2_bias_to_fp16 = const()[name = tensor("layers_3_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172074688)))]; + tensor hidden_states_11_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = hidden_states_11_dilations_0, groups = hidden_states_11_groups_0, pad = hidden_states_11_pad_0, pad_type = hidden_states_11_pad_type_0, strides = hidden_states_11_strides_0, weight = layers_3_fc2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("hidden_states_11_cast_fp16")]; + tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = hidden_states_11_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; + tensor var_650 = const()[name = tensor("op_650"), val = tensor(3)]; + tensor out_17_axes_0 = const()[name = tensor("out_17_axes_0"), val = tensor([1])]; + tensor var_669_to_fp16 = const()[name = tensor("op_669_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_669_to_fp16, x = inputs_17_cast_fp16)[name = tensor("out_17_cast_fp16")]; + tensor obj_17_gamma_0_to_fp16 = const()[name = tensor("obj_17_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172077312)))]; + tensor obj_17_beta_0_to_fp16 = const()[name = tensor("obj_17_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172079936)))]; + tensor obj_17_epsilon_0_to_fp16 = const()[name = tensor("obj_17_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_17_cast_fp16 = batch_norm(beta = obj_17_beta_0_to_fp16, epsilon = obj_17_epsilon_0_to_fp16, gamma = obj_17_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor("obj_17_cast_fp16")]; + tensor query_9_pad_type_0 = const()[name = tensor("query_9_pad_type_0"), val = tensor("valid")]; + tensor query_9_strides_0 = const()[name = tensor("query_9_strides_0"), val = tensor([1, 1])]; + tensor query_9_pad_0 = const()[name = tensor("query_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_9_dilations_0 = const()[name = tensor("query_9_dilations_0"), val = tensor([1, 1])]; + tensor query_9_groups_0 = const()[name = tensor("query_9_groups_0"), val = tensor(1)]; + tensor layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172082560)))]; + tensor layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(175359424)))]; + tensor query_9_cast_fp16 = conv(bias = layers_4_self_attn_q_proj_bias_to_fp16, dilations = query_9_dilations_0, groups = query_9_groups_0, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = query_9_strides_0, weight = layers_4_self_attn_q_proj_weight_to_fp16, x = obj_17_cast_fp16)[name = tensor("query_9_cast_fp16")]; + tensor key_9_pad_type_0 = const()[name = tensor("key_9_pad_type_0"), val = tensor("valid")]; + tensor key_9_strides_0 = const()[name = tensor("key_9_strides_0"), val = tensor([1, 1])]; + tensor key_9_pad_0 = const()[name = tensor("key_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_9_dilations_0 = const()[name = tensor("key_9_dilations_0"), val = tensor([1, 1])]; + tensor key_9_groups_0 = const()[name = tensor("key_9_groups_0"), val = tensor(1)]; + tensor layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(175362048)))]; + tensor key_9_cast_fp16 = conv(dilations = key_9_dilations_0, groups = key_9_groups_0, pad = key_9_pad_0, pad_type = key_9_pad_type_0, strides = key_9_strides_0, weight = layers_4_self_attn_k_proj_weight_to_fp16, x = obj_17_cast_fp16)[name = tensor("key_9_cast_fp16")]; + tensor value_9_pad_type_0 = const()[name = tensor("value_9_pad_type_0"), val = tensor("valid")]; + tensor value_9_strides_0 = const()[name = tensor("value_9_strides_0"), val = tensor([1, 1])]; + tensor value_9_pad_0 = const()[name = tensor("value_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_9_dilations_0 = const()[name = tensor("value_9_dilations_0"), val = tensor([1, 1])]; + tensor value_9_groups_0 = const()[name = tensor("value_9_groups_0"), val = tensor(1)]; + tensor layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178638912)))]; + tensor layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181915776)))]; + tensor value_9_cast_fp16 = conv(bias = layers_4_self_attn_v_proj_bias_to_fp16, dilations = value_9_dilations_0, groups = value_9_groups_0, pad = value_9_pad_0, pad_type = value_9_pad_type_0, strides = value_9_strides_0, weight = layers_4_self_attn_v_proj_weight_to_fp16, x = obj_17_cast_fp16)[name = tensor("value_9_cast_fp16")]; + tensor var_704 = const()[name = tensor("op_704"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_9_cast_fp16 = reshape(shape = var_704, x = query_9_cast_fp16)[name = tensor("mh_q_9_cast_fp16")]; + tensor var_706_to_fp16 = const()[name = tensor("op_706_to_fp16"), val = tensor(0x1p-3)]; + tensor var_707_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_706_to_fp16)[name = tensor("op_707_cast_fp16")]; + tensor var_708 = const()[name = tensor("op_708"), val = tensor([1, 20, 64, -1])]; + tensor var_709_cast_fp16 = reshape(shape = var_708, x = key_9_cast_fp16)[name = tensor("op_709_cast_fp16")]; + tensor mh_w_9_transpose_x_0 = const()[name = tensor("mh_w_9_transpose_x_0"), val = tensor(true)]; + tensor mh_w_9_transpose_y_0 = const()[name = tensor("mh_w_9_transpose_y_0"), val = tensor(false)]; + tensor mh_w_9_cast_fp16 = matmul(transpose_x = mh_w_9_transpose_x_0, transpose_y = mh_w_9_transpose_y_0, x = var_707_cast_fp16, y = var_709_cast_fp16)[name = tensor("mh_w_9_cast_fp16")]; + tensor var_712_cast_fp16 = softmax(axis = var_650, x = mh_w_9_cast_fp16)[name = tensor("op_712_cast_fp16")]; + tensor var_713 = const()[name = tensor("op_713"), val = tensor([1, 20, 64, -1])]; + tensor var_714_cast_fp16 = reshape(shape = var_713, x = value_9_cast_fp16)[name = tensor("op_714_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(true)]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_714_cast_fp16, y = var_712_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_717 = const()[name = tensor("op_717"), val = tensor([1, 1280, 1, -1])]; + tensor input_33_cast_fp16 = reshape(shape = var_717, x = attn_9_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor obj_19_pad_type_0 = const()[name = tensor("obj_19_pad_type_0"), val = tensor("valid")]; + tensor obj_19_strides_0 = const()[name = tensor("obj_19_strides_0"), val = tensor([1, 1])]; + tensor obj_19_pad_0 = const()[name = tensor("obj_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_19_dilations_0 = const()[name = tensor("obj_19_dilations_0"), val = tensor([1, 1])]; + tensor obj_19_groups_0 = const()[name = tensor("obj_19_groups_0"), val = tensor(1)]; + tensor layers_4_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181918400)))]; + tensor layers_4_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185195264)))]; + tensor obj_19_cast_fp16 = conv(bias = layers_4_self_attn_o_proj_bias_to_fp16, dilations = obj_19_dilations_0, groups = obj_19_groups_0, pad = obj_19_pad_0, pad_type = obj_19_pad_type_0, strides = obj_19_strides_0, weight = layers_4_self_attn_o_proj_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("obj_19_cast_fp16")]; + tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = obj_19_cast_fp16)[name = tensor("inputs_19_cast_fp16")]; + tensor out_19_axes_0 = const()[name = tensor("out_19_axes_0"), val = tensor([1])]; + tensor var_735_to_fp16 = const()[name = tensor("op_735_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_735_to_fp16, x = inputs_19_cast_fp16)[name = tensor("out_19_cast_fp16")]; + tensor input_35_gamma_0_to_fp16 = const()[name = tensor("input_35_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185197888)))]; + tensor input_35_beta_0_to_fp16 = const()[name = tensor("input_35_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185200512)))]; + tensor input_35_epsilon_0_to_fp16 = const()[name = tensor("input_35_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = batch_norm(beta = input_35_beta_0_to_fp16, epsilon = input_35_epsilon_0_to_fp16, gamma = input_35_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor input_37_pad_type_0 = const()[name = tensor("input_37_pad_type_0"), val = tensor("valid")]; + tensor input_37_strides_0 = const()[name = tensor("input_37_strides_0"), val = tensor([1, 1])]; + tensor input_37_pad_0 = const()[name = tensor("input_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_37_dilations_0 = const()[name = tensor("input_37_dilations_0"), val = tensor([1, 1])]; + tensor input_37_groups_0 = const()[name = tensor("input_37_groups_0"), val = tensor(1)]; + tensor layers_4_fc1_weight_to_fp16 = const()[name = tensor("layers_4_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185203136)))]; + tensor layers_4_fc1_bias_to_fp16 = const()[name = tensor("layers_4_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198310400)))]; + tensor input_37_cast_fp16 = conv(bias = layers_4_fc1_bias_to_fp16, dilations = input_37_dilations_0, groups = input_37_groups_0, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = input_37_strides_0, weight = layers_4_fc1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_mode_0 = const()[name = tensor("input_39_mode_0"), val = tensor("EXACT")]; + tensor input_39_cast_fp16 = gelu(mode = input_39_mode_0, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor hidden_states_13_pad_type_0 = const()[name = tensor("hidden_states_13_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_13_strides_0 = const()[name = tensor("hidden_states_13_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_13_pad_0 = const()[name = tensor("hidden_states_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_13_dilations_0 = const()[name = tensor("hidden_states_13_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_13_groups_0 = const()[name = tensor("hidden_states_13_groups_0"), val = tensor(1)]; + tensor layers_4_fc2_weight_to_fp16 = const()[name = tensor("layers_4_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198320704)))]; + tensor layers_4_fc2_bias_to_fp16 = const()[name = tensor("layers_4_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211427968)))]; + tensor hidden_states_13_cast_fp16 = conv(bias = layers_4_fc2_bias_to_fp16, dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_4_fc2_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("hidden_states_13_cast_fp16")]; + tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = hidden_states_13_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; + tensor var_768 = const()[name = tensor("op_768"), val = tensor(3)]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([1])]; + tensor var_787_to_fp16 = const()[name = tensor("op_787_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_787_to_fp16, x = inputs_21_cast_fp16)[name = tensor("out_21_cast_fp16")]; + tensor obj_21_gamma_0_to_fp16 = const()[name = tensor("obj_21_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211430592)))]; + tensor obj_21_beta_0_to_fp16 = const()[name = tensor("obj_21_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211433216)))]; + tensor obj_21_epsilon_0_to_fp16 = const()[name = tensor("obj_21_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_21_cast_fp16 = batch_norm(beta = obj_21_beta_0_to_fp16, epsilon = obj_21_epsilon_0_to_fp16, gamma = obj_21_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor("obj_21_cast_fp16")]; + tensor query_11_pad_type_0 = const()[name = tensor("query_11_pad_type_0"), val = tensor("valid")]; + tensor query_11_strides_0 = const()[name = tensor("query_11_strides_0"), val = tensor([1, 1])]; + tensor query_11_pad_0 = const()[name = tensor("query_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_11_dilations_0 = const()[name = tensor("query_11_dilations_0"), val = tensor([1, 1])]; + tensor query_11_groups_0 = const()[name = tensor("query_11_groups_0"), val = tensor(1)]; + tensor layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211435840)))]; + tensor layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214712704)))]; + tensor query_11_cast_fp16 = conv(bias = layers_5_self_attn_q_proj_bias_to_fp16, dilations = query_11_dilations_0, groups = query_11_groups_0, pad = query_11_pad_0, pad_type = query_11_pad_type_0, strides = query_11_strides_0, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = obj_21_cast_fp16)[name = tensor("query_11_cast_fp16")]; + tensor key_11_pad_type_0 = const()[name = tensor("key_11_pad_type_0"), val = tensor("valid")]; + tensor key_11_strides_0 = const()[name = tensor("key_11_strides_0"), val = tensor([1, 1])]; + tensor key_11_pad_0 = const()[name = tensor("key_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_11_dilations_0 = const()[name = tensor("key_11_dilations_0"), val = tensor([1, 1])]; + tensor key_11_groups_0 = const()[name = tensor("key_11_groups_0"), val = tensor(1)]; + tensor layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214715328)))]; + tensor key_11_cast_fp16 = conv(dilations = key_11_dilations_0, groups = key_11_groups_0, pad = key_11_pad_0, pad_type = key_11_pad_type_0, strides = key_11_strides_0, weight = layers_5_self_attn_k_proj_weight_to_fp16, x = obj_21_cast_fp16)[name = tensor("key_11_cast_fp16")]; + tensor value_11_pad_type_0 = const()[name = tensor("value_11_pad_type_0"), val = tensor("valid")]; + tensor value_11_strides_0 = const()[name = tensor("value_11_strides_0"), val = tensor([1, 1])]; + tensor value_11_pad_0 = const()[name = tensor("value_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_11_dilations_0 = const()[name = tensor("value_11_dilations_0"), val = tensor([1, 1])]; + tensor value_11_groups_0 = const()[name = tensor("value_11_groups_0"), val = tensor(1)]; + tensor layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217992192)))]; + tensor layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221269056)))]; + tensor value_11_cast_fp16 = conv(bias = layers_5_self_attn_v_proj_bias_to_fp16, dilations = value_11_dilations_0, groups = value_11_groups_0, pad = value_11_pad_0, pad_type = value_11_pad_type_0, strides = value_11_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = obj_21_cast_fp16)[name = tensor("value_11_cast_fp16")]; + tensor var_822 = const()[name = tensor("op_822"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_11_cast_fp16 = reshape(shape = var_822, x = query_11_cast_fp16)[name = tensor("mh_q_11_cast_fp16")]; + tensor var_824_to_fp16 = const()[name = tensor("op_824_to_fp16"), val = tensor(0x1p-3)]; + tensor var_825_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_824_to_fp16)[name = tensor("op_825_cast_fp16")]; + tensor var_826 = const()[name = tensor("op_826"), val = tensor([1, 20, 64, -1])]; + tensor var_827_cast_fp16 = reshape(shape = var_826, x = key_11_cast_fp16)[name = tensor("op_827_cast_fp16")]; + tensor mh_w_11_transpose_x_0 = const()[name = tensor("mh_w_11_transpose_x_0"), val = tensor(true)]; + tensor mh_w_11_transpose_y_0 = const()[name = tensor("mh_w_11_transpose_y_0"), val = tensor(false)]; + tensor mh_w_11_cast_fp16 = matmul(transpose_x = mh_w_11_transpose_x_0, transpose_y = mh_w_11_transpose_y_0, x = var_825_cast_fp16, y = var_827_cast_fp16)[name = tensor("mh_w_11_cast_fp16")]; + tensor var_830_cast_fp16 = softmax(axis = var_768, x = mh_w_11_cast_fp16)[name = tensor("op_830_cast_fp16")]; + tensor var_831 = const()[name = tensor("op_831"), val = tensor([1, 20, 64, -1])]; + tensor var_832_cast_fp16 = reshape(shape = var_831, x = value_11_cast_fp16)[name = tensor("op_832_cast_fp16")]; + tensor attn_11_transpose_x_0 = const()[name = tensor("attn_11_transpose_x_0"), val = tensor(false)]; + tensor attn_11_transpose_y_0 = const()[name = tensor("attn_11_transpose_y_0"), val = tensor(true)]; + tensor attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_832_cast_fp16, y = var_830_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_835 = const()[name = tensor("op_835"), val = tensor([1, 1280, 1, -1])]; + tensor input_41_cast_fp16 = reshape(shape = var_835, x = attn_11_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor obj_23_pad_type_0 = const()[name = tensor("obj_23_pad_type_0"), val = tensor("valid")]; + tensor obj_23_strides_0 = const()[name = tensor("obj_23_strides_0"), val = tensor([1, 1])]; + tensor obj_23_pad_0 = const()[name = tensor("obj_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_23_dilations_0 = const()[name = tensor("obj_23_dilations_0"), val = tensor([1, 1])]; + tensor obj_23_groups_0 = const()[name = tensor("obj_23_groups_0"), val = tensor(1)]; + tensor layers_5_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221271680)))]; + tensor layers_5_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224548544)))]; + tensor obj_23_cast_fp16 = conv(bias = layers_5_self_attn_o_proj_bias_to_fp16, dilations = obj_23_dilations_0, groups = obj_23_groups_0, pad = obj_23_pad_0, pad_type = obj_23_pad_type_0, strides = obj_23_strides_0, weight = layers_5_self_attn_o_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("obj_23_cast_fp16")]; + tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_23_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; + tensor out_23_axes_0 = const()[name = tensor("out_23_axes_0"), val = tensor([1])]; + tensor var_853_to_fp16 = const()[name = tensor("op_853_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_853_to_fp16, x = inputs_23_cast_fp16)[name = tensor("out_23_cast_fp16")]; + tensor input_43_gamma_0_to_fp16 = const()[name = tensor("input_43_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224551168)))]; + tensor input_43_beta_0_to_fp16 = const()[name = tensor("input_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224553792)))]; + tensor input_43_epsilon_0_to_fp16 = const()[name = tensor("input_43_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_43_cast_fp16 = batch_norm(beta = input_43_beta_0_to_fp16, epsilon = input_43_epsilon_0_to_fp16, gamma = input_43_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor input_45_pad_type_0 = const()[name = tensor("input_45_pad_type_0"), val = tensor("valid")]; + tensor input_45_strides_0 = const()[name = tensor("input_45_strides_0"), val = tensor([1, 1])]; + tensor input_45_pad_0 = const()[name = tensor("input_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_45_dilations_0 = const()[name = tensor("input_45_dilations_0"), val = tensor([1, 1])]; + tensor input_45_groups_0 = const()[name = tensor("input_45_groups_0"), val = tensor(1)]; + tensor layers_5_fc1_weight_to_fp16 = const()[name = tensor("layers_5_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224556416)))]; + tensor layers_5_fc1_bias_to_fp16 = const()[name = tensor("layers_5_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237663680)))]; + tensor input_45_cast_fp16 = conv(bias = layers_5_fc1_bias_to_fp16, dilations = input_45_dilations_0, groups = input_45_groups_0, pad = input_45_pad_0, pad_type = input_45_pad_type_0, strides = input_45_strides_0, weight = layers_5_fc1_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor input_47_mode_0 = const()[name = tensor("input_47_mode_0"), val = tensor("EXACT")]; + tensor input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor hidden_states_15_pad_type_0 = const()[name = tensor("hidden_states_15_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_15_strides_0 = const()[name = tensor("hidden_states_15_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_15_pad_0 = const()[name = tensor("hidden_states_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_15_dilations_0 = const()[name = tensor("hidden_states_15_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_15_groups_0 = const()[name = tensor("hidden_states_15_groups_0"), val = tensor(1)]; + tensor layers_5_fc2_weight_to_fp16 = const()[name = tensor("layers_5_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237673984)))]; + tensor layers_5_fc2_bias_to_fp16 = const()[name = tensor("layers_5_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250781248)))]; + tensor hidden_states_15_cast_fp16 = conv(bias = layers_5_fc2_bias_to_fp16, dilations = hidden_states_15_dilations_0, groups = hidden_states_15_groups_0, pad = hidden_states_15_pad_0, pad_type = hidden_states_15_pad_type_0, strides = hidden_states_15_strides_0, weight = layers_5_fc2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("hidden_states_15_cast_fp16")]; + tensor inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_15_cast_fp16)[name = tensor("inputs_25_cast_fp16")]; + tensor var_886 = const()[name = tensor("op_886"), val = tensor(3)]; + tensor out_25_axes_0 = const()[name = tensor("out_25_axes_0"), val = tensor([1])]; + tensor var_905_to_fp16 = const()[name = tensor("op_905_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_905_to_fp16, x = inputs_25_cast_fp16)[name = tensor("out_25_cast_fp16")]; + tensor obj_25_gamma_0_to_fp16 = const()[name = tensor("obj_25_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250783872)))]; + tensor obj_25_beta_0_to_fp16 = const()[name = tensor("obj_25_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250786496)))]; + tensor obj_25_epsilon_0_to_fp16 = const()[name = tensor("obj_25_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_25_cast_fp16 = batch_norm(beta = obj_25_beta_0_to_fp16, epsilon = obj_25_epsilon_0_to_fp16, gamma = obj_25_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_25_cast_fp16)[name = tensor("obj_25_cast_fp16")]; + tensor query_13_pad_type_0 = const()[name = tensor("query_13_pad_type_0"), val = tensor("valid")]; + tensor query_13_strides_0 = const()[name = tensor("query_13_strides_0"), val = tensor([1, 1])]; + tensor query_13_pad_0 = const()[name = tensor("query_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_13_dilations_0 = const()[name = tensor("query_13_dilations_0"), val = tensor([1, 1])]; + tensor query_13_groups_0 = const()[name = tensor("query_13_groups_0"), val = tensor(1)]; + tensor layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250789120)))]; + tensor layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254065984)))]; + tensor query_13_cast_fp16 = conv(bias = layers_6_self_attn_q_proj_bias_to_fp16, dilations = query_13_dilations_0, groups = query_13_groups_0, pad = query_13_pad_0, pad_type = query_13_pad_type_0, strides = query_13_strides_0, weight = layers_6_self_attn_q_proj_weight_to_fp16, x = obj_25_cast_fp16)[name = tensor("query_13_cast_fp16")]; + tensor key_13_pad_type_0 = const()[name = tensor("key_13_pad_type_0"), val = tensor("valid")]; + tensor key_13_strides_0 = const()[name = tensor("key_13_strides_0"), val = tensor([1, 1])]; + tensor key_13_pad_0 = const()[name = tensor("key_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_13_dilations_0 = const()[name = tensor("key_13_dilations_0"), val = tensor([1, 1])]; + tensor key_13_groups_0 = const()[name = tensor("key_13_groups_0"), val = tensor(1)]; + tensor layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254068608)))]; + tensor key_13_cast_fp16 = conv(dilations = key_13_dilations_0, groups = key_13_groups_0, pad = key_13_pad_0, pad_type = key_13_pad_type_0, strides = key_13_strides_0, weight = layers_6_self_attn_k_proj_weight_to_fp16, x = obj_25_cast_fp16)[name = tensor("key_13_cast_fp16")]; + tensor value_13_pad_type_0 = const()[name = tensor("value_13_pad_type_0"), val = tensor("valid")]; + tensor value_13_strides_0 = const()[name = tensor("value_13_strides_0"), val = tensor([1, 1])]; + tensor value_13_pad_0 = const()[name = tensor("value_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_13_dilations_0 = const()[name = tensor("value_13_dilations_0"), val = tensor([1, 1])]; + tensor value_13_groups_0 = const()[name = tensor("value_13_groups_0"), val = tensor(1)]; + tensor layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257345472)))]; + tensor layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260622336)))]; + tensor value_13_cast_fp16 = conv(bias = layers_6_self_attn_v_proj_bias_to_fp16, dilations = value_13_dilations_0, groups = value_13_groups_0, pad = value_13_pad_0, pad_type = value_13_pad_type_0, strides = value_13_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = obj_25_cast_fp16)[name = tensor("value_13_cast_fp16")]; + tensor var_940 = const()[name = tensor("op_940"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_13_cast_fp16 = reshape(shape = var_940, x = query_13_cast_fp16)[name = tensor("mh_q_13_cast_fp16")]; + tensor var_942_to_fp16 = const()[name = tensor("op_942_to_fp16"), val = tensor(0x1p-3)]; + tensor var_943_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_942_to_fp16)[name = tensor("op_943_cast_fp16")]; + tensor var_944 = const()[name = tensor("op_944"), val = tensor([1, 20, 64, -1])]; + tensor var_945_cast_fp16 = reshape(shape = var_944, x = key_13_cast_fp16)[name = tensor("op_945_cast_fp16")]; + tensor mh_w_13_transpose_x_0 = const()[name = tensor("mh_w_13_transpose_x_0"), val = tensor(true)]; + tensor mh_w_13_transpose_y_0 = const()[name = tensor("mh_w_13_transpose_y_0"), val = tensor(false)]; + tensor mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_943_cast_fp16, y = var_945_cast_fp16)[name = tensor("mh_w_13_cast_fp16")]; + tensor var_948_cast_fp16 = softmax(axis = var_886, x = mh_w_13_cast_fp16)[name = tensor("op_948_cast_fp16")]; + tensor var_949 = const()[name = tensor("op_949"), val = tensor([1, 20, 64, -1])]; + tensor var_950_cast_fp16 = reshape(shape = var_949, x = value_13_cast_fp16)[name = tensor("op_950_cast_fp16")]; + tensor attn_13_transpose_x_0 = const()[name = tensor("attn_13_transpose_x_0"), val = tensor(false)]; + tensor attn_13_transpose_y_0 = const()[name = tensor("attn_13_transpose_y_0"), val = tensor(true)]; + tensor attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_950_cast_fp16, y = var_948_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_953 = const()[name = tensor("op_953"), val = tensor([1, 1280, 1, -1])]; + tensor input_49_cast_fp16 = reshape(shape = var_953, x = attn_13_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor obj_27_pad_type_0 = const()[name = tensor("obj_27_pad_type_0"), val = tensor("valid")]; + tensor obj_27_strides_0 = const()[name = tensor("obj_27_strides_0"), val = tensor([1, 1])]; + tensor obj_27_pad_0 = const()[name = tensor("obj_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_27_dilations_0 = const()[name = tensor("obj_27_dilations_0"), val = tensor([1, 1])]; + tensor obj_27_groups_0 = const()[name = tensor("obj_27_groups_0"), val = tensor(1)]; + tensor layers_6_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260624960)))]; + tensor layers_6_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263901824)))]; + tensor obj_27_cast_fp16 = conv(bias = layers_6_self_attn_o_proj_bias_to_fp16, dilations = obj_27_dilations_0, groups = obj_27_groups_0, pad = obj_27_pad_0, pad_type = obj_27_pad_type_0, strides = obj_27_strides_0, weight = layers_6_self_attn_o_proj_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("obj_27_cast_fp16")]; + tensor inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = obj_27_cast_fp16)[name = tensor("inputs_27_cast_fp16")]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([1])]; + tensor var_971_to_fp16 = const()[name = tensor("op_971_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_971_to_fp16, x = inputs_27_cast_fp16)[name = tensor("out_27_cast_fp16")]; + tensor input_51_gamma_0_to_fp16 = const()[name = tensor("input_51_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263904448)))]; + tensor input_51_beta_0_to_fp16 = const()[name = tensor("input_51_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263907072)))]; + tensor input_51_epsilon_0_to_fp16 = const()[name = tensor("input_51_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_51_cast_fp16 = batch_norm(beta = input_51_beta_0_to_fp16, epsilon = input_51_epsilon_0_to_fp16, gamma = input_51_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_27_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor input_53_pad_type_0 = const()[name = tensor("input_53_pad_type_0"), val = tensor("valid")]; + tensor input_53_strides_0 = const()[name = tensor("input_53_strides_0"), val = tensor([1, 1])]; + tensor input_53_pad_0 = const()[name = tensor("input_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_53_dilations_0 = const()[name = tensor("input_53_dilations_0"), val = tensor([1, 1])]; + tensor input_53_groups_0 = const()[name = tensor("input_53_groups_0"), val = tensor(1)]; + tensor layers_6_fc1_weight_to_fp16 = const()[name = tensor("layers_6_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263909696)))]; + tensor layers_6_fc1_bias_to_fp16 = const()[name = tensor("layers_6_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277016960)))]; + tensor input_53_cast_fp16 = conv(bias = layers_6_fc1_bias_to_fp16, dilations = input_53_dilations_0, groups = input_53_groups_0, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = input_53_strides_0, weight = layers_6_fc1_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor input_55_mode_0 = const()[name = tensor("input_55_mode_0"), val = tensor("EXACT")]; + tensor input_55_cast_fp16 = gelu(mode = input_55_mode_0, x = input_53_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor hidden_states_17_pad_type_0 = const()[name = tensor("hidden_states_17_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_17_strides_0 = const()[name = tensor("hidden_states_17_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_17_pad_0 = const()[name = tensor("hidden_states_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_17_dilations_0 = const()[name = tensor("hidden_states_17_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_17_groups_0 = const()[name = tensor("hidden_states_17_groups_0"), val = tensor(1)]; + tensor layers_6_fc2_weight_to_fp16 = const()[name = tensor("layers_6_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277027264)))]; + tensor layers_6_fc2_bias_to_fp16 = const()[name = tensor("layers_6_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290134528)))]; + tensor hidden_states_17_cast_fp16 = conv(bias = layers_6_fc2_bias_to_fp16, dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_6_fc2_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("hidden_states_17_cast_fp16")]; + tensor inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = hidden_states_17_cast_fp16)[name = tensor("inputs_29_cast_fp16")]; + tensor var_1004 = const()[name = tensor("op_1004"), val = tensor(3)]; + tensor out_29_axes_0 = const()[name = tensor("out_29_axes_0"), val = tensor([1])]; + tensor var_1023_to_fp16 = const()[name = tensor("op_1023_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_1023_to_fp16, x = inputs_29_cast_fp16)[name = tensor("out_29_cast_fp16")]; + tensor obj_29_gamma_0_to_fp16 = const()[name = tensor("obj_29_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290137152)))]; + tensor obj_29_beta_0_to_fp16 = const()[name = tensor("obj_29_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290139776)))]; + tensor obj_29_epsilon_0_to_fp16 = const()[name = tensor("obj_29_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_29_cast_fp16 = batch_norm(beta = obj_29_beta_0_to_fp16, epsilon = obj_29_epsilon_0_to_fp16, gamma = obj_29_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_29_cast_fp16)[name = tensor("obj_29_cast_fp16")]; + tensor query_15_pad_type_0 = const()[name = tensor("query_15_pad_type_0"), val = tensor("valid")]; + tensor query_15_strides_0 = const()[name = tensor("query_15_strides_0"), val = tensor([1, 1])]; + tensor query_15_pad_0 = const()[name = tensor("query_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_15_dilations_0 = const()[name = tensor("query_15_dilations_0"), val = tensor([1, 1])]; + tensor query_15_groups_0 = const()[name = tensor("query_15_groups_0"), val = tensor(1)]; + tensor layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290142400)))]; + tensor layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293419264)))]; + tensor query_15_cast_fp16 = conv(bias = layers_7_self_attn_q_proj_bias_to_fp16, dilations = query_15_dilations_0, groups = query_15_groups_0, pad = query_15_pad_0, pad_type = query_15_pad_type_0, strides = query_15_strides_0, weight = layers_7_self_attn_q_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("query_15_cast_fp16")]; + tensor key_15_pad_type_0 = const()[name = tensor("key_15_pad_type_0"), val = tensor("valid")]; + tensor key_15_strides_0 = const()[name = tensor("key_15_strides_0"), val = tensor([1, 1])]; + tensor key_15_pad_0 = const()[name = tensor("key_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_15_dilations_0 = const()[name = tensor("key_15_dilations_0"), val = tensor([1, 1])]; + tensor key_15_groups_0 = const()[name = tensor("key_15_groups_0"), val = tensor(1)]; + tensor layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293421888)))]; + tensor key_15_cast_fp16 = conv(dilations = key_15_dilations_0, groups = key_15_groups_0, pad = key_15_pad_0, pad_type = key_15_pad_type_0, strides = key_15_strides_0, weight = layers_7_self_attn_k_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("key_15_cast_fp16")]; + tensor value_15_pad_type_0 = const()[name = tensor("value_15_pad_type_0"), val = tensor("valid")]; + tensor value_15_strides_0 = const()[name = tensor("value_15_strides_0"), val = tensor([1, 1])]; + tensor value_15_pad_0 = const()[name = tensor("value_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_15_dilations_0 = const()[name = tensor("value_15_dilations_0"), val = tensor([1, 1])]; + tensor value_15_groups_0 = const()[name = tensor("value_15_groups_0"), val = tensor(1)]; + tensor layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296698752)))]; + tensor layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299975616)))]; + tensor value_15_cast_fp16 = conv(bias = layers_7_self_attn_v_proj_bias_to_fp16, dilations = value_15_dilations_0, groups = value_15_groups_0, pad = value_15_pad_0, pad_type = value_15_pad_type_0, strides = value_15_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("value_15_cast_fp16")]; + tensor var_1058 = const()[name = tensor("op_1058"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_15_cast_fp16 = reshape(shape = var_1058, x = query_15_cast_fp16)[name = tensor("mh_q_15_cast_fp16")]; + tensor var_1060_to_fp16 = const()[name = tensor("op_1060_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1061_cast_fp16 = mul(x = mh_q_15_cast_fp16, y = var_1060_to_fp16)[name = tensor("op_1061_cast_fp16")]; + tensor var_1062 = const()[name = tensor("op_1062"), val = tensor([1, 20, 64, -1])]; + tensor var_1063_cast_fp16 = reshape(shape = var_1062, x = key_15_cast_fp16)[name = tensor("op_1063_cast_fp16")]; + tensor mh_w_15_transpose_x_0 = const()[name = tensor("mh_w_15_transpose_x_0"), val = tensor(true)]; + tensor mh_w_15_transpose_y_0 = const()[name = tensor("mh_w_15_transpose_y_0"), val = tensor(false)]; + tensor mh_w_15_cast_fp16 = matmul(transpose_x = mh_w_15_transpose_x_0, transpose_y = mh_w_15_transpose_y_0, x = var_1061_cast_fp16, y = var_1063_cast_fp16)[name = tensor("mh_w_15_cast_fp16")]; + tensor var_1066_cast_fp16 = softmax(axis = var_1004, x = mh_w_15_cast_fp16)[name = tensor("op_1066_cast_fp16")]; + tensor var_1067 = const()[name = tensor("op_1067"), val = tensor([1, 20, 64, -1])]; + tensor var_1068_cast_fp16 = reshape(shape = var_1067, x = value_15_cast_fp16)[name = tensor("op_1068_cast_fp16")]; + tensor attn_15_transpose_x_0 = const()[name = tensor("attn_15_transpose_x_0"), val = tensor(false)]; + tensor attn_15_transpose_y_0 = const()[name = tensor("attn_15_transpose_y_0"), val = tensor(true)]; + tensor attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = var_1068_cast_fp16, y = var_1066_cast_fp16)[name = tensor("attn_15_cast_fp16")]; + tensor var_1071 = const()[name = tensor("op_1071"), val = tensor([1, 1280, 1, -1])]; + tensor input_57_cast_fp16 = reshape(shape = var_1071, x = attn_15_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor obj_31_pad_type_0 = const()[name = tensor("obj_31_pad_type_0"), val = tensor("valid")]; + tensor obj_31_strides_0 = const()[name = tensor("obj_31_strides_0"), val = tensor([1, 1])]; + tensor obj_31_pad_0 = const()[name = tensor("obj_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_31_dilations_0 = const()[name = tensor("obj_31_dilations_0"), val = tensor([1, 1])]; + tensor obj_31_groups_0 = const()[name = tensor("obj_31_groups_0"), val = tensor(1)]; + tensor layers_7_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299978240)))]; + tensor layers_7_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303255104)))]; + tensor obj_31_cast_fp16 = conv(bias = layers_7_self_attn_o_proj_bias_to_fp16, dilations = obj_31_dilations_0, groups = obj_31_groups_0, pad = obj_31_pad_0, pad_type = obj_31_pad_type_0, strides = obj_31_strides_0, weight = layers_7_self_attn_o_proj_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("obj_31_cast_fp16")]; + tensor inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = obj_31_cast_fp16)[name = tensor("inputs_31_cast_fp16")]; + tensor out_31_axes_0 = const()[name = tensor("out_31_axes_0"), val = tensor([1])]; + tensor var_1089_to_fp16 = const()[name = tensor("op_1089_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_1089_to_fp16, x = inputs_31_cast_fp16)[name = tensor("out_31_cast_fp16")]; + tensor input_59_gamma_0_to_fp16 = const()[name = tensor("input_59_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303257728)))]; + tensor input_59_beta_0_to_fp16 = const()[name = tensor("input_59_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303260352)))]; + tensor input_59_epsilon_0_to_fp16 = const()[name = tensor("input_59_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = batch_norm(beta = input_59_beta_0_to_fp16, epsilon = input_59_epsilon_0_to_fp16, gamma = input_59_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_31_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor input_61_pad_type_0 = const()[name = tensor("input_61_pad_type_0"), val = tensor("valid")]; + tensor input_61_strides_0 = const()[name = tensor("input_61_strides_0"), val = tensor([1, 1])]; + tensor input_61_pad_0 = const()[name = tensor("input_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_61_dilations_0 = const()[name = tensor("input_61_dilations_0"), val = tensor([1, 1])]; + tensor input_61_groups_0 = const()[name = tensor("input_61_groups_0"), val = tensor(1)]; + tensor layers_7_fc1_weight_to_fp16 = const()[name = tensor("layers_7_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303262976)))]; + tensor layers_7_fc1_bias_to_fp16 = const()[name = tensor("layers_7_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316370240)))]; + tensor input_61_cast_fp16 = conv(bias = layers_7_fc1_bias_to_fp16, dilations = input_61_dilations_0, groups = input_61_groups_0, pad = input_61_pad_0, pad_type = input_61_pad_type_0, strides = input_61_strides_0, weight = layers_7_fc1_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor input_63_mode_0 = const()[name = tensor("input_63_mode_0"), val = tensor("EXACT")]; + tensor input_63_cast_fp16 = gelu(mode = input_63_mode_0, x = input_61_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor hidden_states_19_pad_type_0 = const()[name = tensor("hidden_states_19_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_19_strides_0 = const()[name = tensor("hidden_states_19_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_19_pad_0 = const()[name = tensor("hidden_states_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_19_dilations_0 = const()[name = tensor("hidden_states_19_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_19_groups_0 = const()[name = tensor("hidden_states_19_groups_0"), val = tensor(1)]; + tensor layers_7_fc2_weight_to_fp16 = const()[name = tensor("layers_7_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316380544)))]; + tensor layers_7_fc2_bias_to_fp16 = const()[name = tensor("layers_7_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329487808)))]; + tensor hidden_states_19_cast_fp16 = conv(bias = layers_7_fc2_bias_to_fp16, dilations = hidden_states_19_dilations_0, groups = hidden_states_19_groups_0, pad = hidden_states_19_pad_0, pad_type = hidden_states_19_pad_type_0, strides = hidden_states_19_strides_0, weight = layers_7_fc2_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("hidden_states_19_cast_fp16")]; + tensor inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = hidden_states_19_cast_fp16)[name = tensor("inputs_33_cast_fp16")]; + tensor var_1122 = const()[name = tensor("op_1122"), val = tensor(3)]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([1])]; + tensor var_1141_to_fp16 = const()[name = tensor("op_1141_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_1141_to_fp16, x = inputs_33_cast_fp16)[name = tensor("out_33_cast_fp16")]; + tensor obj_33_gamma_0_to_fp16 = const()[name = tensor("obj_33_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329490432)))]; + tensor obj_33_beta_0_to_fp16 = const()[name = tensor("obj_33_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329493056)))]; + tensor obj_33_epsilon_0_to_fp16 = const()[name = tensor("obj_33_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_33_cast_fp16 = batch_norm(beta = obj_33_beta_0_to_fp16, epsilon = obj_33_epsilon_0_to_fp16, gamma = obj_33_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_33_cast_fp16)[name = tensor("obj_33_cast_fp16")]; + tensor query_17_pad_type_0 = const()[name = tensor("query_17_pad_type_0"), val = tensor("valid")]; + tensor query_17_strides_0 = const()[name = tensor("query_17_strides_0"), val = tensor([1, 1])]; + tensor query_17_pad_0 = const()[name = tensor("query_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_17_dilations_0 = const()[name = tensor("query_17_dilations_0"), val = tensor([1, 1])]; + tensor query_17_groups_0 = const()[name = tensor("query_17_groups_0"), val = tensor(1)]; + tensor layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329495680)))]; + tensor layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332772544)))]; + tensor query_17_cast_fp16 = conv(bias = layers_8_self_attn_q_proj_bias_to_fp16, dilations = query_17_dilations_0, groups = query_17_groups_0, pad = query_17_pad_0, pad_type = query_17_pad_type_0, strides = query_17_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = obj_33_cast_fp16)[name = tensor("query_17_cast_fp16")]; + tensor key_17_pad_type_0 = const()[name = tensor("key_17_pad_type_0"), val = tensor("valid")]; + tensor key_17_strides_0 = const()[name = tensor("key_17_strides_0"), val = tensor([1, 1])]; + tensor key_17_pad_0 = const()[name = tensor("key_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_17_dilations_0 = const()[name = tensor("key_17_dilations_0"), val = tensor([1, 1])]; + tensor key_17_groups_0 = const()[name = tensor("key_17_groups_0"), val = tensor(1)]; + tensor layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332775168)))]; + tensor key_17_cast_fp16 = conv(dilations = key_17_dilations_0, groups = key_17_groups_0, pad = key_17_pad_0, pad_type = key_17_pad_type_0, strides = key_17_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = obj_33_cast_fp16)[name = tensor("key_17_cast_fp16")]; + tensor value_17_pad_type_0 = const()[name = tensor("value_17_pad_type_0"), val = tensor("valid")]; + tensor value_17_strides_0 = const()[name = tensor("value_17_strides_0"), val = tensor([1, 1])]; + tensor value_17_pad_0 = const()[name = tensor("value_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_17_dilations_0 = const()[name = tensor("value_17_dilations_0"), val = tensor([1, 1])]; + tensor value_17_groups_0 = const()[name = tensor("value_17_groups_0"), val = tensor(1)]; + tensor layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336052032)))]; + tensor layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339328896)))]; + tensor value_17_cast_fp16 = conv(bias = layers_8_self_attn_v_proj_bias_to_fp16, dilations = value_17_dilations_0, groups = value_17_groups_0, pad = value_17_pad_0, pad_type = value_17_pad_type_0, strides = value_17_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = obj_33_cast_fp16)[name = tensor("value_17_cast_fp16")]; + tensor var_1176 = const()[name = tensor("op_1176"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_17_cast_fp16 = reshape(shape = var_1176, x = query_17_cast_fp16)[name = tensor("mh_q_17_cast_fp16")]; + tensor var_1178_to_fp16 = const()[name = tensor("op_1178_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1179_cast_fp16 = mul(x = mh_q_17_cast_fp16, y = var_1178_to_fp16)[name = tensor("op_1179_cast_fp16")]; + tensor var_1180 = const()[name = tensor("op_1180"), val = tensor([1, 20, 64, -1])]; + tensor var_1181_cast_fp16 = reshape(shape = var_1180, x = key_17_cast_fp16)[name = tensor("op_1181_cast_fp16")]; + tensor mh_w_17_transpose_x_0 = const()[name = tensor("mh_w_17_transpose_x_0"), val = tensor(true)]; + tensor mh_w_17_transpose_y_0 = const()[name = tensor("mh_w_17_transpose_y_0"), val = tensor(false)]; + tensor mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_1179_cast_fp16, y = var_1181_cast_fp16)[name = tensor("mh_w_17_cast_fp16")]; + tensor var_1184_cast_fp16 = softmax(axis = var_1122, x = mh_w_17_cast_fp16)[name = tensor("op_1184_cast_fp16")]; + tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([1, 20, 64, -1])]; + tensor var_1186_cast_fp16 = reshape(shape = var_1185, x = value_17_cast_fp16)[name = tensor("op_1186_cast_fp16")]; + tensor attn_17_transpose_x_0 = const()[name = tensor("attn_17_transpose_x_0"), val = tensor(false)]; + tensor attn_17_transpose_y_0 = const()[name = tensor("attn_17_transpose_y_0"), val = tensor(true)]; + tensor attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = var_1186_cast_fp16, y = var_1184_cast_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor var_1189 = const()[name = tensor("op_1189"), val = tensor([1, 1280, 1, -1])]; + tensor input_65_cast_fp16 = reshape(shape = var_1189, x = attn_17_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor obj_35_pad_type_0 = const()[name = tensor("obj_35_pad_type_0"), val = tensor("valid")]; + tensor obj_35_strides_0 = const()[name = tensor("obj_35_strides_0"), val = tensor([1, 1])]; + tensor obj_35_pad_0 = const()[name = tensor("obj_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_35_dilations_0 = const()[name = tensor("obj_35_dilations_0"), val = tensor([1, 1])]; + tensor obj_35_groups_0 = const()[name = tensor("obj_35_groups_0"), val = tensor(1)]; + tensor layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339331520)))]; + tensor layers_8_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342608384)))]; + tensor obj_35_cast_fp16 = conv(bias = layers_8_self_attn_o_proj_bias_to_fp16, dilations = obj_35_dilations_0, groups = obj_35_groups_0, pad = obj_35_pad_0, pad_type = obj_35_pad_type_0, strides = obj_35_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("obj_35_cast_fp16")]; + tensor inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = obj_35_cast_fp16)[name = tensor("inputs_35_cast_fp16")]; + tensor out_35_axes_0 = const()[name = tensor("out_35_axes_0"), val = tensor([1])]; + tensor var_1207_to_fp16 = const()[name = tensor("op_1207_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_1207_to_fp16, x = inputs_35_cast_fp16)[name = tensor("out_35_cast_fp16")]; + tensor input_67_gamma_0_to_fp16 = const()[name = tensor("input_67_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342611008)))]; + tensor input_67_beta_0_to_fp16 = const()[name = tensor("input_67_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342613632)))]; + tensor input_67_epsilon_0_to_fp16 = const()[name = tensor("input_67_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_67_cast_fp16 = batch_norm(beta = input_67_beta_0_to_fp16, epsilon = input_67_epsilon_0_to_fp16, gamma = input_67_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_35_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1, 1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1, 1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor layers_8_fc1_weight_to_fp16 = const()[name = tensor("layers_8_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342616256)))]; + tensor layers_8_fc1_bias_to_fp16 = const()[name = tensor("layers_8_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355723520)))]; + tensor input_69_cast_fp16 = conv(bias = layers_8_fc1_bias_to_fp16, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = layers_8_fc1_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_mode_0 = const()[name = tensor("input_71_mode_0"), val = tensor("EXACT")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor hidden_states_21_pad_type_0 = const()[name = tensor("hidden_states_21_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_21_strides_0 = const()[name = tensor("hidden_states_21_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_21_pad_0 = const()[name = tensor("hidden_states_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_21_dilations_0 = const()[name = tensor("hidden_states_21_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_21_groups_0 = const()[name = tensor("hidden_states_21_groups_0"), val = tensor(1)]; + tensor layers_8_fc2_weight_to_fp16 = const()[name = tensor("layers_8_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355733824)))]; + tensor layers_8_fc2_bias_to_fp16 = const()[name = tensor("layers_8_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368841088)))]; + tensor hidden_states_21_cast_fp16 = conv(bias = layers_8_fc2_bias_to_fp16, dilations = hidden_states_21_dilations_0, groups = hidden_states_21_groups_0, pad = hidden_states_21_pad_0, pad_type = hidden_states_21_pad_type_0, strides = hidden_states_21_strides_0, weight = layers_8_fc2_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("hidden_states_21_cast_fp16")]; + tensor inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = hidden_states_21_cast_fp16)[name = tensor("inputs_37_cast_fp16")]; + tensor var_1240 = const()[name = tensor("op_1240"), val = tensor(3)]; + tensor out_37_axes_0 = const()[name = tensor("out_37_axes_0"), val = tensor([1])]; + tensor var_1259_to_fp16 = const()[name = tensor("op_1259_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_1259_to_fp16, x = inputs_37_cast_fp16)[name = tensor("out_37_cast_fp16")]; + tensor obj_37_gamma_0_to_fp16 = const()[name = tensor("obj_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368843712)))]; + tensor obj_37_beta_0_to_fp16 = const()[name = tensor("obj_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368846336)))]; + tensor obj_37_epsilon_0_to_fp16 = const()[name = tensor("obj_37_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_37_cast_fp16 = batch_norm(beta = obj_37_beta_0_to_fp16, epsilon = obj_37_epsilon_0_to_fp16, gamma = obj_37_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_37_cast_fp16)[name = tensor("obj_37_cast_fp16")]; + tensor query_19_pad_type_0 = const()[name = tensor("query_19_pad_type_0"), val = tensor("valid")]; + tensor query_19_strides_0 = const()[name = tensor("query_19_strides_0"), val = tensor([1, 1])]; + tensor query_19_pad_0 = const()[name = tensor("query_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_19_dilations_0 = const()[name = tensor("query_19_dilations_0"), val = tensor([1, 1])]; + tensor query_19_groups_0 = const()[name = tensor("query_19_groups_0"), val = tensor(1)]; + tensor layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368848960)))]; + tensor layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(372125824)))]; + tensor query_19_cast_fp16 = conv(bias = layers_9_self_attn_q_proj_bias_to_fp16, dilations = query_19_dilations_0, groups = query_19_groups_0, pad = query_19_pad_0, pad_type = query_19_pad_type_0, strides = query_19_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor("query_19_cast_fp16")]; + tensor key_19_pad_type_0 = const()[name = tensor("key_19_pad_type_0"), val = tensor("valid")]; + tensor key_19_strides_0 = const()[name = tensor("key_19_strides_0"), val = tensor([1, 1])]; + tensor key_19_pad_0 = const()[name = tensor("key_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_19_dilations_0 = const()[name = tensor("key_19_dilations_0"), val = tensor([1, 1])]; + tensor key_19_groups_0 = const()[name = tensor("key_19_groups_0"), val = tensor(1)]; + tensor layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(372128448)))]; + tensor key_19_cast_fp16 = conv(dilations = key_19_dilations_0, groups = key_19_groups_0, pad = key_19_pad_0, pad_type = key_19_pad_type_0, strides = key_19_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor("key_19_cast_fp16")]; + tensor value_19_pad_type_0 = const()[name = tensor("value_19_pad_type_0"), val = tensor("valid")]; + tensor value_19_strides_0 = const()[name = tensor("value_19_strides_0"), val = tensor([1, 1])]; + tensor value_19_pad_0 = const()[name = tensor("value_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_19_dilations_0 = const()[name = tensor("value_19_dilations_0"), val = tensor([1, 1])]; + tensor value_19_groups_0 = const()[name = tensor("value_19_groups_0"), val = tensor(1)]; + tensor layers_9_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375405312)))]; + tensor layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378682176)))]; + tensor value_19_cast_fp16 = conv(bias = layers_9_self_attn_v_proj_bias_to_fp16, dilations = value_19_dilations_0, groups = value_19_groups_0, pad = value_19_pad_0, pad_type = value_19_pad_type_0, strides = value_19_strides_0, weight = layers_9_self_attn_v_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor("value_19_cast_fp16")]; + tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_19_cast_fp16 = reshape(shape = var_1294, x = query_19_cast_fp16)[name = tensor("mh_q_19_cast_fp16")]; + tensor var_1296_to_fp16 = const()[name = tensor("op_1296_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1297_cast_fp16 = mul(x = mh_q_19_cast_fp16, y = var_1296_to_fp16)[name = tensor("op_1297_cast_fp16")]; + tensor var_1298 = const()[name = tensor("op_1298"), val = tensor([1, 20, 64, -1])]; + tensor var_1299_cast_fp16 = reshape(shape = var_1298, x = key_19_cast_fp16)[name = tensor("op_1299_cast_fp16")]; + tensor mh_w_19_transpose_x_0 = const()[name = tensor("mh_w_19_transpose_x_0"), val = tensor(true)]; + tensor mh_w_19_transpose_y_0 = const()[name = tensor("mh_w_19_transpose_y_0"), val = tensor(false)]; + tensor mh_w_19_cast_fp16 = matmul(transpose_x = mh_w_19_transpose_x_0, transpose_y = mh_w_19_transpose_y_0, x = var_1297_cast_fp16, y = var_1299_cast_fp16)[name = tensor("mh_w_19_cast_fp16")]; + tensor var_1302_cast_fp16 = softmax(axis = var_1240, x = mh_w_19_cast_fp16)[name = tensor("op_1302_cast_fp16")]; + tensor var_1303 = const()[name = tensor("op_1303"), val = tensor([1, 20, 64, -1])]; + tensor var_1304_cast_fp16 = reshape(shape = var_1303, x = value_19_cast_fp16)[name = tensor("op_1304_cast_fp16")]; + tensor attn_19_transpose_x_0 = const()[name = tensor("attn_19_transpose_x_0"), val = tensor(false)]; + tensor attn_19_transpose_y_0 = const()[name = tensor("attn_19_transpose_y_0"), val = tensor(true)]; + tensor attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = var_1304_cast_fp16, y = var_1302_cast_fp16)[name = tensor("attn_19_cast_fp16")]; + tensor var_1307 = const()[name = tensor("op_1307"), val = tensor([1, 1280, 1, -1])]; + tensor input_73_cast_fp16 = reshape(shape = var_1307, x = attn_19_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor obj_39_pad_type_0 = const()[name = tensor("obj_39_pad_type_0"), val = tensor("valid")]; + tensor obj_39_strides_0 = const()[name = tensor("obj_39_strides_0"), val = tensor([1, 1])]; + tensor obj_39_pad_0 = const()[name = tensor("obj_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_39_dilations_0 = const()[name = tensor("obj_39_dilations_0"), val = tensor([1, 1])]; + tensor obj_39_groups_0 = const()[name = tensor("obj_39_groups_0"), val = tensor(1)]; + tensor layers_9_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378684800)))]; + tensor layers_9_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381961664)))]; + tensor obj_39_cast_fp16 = conv(bias = layers_9_self_attn_o_proj_bias_to_fp16, dilations = obj_39_dilations_0, groups = obj_39_groups_0, pad = obj_39_pad_0, pad_type = obj_39_pad_type_0, strides = obj_39_strides_0, weight = layers_9_self_attn_o_proj_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("obj_39_cast_fp16")]; + tensor inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = obj_39_cast_fp16)[name = tensor("inputs_39_cast_fp16")]; + tensor out_39_axes_0 = const()[name = tensor("out_39_axes_0"), val = tensor([1])]; + tensor var_1325_to_fp16 = const()[name = tensor("op_1325_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_1325_to_fp16, x = inputs_39_cast_fp16)[name = tensor("out_39_cast_fp16")]; + tensor input_75_gamma_0_to_fp16 = const()[name = tensor("input_75_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381964288)))]; + tensor input_75_beta_0_to_fp16 = const()[name = tensor("input_75_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381966912)))]; + tensor input_75_epsilon_0_to_fp16 = const()[name = tensor("input_75_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_75_cast_fp16 = batch_norm(beta = input_75_beta_0_to_fp16, epsilon = input_75_epsilon_0_to_fp16, gamma = input_75_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_39_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_77_pad_type_0 = const()[name = tensor("input_77_pad_type_0"), val = tensor("valid")]; + tensor input_77_strides_0 = const()[name = tensor("input_77_strides_0"), val = tensor([1, 1])]; + tensor input_77_pad_0 = const()[name = tensor("input_77_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_77_dilations_0 = const()[name = tensor("input_77_dilations_0"), val = tensor([1, 1])]; + tensor input_77_groups_0 = const()[name = tensor("input_77_groups_0"), val = tensor(1)]; + tensor layers_9_fc1_weight_to_fp16 = const()[name = tensor("layers_9_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381969536)))]; + tensor layers_9_fc1_bias_to_fp16 = const()[name = tensor("layers_9_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395076800)))]; + tensor input_77_cast_fp16 = conv(bias = layers_9_fc1_bias_to_fp16, dilations = input_77_dilations_0, groups = input_77_groups_0, pad = input_77_pad_0, pad_type = input_77_pad_type_0, strides = input_77_strides_0, weight = layers_9_fc1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor input_79_mode_0 = const()[name = tensor("input_79_mode_0"), val = tensor("EXACT")]; + tensor input_79_cast_fp16 = gelu(mode = input_79_mode_0, x = input_77_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor hidden_states_23_pad_type_0 = const()[name = tensor("hidden_states_23_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_23_strides_0 = const()[name = tensor("hidden_states_23_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_23_pad_0 = const()[name = tensor("hidden_states_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_23_dilations_0 = const()[name = tensor("hidden_states_23_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_23_groups_0 = const()[name = tensor("hidden_states_23_groups_0"), val = tensor(1)]; + tensor layers_9_fc2_weight_to_fp16 = const()[name = tensor("layers_9_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395087104)))]; + tensor layers_9_fc2_bias_to_fp16 = const()[name = tensor("layers_9_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408194368)))]; + tensor hidden_states_23_cast_fp16 = conv(bias = layers_9_fc2_bias_to_fp16, dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_9_fc2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("hidden_states_23_cast_fp16")]; + tensor inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = hidden_states_23_cast_fp16)[name = tensor("inputs_41_cast_fp16")]; + tensor var_1358 = const()[name = tensor("op_1358"), val = tensor(3)]; + tensor out_41_axes_0 = const()[name = tensor("out_41_axes_0"), val = tensor([1])]; + tensor var_1377_to_fp16 = const()[name = tensor("op_1377_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_1377_to_fp16, x = inputs_41_cast_fp16)[name = tensor("out_41_cast_fp16")]; + tensor obj_41_gamma_0_to_fp16 = const()[name = tensor("obj_41_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408196992)))]; + tensor obj_41_beta_0_to_fp16 = const()[name = tensor("obj_41_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408199616)))]; + tensor obj_41_epsilon_0_to_fp16 = const()[name = tensor("obj_41_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_41_cast_fp16 = batch_norm(beta = obj_41_beta_0_to_fp16, epsilon = obj_41_epsilon_0_to_fp16, gamma = obj_41_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_41_cast_fp16)[name = tensor("obj_41_cast_fp16")]; + tensor query_21_pad_type_0 = const()[name = tensor("query_21_pad_type_0"), val = tensor("valid")]; + tensor query_21_strides_0 = const()[name = tensor("query_21_strides_0"), val = tensor([1, 1])]; + tensor query_21_pad_0 = const()[name = tensor("query_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_21_dilations_0 = const()[name = tensor("query_21_dilations_0"), val = tensor([1, 1])]; + tensor query_21_groups_0 = const()[name = tensor("query_21_groups_0"), val = tensor(1)]; + tensor layers_10_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408202240)))]; + tensor layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411479104)))]; + tensor query_21_cast_fp16 = conv(bias = layers_10_self_attn_q_proj_bias_to_fp16, dilations = query_21_dilations_0, groups = query_21_groups_0, pad = query_21_pad_0, pad_type = query_21_pad_type_0, strides = query_21_strides_0, weight = layers_10_self_attn_q_proj_weight_to_fp16, x = obj_41_cast_fp16)[name = tensor("query_21_cast_fp16")]; + tensor key_21_pad_type_0 = const()[name = tensor("key_21_pad_type_0"), val = tensor("valid")]; + tensor key_21_strides_0 = const()[name = tensor("key_21_strides_0"), val = tensor([1, 1])]; + tensor key_21_pad_0 = const()[name = tensor("key_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_21_dilations_0 = const()[name = tensor("key_21_dilations_0"), val = tensor([1, 1])]; + tensor key_21_groups_0 = const()[name = tensor("key_21_groups_0"), val = tensor(1)]; + tensor layers_10_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411481728)))]; + tensor key_21_cast_fp16 = conv(dilations = key_21_dilations_0, groups = key_21_groups_0, pad = key_21_pad_0, pad_type = key_21_pad_type_0, strides = key_21_strides_0, weight = layers_10_self_attn_k_proj_weight_to_fp16, x = obj_41_cast_fp16)[name = tensor("key_21_cast_fp16")]; + tensor value_21_pad_type_0 = const()[name = tensor("value_21_pad_type_0"), val = tensor("valid")]; + tensor value_21_strides_0 = const()[name = tensor("value_21_strides_0"), val = tensor([1, 1])]; + tensor value_21_pad_0 = const()[name = tensor("value_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_21_dilations_0 = const()[name = tensor("value_21_dilations_0"), val = tensor([1, 1])]; + tensor value_21_groups_0 = const()[name = tensor("value_21_groups_0"), val = tensor(1)]; + tensor layers_10_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(414758592)))]; + tensor layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(418035456)))]; + tensor value_21_cast_fp16 = conv(bias = layers_10_self_attn_v_proj_bias_to_fp16, dilations = value_21_dilations_0, groups = value_21_groups_0, pad = value_21_pad_0, pad_type = value_21_pad_type_0, strides = value_21_strides_0, weight = layers_10_self_attn_v_proj_weight_to_fp16, x = obj_41_cast_fp16)[name = tensor("value_21_cast_fp16")]; + tensor var_1412 = const()[name = tensor("op_1412"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_21_cast_fp16 = reshape(shape = var_1412, x = query_21_cast_fp16)[name = tensor("mh_q_21_cast_fp16")]; + tensor var_1414_to_fp16 = const()[name = tensor("op_1414_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1415_cast_fp16 = mul(x = mh_q_21_cast_fp16, y = var_1414_to_fp16)[name = tensor("op_1415_cast_fp16")]; + tensor var_1416 = const()[name = tensor("op_1416"), val = tensor([1, 20, 64, -1])]; + tensor var_1417_cast_fp16 = reshape(shape = var_1416, x = key_21_cast_fp16)[name = tensor("op_1417_cast_fp16")]; + tensor mh_w_21_transpose_x_0 = const()[name = tensor("mh_w_21_transpose_x_0"), val = tensor(true)]; + tensor mh_w_21_transpose_y_0 = const()[name = tensor("mh_w_21_transpose_y_0"), val = tensor(false)]; + tensor mh_w_21_cast_fp16 = matmul(transpose_x = mh_w_21_transpose_x_0, transpose_y = mh_w_21_transpose_y_0, x = var_1415_cast_fp16, y = var_1417_cast_fp16)[name = tensor("mh_w_21_cast_fp16")]; + tensor var_1420_cast_fp16 = softmax(axis = var_1358, x = mh_w_21_cast_fp16)[name = tensor("op_1420_cast_fp16")]; + tensor var_1421 = const()[name = tensor("op_1421"), val = tensor([1, 20, 64, -1])]; + tensor var_1422_cast_fp16 = reshape(shape = var_1421, x = value_21_cast_fp16)[name = tensor("op_1422_cast_fp16")]; + tensor attn_21_transpose_x_0 = const()[name = tensor("attn_21_transpose_x_0"), val = tensor(false)]; + tensor attn_21_transpose_y_0 = const()[name = tensor("attn_21_transpose_y_0"), val = tensor(true)]; + tensor attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = var_1422_cast_fp16, y = var_1420_cast_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor var_1425 = const()[name = tensor("op_1425"), val = tensor([1, 1280, 1, -1])]; + tensor input_81_cast_fp16 = reshape(shape = var_1425, x = attn_21_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor obj_43_pad_type_0 = const()[name = tensor("obj_43_pad_type_0"), val = tensor("valid")]; + tensor obj_43_strides_0 = const()[name = tensor("obj_43_strides_0"), val = tensor([1, 1])]; + tensor obj_43_pad_0 = const()[name = tensor("obj_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_43_dilations_0 = const()[name = tensor("obj_43_dilations_0"), val = tensor([1, 1])]; + tensor obj_43_groups_0 = const()[name = tensor("obj_43_groups_0"), val = tensor(1)]; + tensor layers_10_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(418038080)))]; + tensor layers_10_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421314944)))]; + tensor obj_43_cast_fp16 = conv(bias = layers_10_self_attn_o_proj_bias_to_fp16, dilations = obj_43_dilations_0, groups = obj_43_groups_0, pad = obj_43_pad_0, pad_type = obj_43_pad_type_0, strides = obj_43_strides_0, weight = layers_10_self_attn_o_proj_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("obj_43_cast_fp16")]; + tensor inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = obj_43_cast_fp16)[name = tensor("inputs_43_cast_fp16")]; + tensor out_43_axes_0 = const()[name = tensor("out_43_axes_0"), val = tensor([1])]; + tensor var_1443_to_fp16 = const()[name = tensor("op_1443_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_1443_to_fp16, x = inputs_43_cast_fp16)[name = tensor("out_43_cast_fp16")]; + tensor input_83_gamma_0_to_fp16 = const()[name = tensor("input_83_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421317568)))]; + tensor input_83_beta_0_to_fp16 = const()[name = tensor("input_83_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421320192)))]; + tensor input_83_epsilon_0_to_fp16 = const()[name = tensor("input_83_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_83_cast_fp16 = batch_norm(beta = input_83_beta_0_to_fp16, epsilon = input_83_epsilon_0_to_fp16, gamma = input_83_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_43_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor input_85_pad_type_0 = const()[name = tensor("input_85_pad_type_0"), val = tensor("valid")]; + tensor input_85_strides_0 = const()[name = tensor("input_85_strides_0"), val = tensor([1, 1])]; + tensor input_85_pad_0 = const()[name = tensor("input_85_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_85_dilations_0 = const()[name = tensor("input_85_dilations_0"), val = tensor([1, 1])]; + tensor input_85_groups_0 = const()[name = tensor("input_85_groups_0"), val = tensor(1)]; + tensor layers_10_fc1_weight_to_fp16 = const()[name = tensor("layers_10_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421322816)))]; + tensor layers_10_fc1_bias_to_fp16 = const()[name = tensor("layers_10_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434430080)))]; + tensor input_85_cast_fp16 = conv(bias = layers_10_fc1_bias_to_fp16, dilations = input_85_dilations_0, groups = input_85_groups_0, pad = input_85_pad_0, pad_type = input_85_pad_type_0, strides = input_85_strides_0, weight = layers_10_fc1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor input_87_mode_0 = const()[name = tensor("input_87_mode_0"), val = tensor("EXACT")]; + tensor input_87_cast_fp16 = gelu(mode = input_87_mode_0, x = input_85_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor hidden_states_25_pad_type_0 = const()[name = tensor("hidden_states_25_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_25_strides_0 = const()[name = tensor("hidden_states_25_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_25_pad_0 = const()[name = tensor("hidden_states_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_25_dilations_0 = const()[name = tensor("hidden_states_25_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_25_groups_0 = const()[name = tensor("hidden_states_25_groups_0"), val = tensor(1)]; + tensor layers_10_fc2_weight_to_fp16 = const()[name = tensor("layers_10_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434440384)))]; + tensor layers_10_fc2_bias_to_fp16 = const()[name = tensor("layers_10_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447547648)))]; + tensor hidden_states_25_cast_fp16 = conv(bias = layers_10_fc2_bias_to_fp16, dilations = hidden_states_25_dilations_0, groups = hidden_states_25_groups_0, pad = hidden_states_25_pad_0, pad_type = hidden_states_25_pad_type_0, strides = hidden_states_25_strides_0, weight = layers_10_fc2_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("hidden_states_25_cast_fp16")]; + tensor inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = hidden_states_25_cast_fp16)[name = tensor("inputs_45_cast_fp16")]; + tensor var_1476 = const()[name = tensor("op_1476"), val = tensor(3)]; + tensor out_45_axes_0 = const()[name = tensor("out_45_axes_0"), val = tensor([1])]; + tensor var_1495_to_fp16 = const()[name = tensor("op_1495_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_1495_to_fp16, x = inputs_45_cast_fp16)[name = tensor("out_45_cast_fp16")]; + tensor obj_45_gamma_0_to_fp16 = const()[name = tensor("obj_45_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447550272)))]; + tensor obj_45_beta_0_to_fp16 = const()[name = tensor("obj_45_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447552896)))]; + tensor obj_45_epsilon_0_to_fp16 = const()[name = tensor("obj_45_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_45_cast_fp16 = batch_norm(beta = obj_45_beta_0_to_fp16, epsilon = obj_45_epsilon_0_to_fp16, gamma = obj_45_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_45_cast_fp16)[name = tensor("obj_45_cast_fp16")]; + tensor query_23_pad_type_0 = const()[name = tensor("query_23_pad_type_0"), val = tensor("valid")]; + tensor query_23_strides_0 = const()[name = tensor("query_23_strides_0"), val = tensor([1, 1])]; + tensor query_23_pad_0 = const()[name = tensor("query_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_23_dilations_0 = const()[name = tensor("query_23_dilations_0"), val = tensor([1, 1])]; + tensor query_23_groups_0 = const()[name = tensor("query_23_groups_0"), val = tensor(1)]; + tensor layers_11_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447555520)))]; + tensor layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450832384)))]; + tensor query_23_cast_fp16 = conv(bias = layers_11_self_attn_q_proj_bias_to_fp16, dilations = query_23_dilations_0, groups = query_23_groups_0, pad = query_23_pad_0, pad_type = query_23_pad_type_0, strides = query_23_strides_0, weight = layers_11_self_attn_q_proj_weight_to_fp16, x = obj_45_cast_fp16)[name = tensor("query_23_cast_fp16")]; + tensor key_23_pad_type_0 = const()[name = tensor("key_23_pad_type_0"), val = tensor("valid")]; + tensor key_23_strides_0 = const()[name = tensor("key_23_strides_0"), val = tensor([1, 1])]; + tensor key_23_pad_0 = const()[name = tensor("key_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_23_dilations_0 = const()[name = tensor("key_23_dilations_0"), val = tensor([1, 1])]; + tensor key_23_groups_0 = const()[name = tensor("key_23_groups_0"), val = tensor(1)]; + tensor layers_11_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450835008)))]; + tensor key_23_cast_fp16 = conv(dilations = key_23_dilations_0, groups = key_23_groups_0, pad = key_23_pad_0, pad_type = key_23_pad_type_0, strides = key_23_strides_0, weight = layers_11_self_attn_k_proj_weight_to_fp16, x = obj_45_cast_fp16)[name = tensor("key_23_cast_fp16")]; + tensor value_23_pad_type_0 = const()[name = tensor("value_23_pad_type_0"), val = tensor("valid")]; + tensor value_23_strides_0 = const()[name = tensor("value_23_strides_0"), val = tensor([1, 1])]; + tensor value_23_pad_0 = const()[name = tensor("value_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_23_dilations_0 = const()[name = tensor("value_23_dilations_0"), val = tensor([1, 1])]; + tensor value_23_groups_0 = const()[name = tensor("value_23_groups_0"), val = tensor(1)]; + tensor layers_11_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454111872)))]; + tensor layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457388736)))]; + tensor value_23_cast_fp16 = conv(bias = layers_11_self_attn_v_proj_bias_to_fp16, dilations = value_23_dilations_0, groups = value_23_groups_0, pad = value_23_pad_0, pad_type = value_23_pad_type_0, strides = value_23_strides_0, weight = layers_11_self_attn_v_proj_weight_to_fp16, x = obj_45_cast_fp16)[name = tensor("value_23_cast_fp16")]; + tensor var_1530 = const()[name = tensor("op_1530"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_23_cast_fp16 = reshape(shape = var_1530, x = query_23_cast_fp16)[name = tensor("mh_q_23_cast_fp16")]; + tensor var_1532_to_fp16 = const()[name = tensor("op_1532_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1533_cast_fp16 = mul(x = mh_q_23_cast_fp16, y = var_1532_to_fp16)[name = tensor("op_1533_cast_fp16")]; + tensor var_1534 = const()[name = tensor("op_1534"), val = tensor([1, 20, 64, -1])]; + tensor var_1535_cast_fp16 = reshape(shape = var_1534, x = key_23_cast_fp16)[name = tensor("op_1535_cast_fp16")]; + tensor mh_w_23_transpose_x_0 = const()[name = tensor("mh_w_23_transpose_x_0"), val = tensor(true)]; + tensor mh_w_23_transpose_y_0 = const()[name = tensor("mh_w_23_transpose_y_0"), val = tensor(false)]; + tensor mh_w_23_cast_fp16 = matmul(transpose_x = mh_w_23_transpose_x_0, transpose_y = mh_w_23_transpose_y_0, x = var_1533_cast_fp16, y = var_1535_cast_fp16)[name = tensor("mh_w_23_cast_fp16")]; + tensor var_1538_cast_fp16 = softmax(axis = var_1476, x = mh_w_23_cast_fp16)[name = tensor("op_1538_cast_fp16")]; + tensor var_1539 = const()[name = tensor("op_1539"), val = tensor([1, 20, 64, -1])]; + tensor var_1540_cast_fp16 = reshape(shape = var_1539, x = value_23_cast_fp16)[name = tensor("op_1540_cast_fp16")]; + tensor attn_23_transpose_x_0 = const()[name = tensor("attn_23_transpose_x_0"), val = tensor(false)]; + tensor attn_23_transpose_y_0 = const()[name = tensor("attn_23_transpose_y_0"), val = tensor(true)]; + tensor attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = var_1540_cast_fp16, y = var_1538_cast_fp16)[name = tensor("attn_23_cast_fp16")]; + tensor var_1543 = const()[name = tensor("op_1543"), val = tensor([1, 1280, 1, -1])]; + tensor input_89_cast_fp16 = reshape(shape = var_1543, x = attn_23_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor obj_47_pad_type_0 = const()[name = tensor("obj_47_pad_type_0"), val = tensor("valid")]; + tensor obj_47_strides_0 = const()[name = tensor("obj_47_strides_0"), val = tensor([1, 1])]; + tensor obj_47_pad_0 = const()[name = tensor("obj_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_47_dilations_0 = const()[name = tensor("obj_47_dilations_0"), val = tensor([1, 1])]; + tensor obj_47_groups_0 = const()[name = tensor("obj_47_groups_0"), val = tensor(1)]; + tensor layers_11_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457391360)))]; + tensor layers_11_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460668224)))]; + tensor obj_47_cast_fp16 = conv(bias = layers_11_self_attn_o_proj_bias_to_fp16, dilations = obj_47_dilations_0, groups = obj_47_groups_0, pad = obj_47_pad_0, pad_type = obj_47_pad_type_0, strides = obj_47_strides_0, weight = layers_11_self_attn_o_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("obj_47_cast_fp16")]; + tensor inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = obj_47_cast_fp16)[name = tensor("inputs_47_cast_fp16")]; + tensor out_47_axes_0 = const()[name = tensor("out_47_axes_0"), val = tensor([1])]; + tensor var_1561_to_fp16 = const()[name = tensor("op_1561_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_1561_to_fp16, x = inputs_47_cast_fp16)[name = tensor("out_47_cast_fp16")]; + tensor input_91_gamma_0_to_fp16 = const()[name = tensor("input_91_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460670848)))]; + tensor input_91_beta_0_to_fp16 = const()[name = tensor("input_91_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460673472)))]; + tensor input_91_epsilon_0_to_fp16 = const()[name = tensor("input_91_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_91_cast_fp16 = batch_norm(beta = input_91_beta_0_to_fp16, epsilon = input_91_epsilon_0_to_fp16, gamma = input_91_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_47_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor input_93_pad_type_0 = const()[name = tensor("input_93_pad_type_0"), val = tensor("valid")]; + tensor input_93_strides_0 = const()[name = tensor("input_93_strides_0"), val = tensor([1, 1])]; + tensor input_93_pad_0 = const()[name = tensor("input_93_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_93_dilations_0 = const()[name = tensor("input_93_dilations_0"), val = tensor([1, 1])]; + tensor input_93_groups_0 = const()[name = tensor("input_93_groups_0"), val = tensor(1)]; + tensor layers_11_fc1_weight_to_fp16 = const()[name = tensor("layers_11_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460676096)))]; + tensor layers_11_fc1_bias_to_fp16 = const()[name = tensor("layers_11_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(473783360)))]; + tensor input_93_cast_fp16 = conv(bias = layers_11_fc1_bias_to_fp16, dilations = input_93_dilations_0, groups = input_93_groups_0, pad = input_93_pad_0, pad_type = input_93_pad_type_0, strides = input_93_strides_0, weight = layers_11_fc1_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor input_95_mode_0 = const()[name = tensor("input_95_mode_0"), val = tensor("EXACT")]; + tensor input_95_cast_fp16 = gelu(mode = input_95_mode_0, x = input_93_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor hidden_states_27_pad_type_0 = const()[name = tensor("hidden_states_27_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_27_strides_0 = const()[name = tensor("hidden_states_27_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_27_pad_0 = const()[name = tensor("hidden_states_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_27_dilations_0 = const()[name = tensor("hidden_states_27_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_27_groups_0 = const()[name = tensor("hidden_states_27_groups_0"), val = tensor(1)]; + tensor layers_11_fc2_weight_to_fp16 = const()[name = tensor("layers_11_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(473793664)))]; + tensor layers_11_fc2_bias_to_fp16 = const()[name = tensor("layers_11_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486900928)))]; + tensor hidden_states_27_cast_fp16 = conv(bias = layers_11_fc2_bias_to_fp16, dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_11_fc2_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("hidden_states_27_cast_fp16")]; + tensor inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = hidden_states_27_cast_fp16)[name = tensor("inputs_49_cast_fp16")]; + tensor var_1594 = const()[name = tensor("op_1594"), val = tensor(3)]; + tensor out_49_axes_0 = const()[name = tensor("out_49_axes_0"), val = tensor([1])]; + tensor var_1613_to_fp16 = const()[name = tensor("op_1613_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_1613_to_fp16, x = inputs_49_cast_fp16)[name = tensor("out_49_cast_fp16")]; + tensor obj_49_gamma_0_to_fp16 = const()[name = tensor("obj_49_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486903552)))]; + tensor obj_49_beta_0_to_fp16 = const()[name = tensor("obj_49_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486906176)))]; + tensor obj_49_epsilon_0_to_fp16 = const()[name = tensor("obj_49_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_49_cast_fp16 = batch_norm(beta = obj_49_beta_0_to_fp16, epsilon = obj_49_epsilon_0_to_fp16, gamma = obj_49_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_49_cast_fp16)[name = tensor("obj_49_cast_fp16")]; + tensor query_25_pad_type_0 = const()[name = tensor("query_25_pad_type_0"), val = tensor("valid")]; + tensor query_25_strides_0 = const()[name = tensor("query_25_strides_0"), val = tensor([1, 1])]; + tensor query_25_pad_0 = const()[name = tensor("query_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_25_dilations_0 = const()[name = tensor("query_25_dilations_0"), val = tensor([1, 1])]; + tensor query_25_groups_0 = const()[name = tensor("query_25_groups_0"), val = tensor(1)]; + tensor layers_12_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486908800)))]; + tensor layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490185664)))]; + tensor query_25_cast_fp16 = conv(bias = layers_12_self_attn_q_proj_bias_to_fp16, dilations = query_25_dilations_0, groups = query_25_groups_0, pad = query_25_pad_0, pad_type = query_25_pad_type_0, strides = query_25_strides_0, weight = layers_12_self_attn_q_proj_weight_to_fp16, x = obj_49_cast_fp16)[name = tensor("query_25_cast_fp16")]; + tensor key_25_pad_type_0 = const()[name = tensor("key_25_pad_type_0"), val = tensor("valid")]; + tensor key_25_strides_0 = const()[name = tensor("key_25_strides_0"), val = tensor([1, 1])]; + tensor key_25_pad_0 = const()[name = tensor("key_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_25_dilations_0 = const()[name = tensor("key_25_dilations_0"), val = tensor([1, 1])]; + tensor key_25_groups_0 = const()[name = tensor("key_25_groups_0"), val = tensor(1)]; + tensor layers_12_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490188288)))]; + tensor key_25_cast_fp16 = conv(dilations = key_25_dilations_0, groups = key_25_groups_0, pad = key_25_pad_0, pad_type = key_25_pad_type_0, strides = key_25_strides_0, weight = layers_12_self_attn_k_proj_weight_to_fp16, x = obj_49_cast_fp16)[name = tensor("key_25_cast_fp16")]; + tensor value_25_pad_type_0 = const()[name = tensor("value_25_pad_type_0"), val = tensor("valid")]; + tensor value_25_strides_0 = const()[name = tensor("value_25_strides_0"), val = tensor([1, 1])]; + tensor value_25_pad_0 = const()[name = tensor("value_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_25_dilations_0 = const()[name = tensor("value_25_dilations_0"), val = tensor([1, 1])]; + tensor value_25_groups_0 = const()[name = tensor("value_25_groups_0"), val = tensor(1)]; + tensor layers_12_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493465152)))]; + tensor layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496742016)))]; + tensor value_25_cast_fp16 = conv(bias = layers_12_self_attn_v_proj_bias_to_fp16, dilations = value_25_dilations_0, groups = value_25_groups_0, pad = value_25_pad_0, pad_type = value_25_pad_type_0, strides = value_25_strides_0, weight = layers_12_self_attn_v_proj_weight_to_fp16, x = obj_49_cast_fp16)[name = tensor("value_25_cast_fp16")]; + tensor var_1648 = const()[name = tensor("op_1648"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_25_cast_fp16 = reshape(shape = var_1648, x = query_25_cast_fp16)[name = tensor("mh_q_25_cast_fp16")]; + tensor var_1650_to_fp16 = const()[name = tensor("op_1650_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1651_cast_fp16 = mul(x = mh_q_25_cast_fp16, y = var_1650_to_fp16)[name = tensor("op_1651_cast_fp16")]; + tensor var_1652 = const()[name = tensor("op_1652"), val = tensor([1, 20, 64, -1])]; + tensor var_1653_cast_fp16 = reshape(shape = var_1652, x = key_25_cast_fp16)[name = tensor("op_1653_cast_fp16")]; + tensor mh_w_25_transpose_x_0 = const()[name = tensor("mh_w_25_transpose_x_0"), val = tensor(true)]; + tensor mh_w_25_transpose_y_0 = const()[name = tensor("mh_w_25_transpose_y_0"), val = tensor(false)]; + tensor mh_w_25_cast_fp16 = matmul(transpose_x = mh_w_25_transpose_x_0, transpose_y = mh_w_25_transpose_y_0, x = var_1651_cast_fp16, y = var_1653_cast_fp16)[name = tensor("mh_w_25_cast_fp16")]; + tensor var_1656_cast_fp16 = softmax(axis = var_1594, x = mh_w_25_cast_fp16)[name = tensor("op_1656_cast_fp16")]; + tensor var_1657 = const()[name = tensor("op_1657"), val = tensor([1, 20, 64, -1])]; + tensor var_1658_cast_fp16 = reshape(shape = var_1657, x = value_25_cast_fp16)[name = tensor("op_1658_cast_fp16")]; + tensor attn_25_transpose_x_0 = const()[name = tensor("attn_25_transpose_x_0"), val = tensor(false)]; + tensor attn_25_transpose_y_0 = const()[name = tensor("attn_25_transpose_y_0"), val = tensor(true)]; + tensor attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = var_1658_cast_fp16, y = var_1656_cast_fp16)[name = tensor("attn_25_cast_fp16")]; + tensor var_1661 = const()[name = tensor("op_1661"), val = tensor([1, 1280, 1, -1])]; + tensor input_97_cast_fp16 = reshape(shape = var_1661, x = attn_25_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor obj_51_pad_type_0 = const()[name = tensor("obj_51_pad_type_0"), val = tensor("valid")]; + tensor obj_51_strides_0 = const()[name = tensor("obj_51_strides_0"), val = tensor([1, 1])]; + tensor obj_51_pad_0 = const()[name = tensor("obj_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_51_dilations_0 = const()[name = tensor("obj_51_dilations_0"), val = tensor([1, 1])]; + tensor obj_51_groups_0 = const()[name = tensor("obj_51_groups_0"), val = tensor(1)]; + tensor layers_12_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496744640)))]; + tensor layers_12_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(500021504)))]; + tensor obj_51_cast_fp16 = conv(bias = layers_12_self_attn_o_proj_bias_to_fp16, dilations = obj_51_dilations_0, groups = obj_51_groups_0, pad = obj_51_pad_0, pad_type = obj_51_pad_type_0, strides = obj_51_strides_0, weight = layers_12_self_attn_o_proj_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("obj_51_cast_fp16")]; + tensor inputs_51_cast_fp16 = add(x = inputs_49_cast_fp16, y = obj_51_cast_fp16)[name = tensor("inputs_51_cast_fp16")]; + tensor out_51_axes_0 = const()[name = tensor("out_51_axes_0"), val = tensor([1])]; + tensor var_1679_to_fp16 = const()[name = tensor("op_1679_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_1679_to_fp16, x = inputs_51_cast_fp16)[name = tensor("out_51_cast_fp16")]; + tensor input_99_gamma_0_to_fp16 = const()[name = tensor("input_99_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(500024128)))]; + tensor input_99_beta_0_to_fp16 = const()[name = tensor("input_99_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(500026752)))]; + tensor input_99_epsilon_0_to_fp16 = const()[name = tensor("input_99_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_99_cast_fp16 = batch_norm(beta = input_99_beta_0_to_fp16, epsilon = input_99_epsilon_0_to_fp16, gamma = input_99_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_51_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor input_101_pad_type_0 = const()[name = tensor("input_101_pad_type_0"), val = tensor("valid")]; + tensor input_101_strides_0 = const()[name = tensor("input_101_strides_0"), val = tensor([1, 1])]; + tensor input_101_pad_0 = const()[name = tensor("input_101_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_101_dilations_0 = const()[name = tensor("input_101_dilations_0"), val = tensor([1, 1])]; + tensor input_101_groups_0 = const()[name = tensor("input_101_groups_0"), val = tensor(1)]; + tensor layers_12_fc1_weight_to_fp16 = const()[name = tensor("layers_12_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(500029376)))]; + tensor layers_12_fc1_bias_to_fp16 = const()[name = tensor("layers_12_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513136640)))]; + tensor input_101_cast_fp16 = conv(bias = layers_12_fc1_bias_to_fp16, dilations = input_101_dilations_0, groups = input_101_groups_0, pad = input_101_pad_0, pad_type = input_101_pad_type_0, strides = input_101_strides_0, weight = layers_12_fc1_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor input_103_mode_0 = const()[name = tensor("input_103_mode_0"), val = tensor("EXACT")]; + tensor input_103_cast_fp16 = gelu(mode = input_103_mode_0, x = input_101_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor hidden_states_29_pad_type_0 = const()[name = tensor("hidden_states_29_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_29_strides_0 = const()[name = tensor("hidden_states_29_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_29_pad_0 = const()[name = tensor("hidden_states_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_29_dilations_0 = const()[name = tensor("hidden_states_29_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_29_groups_0 = const()[name = tensor("hidden_states_29_groups_0"), val = tensor(1)]; + tensor layers_12_fc2_weight_to_fp16 = const()[name = tensor("layers_12_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513146944)))]; + tensor layers_12_fc2_bias_to_fp16 = const()[name = tensor("layers_12_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(526254208)))]; + tensor hidden_states_29_cast_fp16 = conv(bias = layers_12_fc2_bias_to_fp16, dilations = hidden_states_29_dilations_0, groups = hidden_states_29_groups_0, pad = hidden_states_29_pad_0, pad_type = hidden_states_29_pad_type_0, strides = hidden_states_29_strides_0, weight = layers_12_fc2_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("hidden_states_29_cast_fp16")]; + tensor inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = hidden_states_29_cast_fp16)[name = tensor("inputs_53_cast_fp16")]; + tensor var_1712 = const()[name = tensor("op_1712"), val = tensor(3)]; + tensor out_53_axes_0 = const()[name = tensor("out_53_axes_0"), val = tensor([1])]; + tensor var_1731_to_fp16 = const()[name = tensor("op_1731_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_1731_to_fp16, x = inputs_53_cast_fp16)[name = tensor("out_53_cast_fp16")]; + tensor obj_53_gamma_0_to_fp16 = const()[name = tensor("obj_53_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(526256832)))]; + tensor obj_53_beta_0_to_fp16 = const()[name = tensor("obj_53_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(526259456)))]; + tensor obj_53_epsilon_0_to_fp16 = const()[name = tensor("obj_53_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_53_cast_fp16 = batch_norm(beta = obj_53_beta_0_to_fp16, epsilon = obj_53_epsilon_0_to_fp16, gamma = obj_53_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_53_cast_fp16)[name = tensor("obj_53_cast_fp16")]; + tensor query_27_pad_type_0 = const()[name = tensor("query_27_pad_type_0"), val = tensor("valid")]; + tensor query_27_strides_0 = const()[name = tensor("query_27_strides_0"), val = tensor([1, 1])]; + tensor query_27_pad_0 = const()[name = tensor("query_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_27_dilations_0 = const()[name = tensor("query_27_dilations_0"), val = tensor([1, 1])]; + tensor query_27_groups_0 = const()[name = tensor("query_27_groups_0"), val = tensor(1)]; + tensor layers_13_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(526262080)))]; + tensor layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529538944)))]; + tensor query_27_cast_fp16 = conv(bias = layers_13_self_attn_q_proj_bias_to_fp16, dilations = query_27_dilations_0, groups = query_27_groups_0, pad = query_27_pad_0, pad_type = query_27_pad_type_0, strides = query_27_strides_0, weight = layers_13_self_attn_q_proj_weight_to_fp16, x = obj_53_cast_fp16)[name = tensor("query_27_cast_fp16")]; + tensor key_27_pad_type_0 = const()[name = tensor("key_27_pad_type_0"), val = tensor("valid")]; + tensor key_27_strides_0 = const()[name = tensor("key_27_strides_0"), val = tensor([1, 1])]; + tensor key_27_pad_0 = const()[name = tensor("key_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_27_dilations_0 = const()[name = tensor("key_27_dilations_0"), val = tensor([1, 1])]; + tensor key_27_groups_0 = const()[name = tensor("key_27_groups_0"), val = tensor(1)]; + tensor layers_13_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529541568)))]; + tensor key_27_cast_fp16 = conv(dilations = key_27_dilations_0, groups = key_27_groups_0, pad = key_27_pad_0, pad_type = key_27_pad_type_0, strides = key_27_strides_0, weight = layers_13_self_attn_k_proj_weight_to_fp16, x = obj_53_cast_fp16)[name = tensor("key_27_cast_fp16")]; + tensor value_27_pad_type_0 = const()[name = tensor("value_27_pad_type_0"), val = tensor("valid")]; + tensor value_27_strides_0 = const()[name = tensor("value_27_strides_0"), val = tensor([1, 1])]; + tensor value_27_pad_0 = const()[name = tensor("value_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_27_dilations_0 = const()[name = tensor("value_27_dilations_0"), val = tensor([1, 1])]; + tensor value_27_groups_0 = const()[name = tensor("value_27_groups_0"), val = tensor(1)]; + tensor layers_13_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(532818432)))]; + tensor layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536095296)))]; + tensor value_27_cast_fp16 = conv(bias = layers_13_self_attn_v_proj_bias_to_fp16, dilations = value_27_dilations_0, groups = value_27_groups_0, pad = value_27_pad_0, pad_type = value_27_pad_type_0, strides = value_27_strides_0, weight = layers_13_self_attn_v_proj_weight_to_fp16, x = obj_53_cast_fp16)[name = tensor("value_27_cast_fp16")]; + tensor var_1766 = const()[name = tensor("op_1766"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_27_cast_fp16 = reshape(shape = var_1766, x = query_27_cast_fp16)[name = tensor("mh_q_27_cast_fp16")]; + tensor var_1768_to_fp16 = const()[name = tensor("op_1768_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1769_cast_fp16 = mul(x = mh_q_27_cast_fp16, y = var_1768_to_fp16)[name = tensor("op_1769_cast_fp16")]; + tensor var_1770 = const()[name = tensor("op_1770"), val = tensor([1, 20, 64, -1])]; + tensor var_1771_cast_fp16 = reshape(shape = var_1770, x = key_27_cast_fp16)[name = tensor("op_1771_cast_fp16")]; + tensor mh_w_27_transpose_x_0 = const()[name = tensor("mh_w_27_transpose_x_0"), val = tensor(true)]; + tensor mh_w_27_transpose_y_0 = const()[name = tensor("mh_w_27_transpose_y_0"), val = tensor(false)]; + tensor mh_w_27_cast_fp16 = matmul(transpose_x = mh_w_27_transpose_x_0, transpose_y = mh_w_27_transpose_y_0, x = var_1769_cast_fp16, y = var_1771_cast_fp16)[name = tensor("mh_w_27_cast_fp16")]; + tensor var_1774_cast_fp16 = softmax(axis = var_1712, x = mh_w_27_cast_fp16)[name = tensor("op_1774_cast_fp16")]; + tensor var_1775 = const()[name = tensor("op_1775"), val = tensor([1, 20, 64, -1])]; + tensor var_1776_cast_fp16 = reshape(shape = var_1775, x = value_27_cast_fp16)[name = tensor("op_1776_cast_fp16")]; + tensor attn_27_transpose_x_0 = const()[name = tensor("attn_27_transpose_x_0"), val = tensor(false)]; + tensor attn_27_transpose_y_0 = const()[name = tensor("attn_27_transpose_y_0"), val = tensor(true)]; + tensor attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = var_1776_cast_fp16, y = var_1774_cast_fp16)[name = tensor("attn_27_cast_fp16")]; + tensor var_1779 = const()[name = tensor("op_1779"), val = tensor([1, 1280, 1, -1])]; + tensor input_105_cast_fp16 = reshape(shape = var_1779, x = attn_27_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor obj_55_pad_type_0 = const()[name = tensor("obj_55_pad_type_0"), val = tensor("valid")]; + tensor obj_55_strides_0 = const()[name = tensor("obj_55_strides_0"), val = tensor([1, 1])]; + tensor obj_55_pad_0 = const()[name = tensor("obj_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_55_dilations_0 = const()[name = tensor("obj_55_dilations_0"), val = tensor([1, 1])]; + tensor obj_55_groups_0 = const()[name = tensor("obj_55_groups_0"), val = tensor(1)]; + tensor layers_13_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536097920)))]; + tensor layers_13_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539374784)))]; + tensor obj_55_cast_fp16 = conv(bias = layers_13_self_attn_o_proj_bias_to_fp16, dilations = obj_55_dilations_0, groups = obj_55_groups_0, pad = obj_55_pad_0, pad_type = obj_55_pad_type_0, strides = obj_55_strides_0, weight = layers_13_self_attn_o_proj_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("obj_55_cast_fp16")]; + tensor inputs_55_cast_fp16 = add(x = inputs_53_cast_fp16, y = obj_55_cast_fp16)[name = tensor("inputs_55_cast_fp16")]; + tensor out_55_axes_0 = const()[name = tensor("out_55_axes_0"), val = tensor([1])]; + tensor var_1797_to_fp16 = const()[name = tensor("op_1797_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_55_cast_fp16 = layer_norm(axes = out_55_axes_0, epsilon = var_1797_to_fp16, x = inputs_55_cast_fp16)[name = tensor("out_55_cast_fp16")]; + tensor input_107_gamma_0_to_fp16 = const()[name = tensor("input_107_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539377408)))]; + tensor input_107_beta_0_to_fp16 = const()[name = tensor("input_107_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539380032)))]; + tensor input_107_epsilon_0_to_fp16 = const()[name = tensor("input_107_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_107_cast_fp16 = batch_norm(beta = input_107_beta_0_to_fp16, epsilon = input_107_epsilon_0_to_fp16, gamma = input_107_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_55_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1, 1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1, 1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor layers_13_fc1_weight_to_fp16 = const()[name = tensor("layers_13_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539382656)))]; + tensor layers_13_fc1_bias_to_fp16 = const()[name = tensor("layers_13_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552489920)))]; + tensor input_109_cast_fp16 = conv(bias = layers_13_fc1_bias_to_fp16, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = layers_13_fc1_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor input_111_mode_0 = const()[name = tensor("input_111_mode_0"), val = tensor("EXACT")]; + tensor input_111_cast_fp16 = gelu(mode = input_111_mode_0, x = input_109_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor hidden_states_31_pad_type_0 = const()[name = tensor("hidden_states_31_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_31_strides_0 = const()[name = tensor("hidden_states_31_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_31_pad_0 = const()[name = tensor("hidden_states_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_31_dilations_0 = const()[name = tensor("hidden_states_31_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_31_groups_0 = const()[name = tensor("hidden_states_31_groups_0"), val = tensor(1)]; + tensor layers_13_fc2_weight_to_fp16 = const()[name = tensor("layers_13_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552500224)))]; + tensor layers_13_fc2_bias_to_fp16 = const()[name = tensor("layers_13_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565607488)))]; + tensor hidden_states_31_cast_fp16 = conv(bias = layers_13_fc2_bias_to_fp16, dilations = hidden_states_31_dilations_0, groups = hidden_states_31_groups_0, pad = hidden_states_31_pad_0, pad_type = hidden_states_31_pad_type_0, strides = hidden_states_31_strides_0, weight = layers_13_fc2_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("hidden_states_31_cast_fp16")]; + tensor inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = hidden_states_31_cast_fp16)[name = tensor("inputs_57_cast_fp16")]; + tensor var_1830 = const()[name = tensor("op_1830"), val = tensor(3)]; + tensor out_57_axes_0 = const()[name = tensor("out_57_axes_0"), val = tensor([1])]; + tensor var_1849_to_fp16 = const()[name = tensor("op_1849_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_57_cast_fp16 = layer_norm(axes = out_57_axes_0, epsilon = var_1849_to_fp16, x = inputs_57_cast_fp16)[name = tensor("out_57_cast_fp16")]; + tensor obj_57_gamma_0_to_fp16 = const()[name = tensor("obj_57_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565610112)))]; + tensor obj_57_beta_0_to_fp16 = const()[name = tensor("obj_57_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565612736)))]; + tensor obj_57_epsilon_0_to_fp16 = const()[name = tensor("obj_57_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_57_cast_fp16 = batch_norm(beta = obj_57_beta_0_to_fp16, epsilon = obj_57_epsilon_0_to_fp16, gamma = obj_57_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_57_cast_fp16)[name = tensor("obj_57_cast_fp16")]; + tensor query_29_pad_type_0 = const()[name = tensor("query_29_pad_type_0"), val = tensor("valid")]; + tensor query_29_strides_0 = const()[name = tensor("query_29_strides_0"), val = tensor([1, 1])]; + tensor query_29_pad_0 = const()[name = tensor("query_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_29_dilations_0 = const()[name = tensor("query_29_dilations_0"), val = tensor([1, 1])]; + tensor query_29_groups_0 = const()[name = tensor("query_29_groups_0"), val = tensor(1)]; + tensor layers_14_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565615360)))]; + tensor layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568892224)))]; + tensor query_29_cast_fp16 = conv(bias = layers_14_self_attn_q_proj_bias_to_fp16, dilations = query_29_dilations_0, groups = query_29_groups_0, pad = query_29_pad_0, pad_type = query_29_pad_type_0, strides = query_29_strides_0, weight = layers_14_self_attn_q_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("query_29_cast_fp16")]; + tensor key_29_pad_type_0 = const()[name = tensor("key_29_pad_type_0"), val = tensor("valid")]; + tensor key_29_strides_0 = const()[name = tensor("key_29_strides_0"), val = tensor([1, 1])]; + tensor key_29_pad_0 = const()[name = tensor("key_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_29_dilations_0 = const()[name = tensor("key_29_dilations_0"), val = tensor([1, 1])]; + tensor key_29_groups_0 = const()[name = tensor("key_29_groups_0"), val = tensor(1)]; + tensor layers_14_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568894848)))]; + tensor key_29_cast_fp16 = conv(dilations = key_29_dilations_0, groups = key_29_groups_0, pad = key_29_pad_0, pad_type = key_29_pad_type_0, strides = key_29_strides_0, weight = layers_14_self_attn_k_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("key_29_cast_fp16")]; + tensor value_29_pad_type_0 = const()[name = tensor("value_29_pad_type_0"), val = tensor("valid")]; + tensor value_29_strides_0 = const()[name = tensor("value_29_strides_0"), val = tensor([1, 1])]; + tensor value_29_pad_0 = const()[name = tensor("value_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_29_dilations_0 = const()[name = tensor("value_29_dilations_0"), val = tensor([1, 1])]; + tensor value_29_groups_0 = const()[name = tensor("value_29_groups_0"), val = tensor(1)]; + tensor layers_14_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(572171712)))]; + tensor layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(575448576)))]; + tensor value_29_cast_fp16 = conv(bias = layers_14_self_attn_v_proj_bias_to_fp16, dilations = value_29_dilations_0, groups = value_29_groups_0, pad = value_29_pad_0, pad_type = value_29_pad_type_0, strides = value_29_strides_0, weight = layers_14_self_attn_v_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("value_29_cast_fp16")]; + tensor var_1884 = const()[name = tensor("op_1884"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_29_cast_fp16 = reshape(shape = var_1884, x = query_29_cast_fp16)[name = tensor("mh_q_29_cast_fp16")]; + tensor var_1886_to_fp16 = const()[name = tensor("op_1886_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1887_cast_fp16 = mul(x = mh_q_29_cast_fp16, y = var_1886_to_fp16)[name = tensor("op_1887_cast_fp16")]; + tensor var_1888 = const()[name = tensor("op_1888"), val = tensor([1, 20, 64, -1])]; + tensor var_1889_cast_fp16 = reshape(shape = var_1888, x = key_29_cast_fp16)[name = tensor("op_1889_cast_fp16")]; + tensor mh_w_29_transpose_x_0 = const()[name = tensor("mh_w_29_transpose_x_0"), val = tensor(true)]; + tensor mh_w_29_transpose_y_0 = const()[name = tensor("mh_w_29_transpose_y_0"), val = tensor(false)]; + tensor mh_w_29_cast_fp16 = matmul(transpose_x = mh_w_29_transpose_x_0, transpose_y = mh_w_29_transpose_y_0, x = var_1887_cast_fp16, y = var_1889_cast_fp16)[name = tensor("mh_w_29_cast_fp16")]; + tensor var_1892_cast_fp16 = softmax(axis = var_1830, x = mh_w_29_cast_fp16)[name = tensor("op_1892_cast_fp16")]; + tensor var_1893 = const()[name = tensor("op_1893"), val = tensor([1, 20, 64, -1])]; + tensor var_1894_cast_fp16 = reshape(shape = var_1893, x = value_29_cast_fp16)[name = tensor("op_1894_cast_fp16")]; + tensor attn_29_transpose_x_0 = const()[name = tensor("attn_29_transpose_x_0"), val = tensor(false)]; + tensor attn_29_transpose_y_0 = const()[name = tensor("attn_29_transpose_y_0"), val = tensor(true)]; + tensor attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = var_1894_cast_fp16, y = var_1892_cast_fp16)[name = tensor("attn_29_cast_fp16")]; + tensor var_1897 = const()[name = tensor("op_1897"), val = tensor([1, 1280, 1, -1])]; + tensor input_113_cast_fp16 = reshape(shape = var_1897, x = attn_29_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor obj_59_pad_type_0 = const()[name = tensor("obj_59_pad_type_0"), val = tensor("valid")]; + tensor obj_59_strides_0 = const()[name = tensor("obj_59_strides_0"), val = tensor([1, 1])]; + tensor obj_59_pad_0 = const()[name = tensor("obj_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_59_dilations_0 = const()[name = tensor("obj_59_dilations_0"), val = tensor([1, 1])]; + tensor obj_59_groups_0 = const()[name = tensor("obj_59_groups_0"), val = tensor(1)]; + tensor layers_14_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(575451200)))]; + tensor layers_14_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578728064)))]; + tensor obj_59_cast_fp16 = conv(bias = layers_14_self_attn_o_proj_bias_to_fp16, dilations = obj_59_dilations_0, groups = obj_59_groups_0, pad = obj_59_pad_0, pad_type = obj_59_pad_type_0, strides = obj_59_strides_0, weight = layers_14_self_attn_o_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("obj_59_cast_fp16")]; + tensor inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = obj_59_cast_fp16)[name = tensor("inputs_59_cast_fp16")]; + tensor out_59_axes_0 = const()[name = tensor("out_59_axes_0"), val = tensor([1])]; + tensor var_1915_to_fp16 = const()[name = tensor("op_1915_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_59_cast_fp16 = layer_norm(axes = out_59_axes_0, epsilon = var_1915_to_fp16, x = inputs_59_cast_fp16)[name = tensor("out_59_cast_fp16")]; + tensor input_115_gamma_0_to_fp16 = const()[name = tensor("input_115_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578730688)))]; + tensor input_115_beta_0_to_fp16 = const()[name = tensor("input_115_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578733312)))]; + tensor input_115_epsilon_0_to_fp16 = const()[name = tensor("input_115_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_115_cast_fp16 = batch_norm(beta = input_115_beta_0_to_fp16, epsilon = input_115_epsilon_0_to_fp16, gamma = input_115_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_59_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor input_117_pad_type_0 = const()[name = tensor("input_117_pad_type_0"), val = tensor("valid")]; + tensor input_117_strides_0 = const()[name = tensor("input_117_strides_0"), val = tensor([1, 1])]; + tensor input_117_pad_0 = const()[name = tensor("input_117_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_117_dilations_0 = const()[name = tensor("input_117_dilations_0"), val = tensor([1, 1])]; + tensor input_117_groups_0 = const()[name = tensor("input_117_groups_0"), val = tensor(1)]; + tensor layers_14_fc1_weight_to_fp16 = const()[name = tensor("layers_14_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578735936)))]; + tensor layers_14_fc1_bias_to_fp16 = const()[name = tensor("layers_14_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591843200)))]; + tensor input_117_cast_fp16 = conv(bias = layers_14_fc1_bias_to_fp16, dilations = input_117_dilations_0, groups = input_117_groups_0, pad = input_117_pad_0, pad_type = input_117_pad_type_0, strides = input_117_strides_0, weight = layers_14_fc1_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_mode_0 = const()[name = tensor("input_119_mode_0"), val = tensor("EXACT")]; + tensor input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor hidden_states_33_pad_type_0 = const()[name = tensor("hidden_states_33_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_33_strides_0 = const()[name = tensor("hidden_states_33_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_33_pad_0 = const()[name = tensor("hidden_states_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_33_dilations_0 = const()[name = tensor("hidden_states_33_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_33_groups_0 = const()[name = tensor("hidden_states_33_groups_0"), val = tensor(1)]; + tensor layers_14_fc2_weight_to_fp16 = const()[name = tensor("layers_14_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591853504)))]; + tensor layers_14_fc2_bias_to_fp16 = const()[name = tensor("layers_14_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604960768)))]; + tensor hidden_states_33_cast_fp16 = conv(bias = layers_14_fc2_bias_to_fp16, dilations = hidden_states_33_dilations_0, groups = hidden_states_33_groups_0, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = hidden_states_33_strides_0, weight = layers_14_fc2_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("hidden_states_33_cast_fp16")]; + tensor inputs_61_cast_fp16 = add(x = inputs_59_cast_fp16, y = hidden_states_33_cast_fp16)[name = tensor("inputs_61_cast_fp16")]; + tensor var_1948 = const()[name = tensor("op_1948"), val = tensor(3)]; + tensor out_61_axes_0 = const()[name = tensor("out_61_axes_0"), val = tensor([1])]; + tensor var_1967_to_fp16 = const()[name = tensor("op_1967_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_61_cast_fp16 = layer_norm(axes = out_61_axes_0, epsilon = var_1967_to_fp16, x = inputs_61_cast_fp16)[name = tensor("out_61_cast_fp16")]; + tensor obj_61_gamma_0_to_fp16 = const()[name = tensor("obj_61_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604963392)))]; + tensor obj_61_beta_0_to_fp16 = const()[name = tensor("obj_61_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604966016)))]; + tensor obj_61_epsilon_0_to_fp16 = const()[name = tensor("obj_61_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_61_cast_fp16 = batch_norm(beta = obj_61_beta_0_to_fp16, epsilon = obj_61_epsilon_0_to_fp16, gamma = obj_61_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_61_cast_fp16)[name = tensor("obj_61_cast_fp16")]; + tensor query_31_pad_type_0 = const()[name = tensor("query_31_pad_type_0"), val = tensor("valid")]; + tensor query_31_strides_0 = const()[name = tensor("query_31_strides_0"), val = tensor([1, 1])]; + tensor query_31_pad_0 = const()[name = tensor("query_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_31_dilations_0 = const()[name = tensor("query_31_dilations_0"), val = tensor([1, 1])]; + tensor query_31_groups_0 = const()[name = tensor("query_31_groups_0"), val = tensor(1)]; + tensor layers_15_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604968640)))]; + tensor layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(608245504)))]; + tensor query_31_cast_fp16 = conv(bias = layers_15_self_attn_q_proj_bias_to_fp16, dilations = query_31_dilations_0, groups = query_31_groups_0, pad = query_31_pad_0, pad_type = query_31_pad_type_0, strides = query_31_strides_0, weight = layers_15_self_attn_q_proj_weight_to_fp16, x = obj_61_cast_fp16)[name = tensor("query_31_cast_fp16")]; + tensor key_31_pad_type_0 = const()[name = tensor("key_31_pad_type_0"), val = tensor("valid")]; + tensor key_31_strides_0 = const()[name = tensor("key_31_strides_0"), val = tensor([1, 1])]; + tensor key_31_pad_0 = const()[name = tensor("key_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_31_dilations_0 = const()[name = tensor("key_31_dilations_0"), val = tensor([1, 1])]; + tensor key_31_groups_0 = const()[name = tensor("key_31_groups_0"), val = tensor(1)]; + tensor layers_15_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(608248128)))]; + tensor key_31_cast_fp16 = conv(dilations = key_31_dilations_0, groups = key_31_groups_0, pad = key_31_pad_0, pad_type = key_31_pad_type_0, strides = key_31_strides_0, weight = layers_15_self_attn_k_proj_weight_to_fp16, x = obj_61_cast_fp16)[name = tensor("key_31_cast_fp16")]; + tensor value_31_pad_type_0 = const()[name = tensor("value_31_pad_type_0"), val = tensor("valid")]; + tensor value_31_strides_0 = const()[name = tensor("value_31_strides_0"), val = tensor([1, 1])]; + tensor value_31_pad_0 = const()[name = tensor("value_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_31_dilations_0 = const()[name = tensor("value_31_dilations_0"), val = tensor([1, 1])]; + tensor value_31_groups_0 = const()[name = tensor("value_31_groups_0"), val = tensor(1)]; + tensor layers_15_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(611524992)))]; + tensor layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614801856)))]; + tensor value_31_cast_fp16 = conv(bias = layers_15_self_attn_v_proj_bias_to_fp16, dilations = value_31_dilations_0, groups = value_31_groups_0, pad = value_31_pad_0, pad_type = value_31_pad_type_0, strides = value_31_strides_0, weight = layers_15_self_attn_v_proj_weight_to_fp16, x = obj_61_cast_fp16)[name = tensor("value_31_cast_fp16")]; + tensor var_2002 = const()[name = tensor("op_2002"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_31_cast_fp16 = reshape(shape = var_2002, x = query_31_cast_fp16)[name = tensor("mh_q_31_cast_fp16")]; + tensor var_2004_to_fp16 = const()[name = tensor("op_2004_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2005_cast_fp16 = mul(x = mh_q_31_cast_fp16, y = var_2004_to_fp16)[name = tensor("op_2005_cast_fp16")]; + tensor var_2006 = const()[name = tensor("op_2006"), val = tensor([1, 20, 64, -1])]; + tensor var_2007_cast_fp16 = reshape(shape = var_2006, x = key_31_cast_fp16)[name = tensor("op_2007_cast_fp16")]; + tensor mh_w_31_transpose_x_0 = const()[name = tensor("mh_w_31_transpose_x_0"), val = tensor(true)]; + tensor mh_w_31_transpose_y_0 = const()[name = tensor("mh_w_31_transpose_y_0"), val = tensor(false)]; + tensor mh_w_31_cast_fp16 = matmul(transpose_x = mh_w_31_transpose_x_0, transpose_y = mh_w_31_transpose_y_0, x = var_2005_cast_fp16, y = var_2007_cast_fp16)[name = tensor("mh_w_31_cast_fp16")]; + tensor var_2010_cast_fp16 = softmax(axis = var_1948, x = mh_w_31_cast_fp16)[name = tensor("op_2010_cast_fp16")]; + tensor var_2011 = const()[name = tensor("op_2011"), val = tensor([1, 20, 64, -1])]; + tensor var_2012_cast_fp16 = reshape(shape = var_2011, x = value_31_cast_fp16)[name = tensor("op_2012_cast_fp16")]; + tensor attn_31_transpose_x_0 = const()[name = tensor("attn_31_transpose_x_0"), val = tensor(false)]; + tensor attn_31_transpose_y_0 = const()[name = tensor("attn_31_transpose_y_0"), val = tensor(true)]; + tensor attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = var_2012_cast_fp16, y = var_2010_cast_fp16)[name = tensor("attn_31_cast_fp16")]; + tensor var_2015 = const()[name = tensor("op_2015"), val = tensor([1, 1280, 1, -1])]; + tensor input_121_cast_fp16 = reshape(shape = var_2015, x = attn_31_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor obj_63_pad_type_0 = const()[name = tensor("obj_63_pad_type_0"), val = tensor("valid")]; + tensor obj_63_strides_0 = const()[name = tensor("obj_63_strides_0"), val = tensor([1, 1])]; + tensor obj_63_pad_0 = const()[name = tensor("obj_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_63_dilations_0 = const()[name = tensor("obj_63_dilations_0"), val = tensor([1, 1])]; + tensor obj_63_groups_0 = const()[name = tensor("obj_63_groups_0"), val = tensor(1)]; + tensor layers_15_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614804480)))]; + tensor layers_15_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(618081344)))]; + tensor obj_63_cast_fp16 = conv(bias = layers_15_self_attn_o_proj_bias_to_fp16, dilations = obj_63_dilations_0, groups = obj_63_groups_0, pad = obj_63_pad_0, pad_type = obj_63_pad_type_0, strides = obj_63_strides_0, weight = layers_15_self_attn_o_proj_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("obj_63_cast_fp16")]; + tensor inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = obj_63_cast_fp16)[name = tensor("inputs_63_cast_fp16")]; + tensor out_63_axes_0 = const()[name = tensor("out_63_axes_0"), val = tensor([1])]; + tensor var_2033_to_fp16 = const()[name = tensor("op_2033_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_63_cast_fp16 = layer_norm(axes = out_63_axes_0, epsilon = var_2033_to_fp16, x = inputs_63_cast_fp16)[name = tensor("out_63_cast_fp16")]; + tensor input_123_gamma_0_to_fp16 = const()[name = tensor("input_123_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(618083968)))]; + tensor input_123_beta_0_to_fp16 = const()[name = tensor("input_123_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(618086592)))]; + tensor input_123_epsilon_0_to_fp16 = const()[name = tensor("input_123_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_123_cast_fp16 = batch_norm(beta = input_123_beta_0_to_fp16, epsilon = input_123_epsilon_0_to_fp16, gamma = input_123_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_63_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor input_125_pad_type_0 = const()[name = tensor("input_125_pad_type_0"), val = tensor("valid")]; + tensor input_125_strides_0 = const()[name = tensor("input_125_strides_0"), val = tensor([1, 1])]; + tensor input_125_pad_0 = const()[name = tensor("input_125_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_125_dilations_0 = const()[name = tensor("input_125_dilations_0"), val = tensor([1, 1])]; + tensor input_125_groups_0 = const()[name = tensor("input_125_groups_0"), val = tensor(1)]; + tensor layers_15_fc1_weight_to_fp16 = const()[name = tensor("layers_15_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(618089216)))]; + tensor layers_15_fc1_bias_to_fp16 = const()[name = tensor("layers_15_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(631196480)))]; + tensor input_125_cast_fp16 = conv(bias = layers_15_fc1_bias_to_fp16, dilations = input_125_dilations_0, groups = input_125_groups_0, pad = input_125_pad_0, pad_type = input_125_pad_type_0, strides = input_125_strides_0, weight = layers_15_fc1_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor input_127_mode_0 = const()[name = tensor("input_127_mode_0"), val = tensor("EXACT")]; + tensor input_127_cast_fp16 = gelu(mode = input_127_mode_0, x = input_125_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor hidden_states_35_pad_type_0 = const()[name = tensor("hidden_states_35_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_35_strides_0 = const()[name = tensor("hidden_states_35_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_35_pad_0 = const()[name = tensor("hidden_states_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_35_dilations_0 = const()[name = tensor("hidden_states_35_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_35_groups_0 = const()[name = tensor("hidden_states_35_groups_0"), val = tensor(1)]; + tensor layers_15_fc2_weight_to_fp16 = const()[name = tensor("layers_15_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(631206784)))]; + tensor layers_15_fc2_bias_to_fp16 = const()[name = tensor("layers_15_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(644314048)))]; + tensor hidden_states_35_cast_fp16 = conv(bias = layers_15_fc2_bias_to_fp16, dilations = hidden_states_35_dilations_0, groups = hidden_states_35_groups_0, pad = hidden_states_35_pad_0, pad_type = hidden_states_35_pad_type_0, strides = hidden_states_35_strides_0, weight = layers_15_fc2_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("hidden_states_35_cast_fp16")]; + tensor inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = hidden_states_35_cast_fp16)[name = tensor("inputs_65_cast_fp16")]; + tensor var_2066 = const()[name = tensor("op_2066"), val = tensor(3)]; + tensor out_65_axes_0 = const()[name = tensor("out_65_axes_0"), val = tensor([1])]; + tensor var_2085_to_fp16 = const()[name = tensor("op_2085_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_65_cast_fp16 = layer_norm(axes = out_65_axes_0, epsilon = var_2085_to_fp16, x = inputs_65_cast_fp16)[name = tensor("out_65_cast_fp16")]; + tensor obj_65_gamma_0_to_fp16 = const()[name = tensor("obj_65_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(644316672)))]; + tensor obj_65_beta_0_to_fp16 = const()[name = tensor("obj_65_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(644319296)))]; + tensor obj_65_epsilon_0_to_fp16 = const()[name = tensor("obj_65_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_65_cast_fp16 = batch_norm(beta = obj_65_beta_0_to_fp16, epsilon = obj_65_epsilon_0_to_fp16, gamma = obj_65_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_65_cast_fp16)[name = tensor("obj_65_cast_fp16")]; + tensor query_33_pad_type_0 = const()[name = tensor("query_33_pad_type_0"), val = tensor("valid")]; + tensor query_33_strides_0 = const()[name = tensor("query_33_strides_0"), val = tensor([1, 1])]; + tensor query_33_pad_0 = const()[name = tensor("query_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_33_dilations_0 = const()[name = tensor("query_33_dilations_0"), val = tensor([1, 1])]; + tensor query_33_groups_0 = const()[name = tensor("query_33_groups_0"), val = tensor(1)]; + tensor layers_16_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(644321920)))]; + tensor layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(647598784)))]; + tensor query_33_cast_fp16 = conv(bias = layers_16_self_attn_q_proj_bias_to_fp16, dilations = query_33_dilations_0, groups = query_33_groups_0, pad = query_33_pad_0, pad_type = query_33_pad_type_0, strides = query_33_strides_0, weight = layers_16_self_attn_q_proj_weight_to_fp16, x = obj_65_cast_fp16)[name = tensor("query_33_cast_fp16")]; + tensor key_33_pad_type_0 = const()[name = tensor("key_33_pad_type_0"), val = tensor("valid")]; + tensor key_33_strides_0 = const()[name = tensor("key_33_strides_0"), val = tensor([1, 1])]; + tensor key_33_pad_0 = const()[name = tensor("key_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_33_dilations_0 = const()[name = tensor("key_33_dilations_0"), val = tensor([1, 1])]; + tensor key_33_groups_0 = const()[name = tensor("key_33_groups_0"), val = tensor(1)]; + tensor layers_16_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(647601408)))]; + tensor key_33_cast_fp16 = conv(dilations = key_33_dilations_0, groups = key_33_groups_0, pad = key_33_pad_0, pad_type = key_33_pad_type_0, strides = key_33_strides_0, weight = layers_16_self_attn_k_proj_weight_to_fp16, x = obj_65_cast_fp16)[name = tensor("key_33_cast_fp16")]; + tensor value_33_pad_type_0 = const()[name = tensor("value_33_pad_type_0"), val = tensor("valid")]; + tensor value_33_strides_0 = const()[name = tensor("value_33_strides_0"), val = tensor([1, 1])]; + tensor value_33_pad_0 = const()[name = tensor("value_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_33_dilations_0 = const()[name = tensor("value_33_dilations_0"), val = tensor([1, 1])]; + tensor value_33_groups_0 = const()[name = tensor("value_33_groups_0"), val = tensor(1)]; + tensor layers_16_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(650878272)))]; + tensor layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(654155136)))]; + tensor value_33_cast_fp16 = conv(bias = layers_16_self_attn_v_proj_bias_to_fp16, dilations = value_33_dilations_0, groups = value_33_groups_0, pad = value_33_pad_0, pad_type = value_33_pad_type_0, strides = value_33_strides_0, weight = layers_16_self_attn_v_proj_weight_to_fp16, x = obj_65_cast_fp16)[name = tensor("value_33_cast_fp16")]; + tensor var_2120 = const()[name = tensor("op_2120"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_33_cast_fp16 = reshape(shape = var_2120, x = query_33_cast_fp16)[name = tensor("mh_q_33_cast_fp16")]; + tensor var_2122_to_fp16 = const()[name = tensor("op_2122_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2123_cast_fp16 = mul(x = mh_q_33_cast_fp16, y = var_2122_to_fp16)[name = tensor("op_2123_cast_fp16")]; + tensor var_2124 = const()[name = tensor("op_2124"), val = tensor([1, 20, 64, -1])]; + tensor var_2125_cast_fp16 = reshape(shape = var_2124, x = key_33_cast_fp16)[name = tensor("op_2125_cast_fp16")]; + tensor mh_w_33_transpose_x_0 = const()[name = tensor("mh_w_33_transpose_x_0"), val = tensor(true)]; + tensor mh_w_33_transpose_y_0 = const()[name = tensor("mh_w_33_transpose_y_0"), val = tensor(false)]; + tensor mh_w_33_cast_fp16 = matmul(transpose_x = mh_w_33_transpose_x_0, transpose_y = mh_w_33_transpose_y_0, x = var_2123_cast_fp16, y = var_2125_cast_fp16)[name = tensor("mh_w_33_cast_fp16")]; + tensor var_2128_cast_fp16 = softmax(axis = var_2066, x = mh_w_33_cast_fp16)[name = tensor("op_2128_cast_fp16")]; + tensor var_2129 = const()[name = tensor("op_2129"), val = tensor([1, 20, 64, -1])]; + tensor var_2130_cast_fp16 = reshape(shape = var_2129, x = value_33_cast_fp16)[name = tensor("op_2130_cast_fp16")]; + tensor attn_33_transpose_x_0 = const()[name = tensor("attn_33_transpose_x_0"), val = tensor(false)]; + tensor attn_33_transpose_y_0 = const()[name = tensor("attn_33_transpose_y_0"), val = tensor(true)]; + tensor attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = var_2130_cast_fp16, y = var_2128_cast_fp16)[name = tensor("attn_33_cast_fp16")]; + tensor var_2133 = const()[name = tensor("op_2133"), val = tensor([1, 1280, 1, -1])]; + tensor input_129_cast_fp16 = reshape(shape = var_2133, x = attn_33_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor obj_67_pad_type_0 = const()[name = tensor("obj_67_pad_type_0"), val = tensor("valid")]; + tensor obj_67_strides_0 = const()[name = tensor("obj_67_strides_0"), val = tensor([1, 1])]; + tensor obj_67_pad_0 = const()[name = tensor("obj_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_67_dilations_0 = const()[name = tensor("obj_67_dilations_0"), val = tensor([1, 1])]; + tensor obj_67_groups_0 = const()[name = tensor("obj_67_groups_0"), val = tensor(1)]; + tensor layers_16_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(654157760)))]; + tensor layers_16_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657434624)))]; + tensor obj_67_cast_fp16 = conv(bias = layers_16_self_attn_o_proj_bias_to_fp16, dilations = obj_67_dilations_0, groups = obj_67_groups_0, pad = obj_67_pad_0, pad_type = obj_67_pad_type_0, strides = obj_67_strides_0, weight = layers_16_self_attn_o_proj_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("obj_67_cast_fp16")]; + tensor inputs_67_cast_fp16 = add(x = inputs_65_cast_fp16, y = obj_67_cast_fp16)[name = tensor("inputs_67_cast_fp16")]; + tensor out_67_axes_0 = const()[name = tensor("out_67_axes_0"), val = tensor([1])]; + tensor var_2151_to_fp16 = const()[name = tensor("op_2151_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_67_cast_fp16 = layer_norm(axes = out_67_axes_0, epsilon = var_2151_to_fp16, x = inputs_67_cast_fp16)[name = tensor("out_67_cast_fp16")]; + tensor input_131_gamma_0_to_fp16 = const()[name = tensor("input_131_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657437248)))]; + tensor input_131_beta_0_to_fp16 = const()[name = tensor("input_131_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657439872)))]; + tensor input_131_epsilon_0_to_fp16 = const()[name = tensor("input_131_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_131_cast_fp16 = batch_norm(beta = input_131_beta_0_to_fp16, epsilon = input_131_epsilon_0_to_fp16, gamma = input_131_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_67_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor input_133_pad_type_0 = const()[name = tensor("input_133_pad_type_0"), val = tensor("valid")]; + tensor input_133_strides_0 = const()[name = tensor("input_133_strides_0"), val = tensor([1, 1])]; + tensor input_133_pad_0 = const()[name = tensor("input_133_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_133_dilations_0 = const()[name = tensor("input_133_dilations_0"), val = tensor([1, 1])]; + tensor input_133_groups_0 = const()[name = tensor("input_133_groups_0"), val = tensor(1)]; + tensor layers_16_fc1_weight_to_fp16 = const()[name = tensor("layers_16_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657442496)))]; + tensor layers_16_fc1_bias_to_fp16 = const()[name = tensor("layers_16_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(670549760)))]; + tensor input_133_cast_fp16 = conv(bias = layers_16_fc1_bias_to_fp16, dilations = input_133_dilations_0, groups = input_133_groups_0, pad = input_133_pad_0, pad_type = input_133_pad_type_0, strides = input_133_strides_0, weight = layers_16_fc1_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor input_135_mode_0 = const()[name = tensor("input_135_mode_0"), val = tensor("EXACT")]; + tensor input_135_cast_fp16 = gelu(mode = input_135_mode_0, x = input_133_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor hidden_states_37_pad_type_0 = const()[name = tensor("hidden_states_37_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_37_strides_0 = const()[name = tensor("hidden_states_37_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_37_pad_0 = const()[name = tensor("hidden_states_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_37_dilations_0 = const()[name = tensor("hidden_states_37_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_37_groups_0 = const()[name = tensor("hidden_states_37_groups_0"), val = tensor(1)]; + tensor layers_16_fc2_weight_to_fp16 = const()[name = tensor("layers_16_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(670560064)))]; + tensor layers_16_fc2_bias_to_fp16 = const()[name = tensor("layers_16_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683667328)))]; + tensor hidden_states_37_cast_fp16 = conv(bias = layers_16_fc2_bias_to_fp16, dilations = hidden_states_37_dilations_0, groups = hidden_states_37_groups_0, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = hidden_states_37_strides_0, weight = layers_16_fc2_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("hidden_states_37_cast_fp16")]; + tensor inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = hidden_states_37_cast_fp16)[name = tensor("inputs_69_cast_fp16")]; + tensor var_2184 = const()[name = tensor("op_2184"), val = tensor(3)]; + tensor out_69_axes_0 = const()[name = tensor("out_69_axes_0"), val = tensor([1])]; + tensor var_2203_to_fp16 = const()[name = tensor("op_2203_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_69_cast_fp16 = layer_norm(axes = out_69_axes_0, epsilon = var_2203_to_fp16, x = inputs_69_cast_fp16)[name = tensor("out_69_cast_fp16")]; + tensor obj_69_gamma_0_to_fp16 = const()[name = tensor("obj_69_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683669952)))]; + tensor obj_69_beta_0_to_fp16 = const()[name = tensor("obj_69_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683672576)))]; + tensor obj_69_epsilon_0_to_fp16 = const()[name = tensor("obj_69_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_69_cast_fp16 = batch_norm(beta = obj_69_beta_0_to_fp16, epsilon = obj_69_epsilon_0_to_fp16, gamma = obj_69_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_69_cast_fp16)[name = tensor("obj_69_cast_fp16")]; + tensor query_35_pad_type_0 = const()[name = tensor("query_35_pad_type_0"), val = tensor("valid")]; + tensor query_35_strides_0 = const()[name = tensor("query_35_strides_0"), val = tensor([1, 1])]; + tensor query_35_pad_0 = const()[name = tensor("query_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_35_dilations_0 = const()[name = tensor("query_35_dilations_0"), val = tensor([1, 1])]; + tensor query_35_groups_0 = const()[name = tensor("query_35_groups_0"), val = tensor(1)]; + tensor layers_17_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683675200)))]; + tensor layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(686952064)))]; + tensor query_35_cast_fp16 = conv(bias = layers_17_self_attn_q_proj_bias_to_fp16, dilations = query_35_dilations_0, groups = query_35_groups_0, pad = query_35_pad_0, pad_type = query_35_pad_type_0, strides = query_35_strides_0, weight = layers_17_self_attn_q_proj_weight_to_fp16, x = obj_69_cast_fp16)[name = tensor("query_35_cast_fp16")]; + tensor key_35_pad_type_0 = const()[name = tensor("key_35_pad_type_0"), val = tensor("valid")]; + tensor key_35_strides_0 = const()[name = tensor("key_35_strides_0"), val = tensor([1, 1])]; + tensor key_35_pad_0 = const()[name = tensor("key_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_35_dilations_0 = const()[name = tensor("key_35_dilations_0"), val = tensor([1, 1])]; + tensor key_35_groups_0 = const()[name = tensor("key_35_groups_0"), val = tensor(1)]; + tensor layers_17_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(686954688)))]; + tensor key_35_cast_fp16 = conv(dilations = key_35_dilations_0, groups = key_35_groups_0, pad = key_35_pad_0, pad_type = key_35_pad_type_0, strides = key_35_strides_0, weight = layers_17_self_attn_k_proj_weight_to_fp16, x = obj_69_cast_fp16)[name = tensor("key_35_cast_fp16")]; + tensor value_35_pad_type_0 = const()[name = tensor("value_35_pad_type_0"), val = tensor("valid")]; + tensor value_35_strides_0 = const()[name = tensor("value_35_strides_0"), val = tensor([1, 1])]; + tensor value_35_pad_0 = const()[name = tensor("value_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_35_dilations_0 = const()[name = tensor("value_35_dilations_0"), val = tensor([1, 1])]; + tensor value_35_groups_0 = const()[name = tensor("value_35_groups_0"), val = tensor(1)]; + tensor layers_17_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(690231552)))]; + tensor layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(693508416)))]; + tensor value_35_cast_fp16 = conv(bias = layers_17_self_attn_v_proj_bias_to_fp16, dilations = value_35_dilations_0, groups = value_35_groups_0, pad = value_35_pad_0, pad_type = value_35_pad_type_0, strides = value_35_strides_0, weight = layers_17_self_attn_v_proj_weight_to_fp16, x = obj_69_cast_fp16)[name = tensor("value_35_cast_fp16")]; + tensor var_2238 = const()[name = tensor("op_2238"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_35_cast_fp16 = reshape(shape = var_2238, x = query_35_cast_fp16)[name = tensor("mh_q_35_cast_fp16")]; + tensor var_2240_to_fp16 = const()[name = tensor("op_2240_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2241_cast_fp16 = mul(x = mh_q_35_cast_fp16, y = var_2240_to_fp16)[name = tensor("op_2241_cast_fp16")]; + tensor var_2242 = const()[name = tensor("op_2242"), val = tensor([1, 20, 64, -1])]; + tensor var_2243_cast_fp16 = reshape(shape = var_2242, x = key_35_cast_fp16)[name = tensor("op_2243_cast_fp16")]; + tensor mh_w_35_transpose_x_0 = const()[name = tensor("mh_w_35_transpose_x_0"), val = tensor(true)]; + tensor mh_w_35_transpose_y_0 = const()[name = tensor("mh_w_35_transpose_y_0"), val = tensor(false)]; + tensor mh_w_35_cast_fp16 = matmul(transpose_x = mh_w_35_transpose_x_0, transpose_y = mh_w_35_transpose_y_0, x = var_2241_cast_fp16, y = var_2243_cast_fp16)[name = tensor("mh_w_35_cast_fp16")]; + tensor var_2246_cast_fp16 = softmax(axis = var_2184, x = mh_w_35_cast_fp16)[name = tensor("op_2246_cast_fp16")]; + tensor var_2247 = const()[name = tensor("op_2247"), val = tensor([1, 20, 64, -1])]; + tensor var_2248_cast_fp16 = reshape(shape = var_2247, x = value_35_cast_fp16)[name = tensor("op_2248_cast_fp16")]; + tensor attn_35_transpose_x_0 = const()[name = tensor("attn_35_transpose_x_0"), val = tensor(false)]; + tensor attn_35_transpose_y_0 = const()[name = tensor("attn_35_transpose_y_0"), val = tensor(true)]; + tensor attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = var_2248_cast_fp16, y = var_2246_cast_fp16)[name = tensor("attn_35_cast_fp16")]; + tensor var_2251 = const()[name = tensor("op_2251"), val = tensor([1, 1280, 1, -1])]; + tensor input_137_cast_fp16 = reshape(shape = var_2251, x = attn_35_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor obj_71_pad_type_0 = const()[name = tensor("obj_71_pad_type_0"), val = tensor("valid")]; + tensor obj_71_strides_0 = const()[name = tensor("obj_71_strides_0"), val = tensor([1, 1])]; + tensor obj_71_pad_0 = const()[name = tensor("obj_71_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_71_dilations_0 = const()[name = tensor("obj_71_dilations_0"), val = tensor([1, 1])]; + tensor obj_71_groups_0 = const()[name = tensor("obj_71_groups_0"), val = tensor(1)]; + tensor layers_17_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(693511040)))]; + tensor layers_17_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696787904)))]; + tensor obj_71_cast_fp16 = conv(bias = layers_17_self_attn_o_proj_bias_to_fp16, dilations = obj_71_dilations_0, groups = obj_71_groups_0, pad = obj_71_pad_0, pad_type = obj_71_pad_type_0, strides = obj_71_strides_0, weight = layers_17_self_attn_o_proj_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("obj_71_cast_fp16")]; + tensor inputs_71_cast_fp16 = add(x = inputs_69_cast_fp16, y = obj_71_cast_fp16)[name = tensor("inputs_71_cast_fp16")]; + tensor out_71_axes_0 = const()[name = tensor("out_71_axes_0"), val = tensor([1])]; + tensor var_2269_to_fp16 = const()[name = tensor("op_2269_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_71_cast_fp16 = layer_norm(axes = out_71_axes_0, epsilon = var_2269_to_fp16, x = inputs_71_cast_fp16)[name = tensor("out_71_cast_fp16")]; + tensor input_139_gamma_0_to_fp16 = const()[name = tensor("input_139_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696790528)))]; + tensor input_139_beta_0_to_fp16 = const()[name = tensor("input_139_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696793152)))]; + tensor input_139_epsilon_0_to_fp16 = const()[name = tensor("input_139_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_139_cast_fp16 = batch_norm(beta = input_139_beta_0_to_fp16, epsilon = input_139_epsilon_0_to_fp16, gamma = input_139_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_71_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor input_141_pad_type_0 = const()[name = tensor("input_141_pad_type_0"), val = tensor("valid")]; + tensor input_141_strides_0 = const()[name = tensor("input_141_strides_0"), val = tensor([1, 1])]; + tensor input_141_pad_0 = const()[name = tensor("input_141_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_141_dilations_0 = const()[name = tensor("input_141_dilations_0"), val = tensor([1, 1])]; + tensor input_141_groups_0 = const()[name = tensor("input_141_groups_0"), val = tensor(1)]; + tensor layers_17_fc1_weight_to_fp16 = const()[name = tensor("layers_17_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696795776)))]; + tensor layers_17_fc1_bias_to_fp16 = const()[name = tensor("layers_17_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(709903040)))]; + tensor input_141_cast_fp16 = conv(bias = layers_17_fc1_bias_to_fp16, dilations = input_141_dilations_0, groups = input_141_groups_0, pad = input_141_pad_0, pad_type = input_141_pad_type_0, strides = input_141_strides_0, weight = layers_17_fc1_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor input_143_mode_0 = const()[name = tensor("input_143_mode_0"), val = tensor("EXACT")]; + tensor input_143_cast_fp16 = gelu(mode = input_143_mode_0, x = input_141_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor hidden_states_39_pad_type_0 = const()[name = tensor("hidden_states_39_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_39_strides_0 = const()[name = tensor("hidden_states_39_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_39_pad_0 = const()[name = tensor("hidden_states_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_39_dilations_0 = const()[name = tensor("hidden_states_39_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_39_groups_0 = const()[name = tensor("hidden_states_39_groups_0"), val = tensor(1)]; + tensor layers_17_fc2_weight_to_fp16 = const()[name = tensor("layers_17_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(709913344)))]; + tensor layers_17_fc2_bias_to_fp16 = const()[name = tensor("layers_17_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(723020608)))]; + tensor hidden_states_39_cast_fp16 = conv(bias = layers_17_fc2_bias_to_fp16, dilations = hidden_states_39_dilations_0, groups = hidden_states_39_groups_0, pad = hidden_states_39_pad_0, pad_type = hidden_states_39_pad_type_0, strides = hidden_states_39_strides_0, weight = layers_17_fc2_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("hidden_states_39_cast_fp16")]; + tensor inputs_73_cast_fp16 = add(x = inputs_71_cast_fp16, y = hidden_states_39_cast_fp16)[name = tensor("inputs_73_cast_fp16")]; + tensor var_2302 = const()[name = tensor("op_2302"), val = tensor(3)]; + tensor out_73_axes_0 = const()[name = tensor("out_73_axes_0"), val = tensor([1])]; + tensor var_2321_to_fp16 = const()[name = tensor("op_2321_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_73_cast_fp16 = layer_norm(axes = out_73_axes_0, epsilon = var_2321_to_fp16, x = inputs_73_cast_fp16)[name = tensor("out_73_cast_fp16")]; + tensor obj_73_gamma_0_to_fp16 = const()[name = tensor("obj_73_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(723023232)))]; + tensor obj_73_beta_0_to_fp16 = const()[name = tensor("obj_73_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(723025856)))]; + tensor obj_73_epsilon_0_to_fp16 = const()[name = tensor("obj_73_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_73_cast_fp16 = batch_norm(beta = obj_73_beta_0_to_fp16, epsilon = obj_73_epsilon_0_to_fp16, gamma = obj_73_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_73_cast_fp16)[name = tensor("obj_73_cast_fp16")]; + tensor query_37_pad_type_0 = const()[name = tensor("query_37_pad_type_0"), val = tensor("valid")]; + tensor query_37_strides_0 = const()[name = tensor("query_37_strides_0"), val = tensor([1, 1])]; + tensor query_37_pad_0 = const()[name = tensor("query_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_37_dilations_0 = const()[name = tensor("query_37_dilations_0"), val = tensor([1, 1])]; + tensor query_37_groups_0 = const()[name = tensor("query_37_groups_0"), val = tensor(1)]; + tensor layers_18_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(723028480)))]; + tensor layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(726305344)))]; + tensor query_37_cast_fp16 = conv(bias = layers_18_self_attn_q_proj_bias_to_fp16, dilations = query_37_dilations_0, groups = query_37_groups_0, pad = query_37_pad_0, pad_type = query_37_pad_type_0, strides = query_37_strides_0, weight = layers_18_self_attn_q_proj_weight_to_fp16, x = obj_73_cast_fp16)[name = tensor("query_37_cast_fp16")]; + tensor key_37_pad_type_0 = const()[name = tensor("key_37_pad_type_0"), val = tensor("valid")]; + tensor key_37_strides_0 = const()[name = tensor("key_37_strides_0"), val = tensor([1, 1])]; + tensor key_37_pad_0 = const()[name = tensor("key_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_37_dilations_0 = const()[name = tensor("key_37_dilations_0"), val = tensor([1, 1])]; + tensor key_37_groups_0 = const()[name = tensor("key_37_groups_0"), val = tensor(1)]; + tensor layers_18_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(726307968)))]; + tensor key_37_cast_fp16 = conv(dilations = key_37_dilations_0, groups = key_37_groups_0, pad = key_37_pad_0, pad_type = key_37_pad_type_0, strides = key_37_strides_0, weight = layers_18_self_attn_k_proj_weight_to_fp16, x = obj_73_cast_fp16)[name = tensor("key_37_cast_fp16")]; + tensor value_37_pad_type_0 = const()[name = tensor("value_37_pad_type_0"), val = tensor("valid")]; + tensor value_37_strides_0 = const()[name = tensor("value_37_strides_0"), val = tensor([1, 1])]; + tensor value_37_pad_0 = const()[name = tensor("value_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_37_dilations_0 = const()[name = tensor("value_37_dilations_0"), val = tensor([1, 1])]; + tensor value_37_groups_0 = const()[name = tensor("value_37_groups_0"), val = tensor(1)]; + tensor layers_18_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(729584832)))]; + tensor layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(732861696)))]; + tensor value_37_cast_fp16 = conv(bias = layers_18_self_attn_v_proj_bias_to_fp16, dilations = value_37_dilations_0, groups = value_37_groups_0, pad = value_37_pad_0, pad_type = value_37_pad_type_0, strides = value_37_strides_0, weight = layers_18_self_attn_v_proj_weight_to_fp16, x = obj_73_cast_fp16)[name = tensor("value_37_cast_fp16")]; + tensor var_2356 = const()[name = tensor("op_2356"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_37_cast_fp16 = reshape(shape = var_2356, x = query_37_cast_fp16)[name = tensor("mh_q_37_cast_fp16")]; + tensor var_2358_to_fp16 = const()[name = tensor("op_2358_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2359_cast_fp16 = mul(x = mh_q_37_cast_fp16, y = var_2358_to_fp16)[name = tensor("op_2359_cast_fp16")]; + tensor var_2360 = const()[name = tensor("op_2360"), val = tensor([1, 20, 64, -1])]; + tensor var_2361_cast_fp16 = reshape(shape = var_2360, x = key_37_cast_fp16)[name = tensor("op_2361_cast_fp16")]; + tensor mh_w_37_transpose_x_0 = const()[name = tensor("mh_w_37_transpose_x_0"), val = tensor(true)]; + tensor mh_w_37_transpose_y_0 = const()[name = tensor("mh_w_37_transpose_y_0"), val = tensor(false)]; + tensor mh_w_37_cast_fp16 = matmul(transpose_x = mh_w_37_transpose_x_0, transpose_y = mh_w_37_transpose_y_0, x = var_2359_cast_fp16, y = var_2361_cast_fp16)[name = tensor("mh_w_37_cast_fp16")]; + tensor var_2364_cast_fp16 = softmax(axis = var_2302, x = mh_w_37_cast_fp16)[name = tensor("op_2364_cast_fp16")]; + tensor var_2365 = const()[name = tensor("op_2365"), val = tensor([1, 20, 64, -1])]; + tensor var_2366_cast_fp16 = reshape(shape = var_2365, x = value_37_cast_fp16)[name = tensor("op_2366_cast_fp16")]; + tensor attn_37_transpose_x_0 = const()[name = tensor("attn_37_transpose_x_0"), val = tensor(false)]; + tensor attn_37_transpose_y_0 = const()[name = tensor("attn_37_transpose_y_0"), val = tensor(true)]; + tensor attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = var_2366_cast_fp16, y = var_2364_cast_fp16)[name = tensor("attn_37_cast_fp16")]; + tensor var_2369 = const()[name = tensor("op_2369"), val = tensor([1, 1280, 1, -1])]; + tensor input_145_cast_fp16 = reshape(shape = var_2369, x = attn_37_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor obj_75_pad_type_0 = const()[name = tensor("obj_75_pad_type_0"), val = tensor("valid")]; + tensor obj_75_strides_0 = const()[name = tensor("obj_75_strides_0"), val = tensor([1, 1])]; + tensor obj_75_pad_0 = const()[name = tensor("obj_75_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_75_dilations_0 = const()[name = tensor("obj_75_dilations_0"), val = tensor([1, 1])]; + tensor obj_75_groups_0 = const()[name = tensor("obj_75_groups_0"), val = tensor(1)]; + tensor layers_18_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(732864320)))]; + tensor layers_18_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(736141184)))]; + tensor obj_75_cast_fp16 = conv(bias = layers_18_self_attn_o_proj_bias_to_fp16, dilations = obj_75_dilations_0, groups = obj_75_groups_0, pad = obj_75_pad_0, pad_type = obj_75_pad_type_0, strides = obj_75_strides_0, weight = layers_18_self_attn_o_proj_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("obj_75_cast_fp16")]; + tensor inputs_75_cast_fp16 = add(x = inputs_73_cast_fp16, y = obj_75_cast_fp16)[name = tensor("inputs_75_cast_fp16")]; + tensor out_75_axes_0 = const()[name = tensor("out_75_axes_0"), val = tensor([1])]; + tensor var_2387_to_fp16 = const()[name = tensor("op_2387_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_75_cast_fp16 = layer_norm(axes = out_75_axes_0, epsilon = var_2387_to_fp16, x = inputs_75_cast_fp16)[name = tensor("out_75_cast_fp16")]; + tensor input_147_gamma_0_to_fp16 = const()[name = tensor("input_147_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(736143808)))]; + tensor input_147_beta_0_to_fp16 = const()[name = tensor("input_147_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(736146432)))]; + tensor input_147_epsilon_0_to_fp16 = const()[name = tensor("input_147_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_147_cast_fp16 = batch_norm(beta = input_147_beta_0_to_fp16, epsilon = input_147_epsilon_0_to_fp16, gamma = input_147_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_75_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1, 1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1, 1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor layers_18_fc1_weight_to_fp16 = const()[name = tensor("layers_18_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(736149056)))]; + tensor layers_18_fc1_bias_to_fp16 = const()[name = tensor("layers_18_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(749256320)))]; + tensor input_149_cast_fp16 = conv(bias = layers_18_fc1_bias_to_fp16, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = layers_18_fc1_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor input_151_mode_0 = const()[name = tensor("input_151_mode_0"), val = tensor("EXACT")]; + tensor input_151_cast_fp16 = gelu(mode = input_151_mode_0, x = input_149_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor hidden_states_41_pad_type_0 = const()[name = tensor("hidden_states_41_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_41_strides_0 = const()[name = tensor("hidden_states_41_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_41_pad_0 = const()[name = tensor("hidden_states_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_41_dilations_0 = const()[name = tensor("hidden_states_41_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_41_groups_0 = const()[name = tensor("hidden_states_41_groups_0"), val = tensor(1)]; + tensor layers_18_fc2_weight_to_fp16 = const()[name = tensor("layers_18_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(749266624)))]; + tensor layers_18_fc2_bias_to_fp16 = const()[name = tensor("layers_18_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762373888)))]; + tensor hidden_states_41_cast_fp16 = conv(bias = layers_18_fc2_bias_to_fp16, dilations = hidden_states_41_dilations_0, groups = hidden_states_41_groups_0, pad = hidden_states_41_pad_0, pad_type = hidden_states_41_pad_type_0, strides = hidden_states_41_strides_0, weight = layers_18_fc2_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("hidden_states_41_cast_fp16")]; + tensor inputs_77_cast_fp16 = add(x = inputs_75_cast_fp16, y = hidden_states_41_cast_fp16)[name = tensor("inputs_77_cast_fp16")]; + tensor var_2420 = const()[name = tensor("op_2420"), val = tensor(3)]; + tensor out_77_axes_0 = const()[name = tensor("out_77_axes_0"), val = tensor([1])]; + tensor var_2439_to_fp16 = const()[name = tensor("op_2439_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_77_cast_fp16 = layer_norm(axes = out_77_axes_0, epsilon = var_2439_to_fp16, x = inputs_77_cast_fp16)[name = tensor("out_77_cast_fp16")]; + tensor obj_77_gamma_0_to_fp16 = const()[name = tensor("obj_77_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762376512)))]; + tensor obj_77_beta_0_to_fp16 = const()[name = tensor("obj_77_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762379136)))]; + tensor obj_77_epsilon_0_to_fp16 = const()[name = tensor("obj_77_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_77_cast_fp16 = batch_norm(beta = obj_77_beta_0_to_fp16, epsilon = obj_77_epsilon_0_to_fp16, gamma = obj_77_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_77_cast_fp16)[name = tensor("obj_77_cast_fp16")]; + tensor query_39_pad_type_0 = const()[name = tensor("query_39_pad_type_0"), val = tensor("valid")]; + tensor query_39_strides_0 = const()[name = tensor("query_39_strides_0"), val = tensor([1, 1])]; + tensor query_39_pad_0 = const()[name = tensor("query_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_39_dilations_0 = const()[name = tensor("query_39_dilations_0"), val = tensor([1, 1])]; + tensor query_39_groups_0 = const()[name = tensor("query_39_groups_0"), val = tensor(1)]; + tensor layers_19_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762381760)))]; + tensor layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(765658624)))]; + tensor query_39_cast_fp16 = conv(bias = layers_19_self_attn_q_proj_bias_to_fp16, dilations = query_39_dilations_0, groups = query_39_groups_0, pad = query_39_pad_0, pad_type = query_39_pad_type_0, strides = query_39_strides_0, weight = layers_19_self_attn_q_proj_weight_to_fp16, x = obj_77_cast_fp16)[name = tensor("query_39_cast_fp16")]; + tensor key_39_pad_type_0 = const()[name = tensor("key_39_pad_type_0"), val = tensor("valid")]; + tensor key_39_strides_0 = const()[name = tensor("key_39_strides_0"), val = tensor([1, 1])]; + tensor key_39_pad_0 = const()[name = tensor("key_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_39_dilations_0 = const()[name = tensor("key_39_dilations_0"), val = tensor([1, 1])]; + tensor key_39_groups_0 = const()[name = tensor("key_39_groups_0"), val = tensor(1)]; + tensor layers_19_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(765661248)))]; + tensor key_39_cast_fp16 = conv(dilations = key_39_dilations_0, groups = key_39_groups_0, pad = key_39_pad_0, pad_type = key_39_pad_type_0, strides = key_39_strides_0, weight = layers_19_self_attn_k_proj_weight_to_fp16, x = obj_77_cast_fp16)[name = tensor("key_39_cast_fp16")]; + tensor value_39_pad_type_0 = const()[name = tensor("value_39_pad_type_0"), val = tensor("valid")]; + tensor value_39_strides_0 = const()[name = tensor("value_39_strides_0"), val = tensor([1, 1])]; + tensor value_39_pad_0 = const()[name = tensor("value_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_39_dilations_0 = const()[name = tensor("value_39_dilations_0"), val = tensor([1, 1])]; + tensor value_39_groups_0 = const()[name = tensor("value_39_groups_0"), val = tensor(1)]; + tensor layers_19_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(768938112)))]; + tensor layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(772214976)))]; + tensor value_39_cast_fp16 = conv(bias = layers_19_self_attn_v_proj_bias_to_fp16, dilations = value_39_dilations_0, groups = value_39_groups_0, pad = value_39_pad_0, pad_type = value_39_pad_type_0, strides = value_39_strides_0, weight = layers_19_self_attn_v_proj_weight_to_fp16, x = obj_77_cast_fp16)[name = tensor("value_39_cast_fp16")]; + tensor var_2474 = const()[name = tensor("op_2474"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_39_cast_fp16 = reshape(shape = var_2474, x = query_39_cast_fp16)[name = tensor("mh_q_39_cast_fp16")]; + tensor var_2476_to_fp16 = const()[name = tensor("op_2476_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2477_cast_fp16 = mul(x = mh_q_39_cast_fp16, y = var_2476_to_fp16)[name = tensor("op_2477_cast_fp16")]; + tensor var_2478 = const()[name = tensor("op_2478"), val = tensor([1, 20, 64, -1])]; + tensor var_2479_cast_fp16 = reshape(shape = var_2478, x = key_39_cast_fp16)[name = tensor("op_2479_cast_fp16")]; + tensor mh_w_39_transpose_x_0 = const()[name = tensor("mh_w_39_transpose_x_0"), val = tensor(true)]; + tensor mh_w_39_transpose_y_0 = const()[name = tensor("mh_w_39_transpose_y_0"), val = tensor(false)]; + tensor mh_w_39_cast_fp16 = matmul(transpose_x = mh_w_39_transpose_x_0, transpose_y = mh_w_39_transpose_y_0, x = var_2477_cast_fp16, y = var_2479_cast_fp16)[name = tensor("mh_w_39_cast_fp16")]; + tensor var_2482_cast_fp16 = softmax(axis = var_2420, x = mh_w_39_cast_fp16)[name = tensor("op_2482_cast_fp16")]; + tensor var_2483 = const()[name = tensor("op_2483"), val = tensor([1, 20, 64, -1])]; + tensor var_2484_cast_fp16 = reshape(shape = var_2483, x = value_39_cast_fp16)[name = tensor("op_2484_cast_fp16")]; + tensor attn_39_transpose_x_0 = const()[name = tensor("attn_39_transpose_x_0"), val = tensor(false)]; + tensor attn_39_transpose_y_0 = const()[name = tensor("attn_39_transpose_y_0"), val = tensor(true)]; + tensor attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = var_2484_cast_fp16, y = var_2482_cast_fp16)[name = tensor("attn_39_cast_fp16")]; + tensor var_2487 = const()[name = tensor("op_2487"), val = tensor([1, 1280, 1, -1])]; + tensor input_153_cast_fp16 = reshape(shape = var_2487, x = attn_39_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor obj_79_pad_type_0 = const()[name = tensor("obj_79_pad_type_0"), val = tensor("valid")]; + tensor obj_79_strides_0 = const()[name = tensor("obj_79_strides_0"), val = tensor([1, 1])]; + tensor obj_79_pad_0 = const()[name = tensor("obj_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_79_dilations_0 = const()[name = tensor("obj_79_dilations_0"), val = tensor([1, 1])]; + tensor obj_79_groups_0 = const()[name = tensor("obj_79_groups_0"), val = tensor(1)]; + tensor layers_19_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(772217600)))]; + tensor layers_19_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775494464)))]; + tensor obj_79_cast_fp16 = conv(bias = layers_19_self_attn_o_proj_bias_to_fp16, dilations = obj_79_dilations_0, groups = obj_79_groups_0, pad = obj_79_pad_0, pad_type = obj_79_pad_type_0, strides = obj_79_strides_0, weight = layers_19_self_attn_o_proj_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("obj_79_cast_fp16")]; + tensor inputs_79_cast_fp16 = add(x = inputs_77_cast_fp16, y = obj_79_cast_fp16)[name = tensor("inputs_79_cast_fp16")]; + tensor out_79_axes_0 = const()[name = tensor("out_79_axes_0"), val = tensor([1])]; + tensor var_2505_to_fp16 = const()[name = tensor("op_2505_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_79_cast_fp16 = layer_norm(axes = out_79_axes_0, epsilon = var_2505_to_fp16, x = inputs_79_cast_fp16)[name = tensor("out_79_cast_fp16")]; + tensor input_155_gamma_0_to_fp16 = const()[name = tensor("input_155_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775497088)))]; + tensor input_155_beta_0_to_fp16 = const()[name = tensor("input_155_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775499712)))]; + tensor input_155_epsilon_0_to_fp16 = const()[name = tensor("input_155_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_155_cast_fp16 = batch_norm(beta = input_155_beta_0_to_fp16, epsilon = input_155_epsilon_0_to_fp16, gamma = input_155_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_79_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor input_157_pad_type_0 = const()[name = tensor("input_157_pad_type_0"), val = tensor("valid")]; + tensor input_157_strides_0 = const()[name = tensor("input_157_strides_0"), val = tensor([1, 1])]; + tensor input_157_pad_0 = const()[name = tensor("input_157_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_157_dilations_0 = const()[name = tensor("input_157_dilations_0"), val = tensor([1, 1])]; + tensor input_157_groups_0 = const()[name = tensor("input_157_groups_0"), val = tensor(1)]; + tensor layers_19_fc1_weight_to_fp16 = const()[name = tensor("layers_19_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775502336)))]; + tensor layers_19_fc1_bias_to_fp16 = const()[name = tensor("layers_19_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(788609600)))]; + tensor input_157_cast_fp16 = conv(bias = layers_19_fc1_bias_to_fp16, dilations = input_157_dilations_0, groups = input_157_groups_0, pad = input_157_pad_0, pad_type = input_157_pad_type_0, strides = input_157_strides_0, weight = layers_19_fc1_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_mode_0 = const()[name = tensor("input_159_mode_0"), val = tensor("EXACT")]; + tensor input_159_cast_fp16 = gelu(mode = input_159_mode_0, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor hidden_states_43_pad_type_0 = const()[name = tensor("hidden_states_43_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_43_strides_0 = const()[name = tensor("hidden_states_43_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_43_pad_0 = const()[name = tensor("hidden_states_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_43_dilations_0 = const()[name = tensor("hidden_states_43_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_43_groups_0 = const()[name = tensor("hidden_states_43_groups_0"), val = tensor(1)]; + tensor layers_19_fc2_weight_to_fp16 = const()[name = tensor("layers_19_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(788619904)))]; + tensor layers_19_fc2_bias_to_fp16 = const()[name = tensor("layers_19_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801727168)))]; + tensor hidden_states_43_cast_fp16 = conv(bias = layers_19_fc2_bias_to_fp16, dilations = hidden_states_43_dilations_0, groups = hidden_states_43_groups_0, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = hidden_states_43_strides_0, weight = layers_19_fc2_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("hidden_states_43_cast_fp16")]; + tensor inputs_81_cast_fp16 = add(x = inputs_79_cast_fp16, y = hidden_states_43_cast_fp16)[name = tensor("inputs_81_cast_fp16")]; + tensor var_2538 = const()[name = tensor("op_2538"), val = tensor(3)]; + tensor out_81_axes_0 = const()[name = tensor("out_81_axes_0"), val = tensor([1])]; + tensor var_2557_to_fp16 = const()[name = tensor("op_2557_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_81_cast_fp16 = layer_norm(axes = out_81_axes_0, epsilon = var_2557_to_fp16, x = inputs_81_cast_fp16)[name = tensor("out_81_cast_fp16")]; + tensor obj_81_gamma_0_to_fp16 = const()[name = tensor("obj_81_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801729792)))]; + tensor obj_81_beta_0_to_fp16 = const()[name = tensor("obj_81_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801732416)))]; + tensor obj_81_epsilon_0_to_fp16 = const()[name = tensor("obj_81_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_81_cast_fp16 = batch_norm(beta = obj_81_beta_0_to_fp16, epsilon = obj_81_epsilon_0_to_fp16, gamma = obj_81_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_81_cast_fp16)[name = tensor("obj_81_cast_fp16")]; + tensor query_41_pad_type_0 = const()[name = tensor("query_41_pad_type_0"), val = tensor("valid")]; + tensor query_41_strides_0 = const()[name = tensor("query_41_strides_0"), val = tensor([1, 1])]; + tensor query_41_pad_0 = const()[name = tensor("query_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_41_dilations_0 = const()[name = tensor("query_41_dilations_0"), val = tensor([1, 1])]; + tensor query_41_groups_0 = const()[name = tensor("query_41_groups_0"), val = tensor(1)]; + tensor layers_20_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801735040)))]; + tensor layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(805011904)))]; + tensor query_41_cast_fp16 = conv(bias = layers_20_self_attn_q_proj_bias_to_fp16, dilations = query_41_dilations_0, groups = query_41_groups_0, pad = query_41_pad_0, pad_type = query_41_pad_type_0, strides = query_41_strides_0, weight = layers_20_self_attn_q_proj_weight_to_fp16, x = obj_81_cast_fp16)[name = tensor("query_41_cast_fp16")]; + tensor key_41_pad_type_0 = const()[name = tensor("key_41_pad_type_0"), val = tensor("valid")]; + tensor key_41_strides_0 = const()[name = tensor("key_41_strides_0"), val = tensor([1, 1])]; + tensor key_41_pad_0 = const()[name = tensor("key_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_41_dilations_0 = const()[name = tensor("key_41_dilations_0"), val = tensor([1, 1])]; + tensor key_41_groups_0 = const()[name = tensor("key_41_groups_0"), val = tensor(1)]; + tensor layers_20_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(805014528)))]; + tensor key_41_cast_fp16 = conv(dilations = key_41_dilations_0, groups = key_41_groups_0, pad = key_41_pad_0, pad_type = key_41_pad_type_0, strides = key_41_strides_0, weight = layers_20_self_attn_k_proj_weight_to_fp16, x = obj_81_cast_fp16)[name = tensor("key_41_cast_fp16")]; + tensor value_41_pad_type_0 = const()[name = tensor("value_41_pad_type_0"), val = tensor("valid")]; + tensor value_41_strides_0 = const()[name = tensor("value_41_strides_0"), val = tensor([1, 1])]; + tensor value_41_pad_0 = const()[name = tensor("value_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_41_dilations_0 = const()[name = tensor("value_41_dilations_0"), val = tensor([1, 1])]; + tensor value_41_groups_0 = const()[name = tensor("value_41_groups_0"), val = tensor(1)]; + tensor layers_20_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(808291392)))]; + tensor layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(811568256)))]; + tensor value_41_cast_fp16 = conv(bias = layers_20_self_attn_v_proj_bias_to_fp16, dilations = value_41_dilations_0, groups = value_41_groups_0, pad = value_41_pad_0, pad_type = value_41_pad_type_0, strides = value_41_strides_0, weight = layers_20_self_attn_v_proj_weight_to_fp16, x = obj_81_cast_fp16)[name = tensor("value_41_cast_fp16")]; + tensor var_2592 = const()[name = tensor("op_2592"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_41_cast_fp16 = reshape(shape = var_2592, x = query_41_cast_fp16)[name = tensor("mh_q_41_cast_fp16")]; + tensor var_2594_to_fp16 = const()[name = tensor("op_2594_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2595_cast_fp16 = mul(x = mh_q_41_cast_fp16, y = var_2594_to_fp16)[name = tensor("op_2595_cast_fp16")]; + tensor var_2596 = const()[name = tensor("op_2596"), val = tensor([1, 20, 64, -1])]; + tensor var_2597_cast_fp16 = reshape(shape = var_2596, x = key_41_cast_fp16)[name = tensor("op_2597_cast_fp16")]; + tensor mh_w_41_transpose_x_0 = const()[name = tensor("mh_w_41_transpose_x_0"), val = tensor(true)]; + tensor mh_w_41_transpose_y_0 = const()[name = tensor("mh_w_41_transpose_y_0"), val = tensor(false)]; + tensor mh_w_41_cast_fp16 = matmul(transpose_x = mh_w_41_transpose_x_0, transpose_y = mh_w_41_transpose_y_0, x = var_2595_cast_fp16, y = var_2597_cast_fp16)[name = tensor("mh_w_41_cast_fp16")]; + tensor var_2600_cast_fp16 = softmax(axis = var_2538, x = mh_w_41_cast_fp16)[name = tensor("op_2600_cast_fp16")]; + tensor var_2601 = const()[name = tensor("op_2601"), val = tensor([1, 20, 64, -1])]; + tensor var_2602_cast_fp16 = reshape(shape = var_2601, x = value_41_cast_fp16)[name = tensor("op_2602_cast_fp16")]; + tensor attn_41_transpose_x_0 = const()[name = tensor("attn_41_transpose_x_0"), val = tensor(false)]; + tensor attn_41_transpose_y_0 = const()[name = tensor("attn_41_transpose_y_0"), val = tensor(true)]; + tensor attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = var_2602_cast_fp16, y = var_2600_cast_fp16)[name = tensor("attn_41_cast_fp16")]; + tensor var_2605 = const()[name = tensor("op_2605"), val = tensor([1, 1280, 1, -1])]; + tensor input_161_cast_fp16 = reshape(shape = var_2605, x = attn_41_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor obj_83_pad_type_0 = const()[name = tensor("obj_83_pad_type_0"), val = tensor("valid")]; + tensor obj_83_strides_0 = const()[name = tensor("obj_83_strides_0"), val = tensor([1, 1])]; + tensor obj_83_pad_0 = const()[name = tensor("obj_83_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_83_dilations_0 = const()[name = tensor("obj_83_dilations_0"), val = tensor([1, 1])]; + tensor obj_83_groups_0 = const()[name = tensor("obj_83_groups_0"), val = tensor(1)]; + tensor layers_20_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(811570880)))]; + tensor layers_20_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814847744)))]; + tensor obj_83_cast_fp16 = conv(bias = layers_20_self_attn_o_proj_bias_to_fp16, dilations = obj_83_dilations_0, groups = obj_83_groups_0, pad = obj_83_pad_0, pad_type = obj_83_pad_type_0, strides = obj_83_strides_0, weight = layers_20_self_attn_o_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("obj_83_cast_fp16")]; + tensor inputs_83_cast_fp16 = add(x = inputs_81_cast_fp16, y = obj_83_cast_fp16)[name = tensor("inputs_83_cast_fp16")]; + tensor out_83_axes_0 = const()[name = tensor("out_83_axes_0"), val = tensor([1])]; + tensor var_2623_to_fp16 = const()[name = tensor("op_2623_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_83_cast_fp16 = layer_norm(axes = out_83_axes_0, epsilon = var_2623_to_fp16, x = inputs_83_cast_fp16)[name = tensor("out_83_cast_fp16")]; + tensor input_163_gamma_0_to_fp16 = const()[name = tensor("input_163_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814850368)))]; + tensor input_163_beta_0_to_fp16 = const()[name = tensor("input_163_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814852992)))]; + tensor input_163_epsilon_0_to_fp16 = const()[name = tensor("input_163_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_163_cast_fp16 = batch_norm(beta = input_163_beta_0_to_fp16, epsilon = input_163_epsilon_0_to_fp16, gamma = input_163_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_83_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor input_165_pad_type_0 = const()[name = tensor("input_165_pad_type_0"), val = tensor("valid")]; + tensor input_165_strides_0 = const()[name = tensor("input_165_strides_0"), val = tensor([1, 1])]; + tensor input_165_pad_0 = const()[name = tensor("input_165_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_165_dilations_0 = const()[name = tensor("input_165_dilations_0"), val = tensor([1, 1])]; + tensor input_165_groups_0 = const()[name = tensor("input_165_groups_0"), val = tensor(1)]; + tensor layers_20_fc1_weight_to_fp16 = const()[name = tensor("layers_20_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814855616)))]; + tensor layers_20_fc1_bias_to_fp16 = const()[name = tensor("layers_20_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(827962880)))]; + tensor input_165_cast_fp16 = conv(bias = layers_20_fc1_bias_to_fp16, dilations = input_165_dilations_0, groups = input_165_groups_0, pad = input_165_pad_0, pad_type = input_165_pad_type_0, strides = input_165_strides_0, weight = layers_20_fc1_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor input_167_mode_0 = const()[name = tensor("input_167_mode_0"), val = tensor("EXACT")]; + tensor input_167_cast_fp16 = gelu(mode = input_167_mode_0, x = input_165_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor hidden_states_45_pad_type_0 = const()[name = tensor("hidden_states_45_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_45_strides_0 = const()[name = tensor("hidden_states_45_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_45_pad_0 = const()[name = tensor("hidden_states_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_45_dilations_0 = const()[name = tensor("hidden_states_45_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_45_groups_0 = const()[name = tensor("hidden_states_45_groups_0"), val = tensor(1)]; + tensor layers_20_fc2_weight_to_fp16 = const()[name = tensor("layers_20_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(827973184)))]; + tensor layers_20_fc2_bias_to_fp16 = const()[name = tensor("layers_20_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(841080448)))]; + tensor hidden_states_45_cast_fp16 = conv(bias = layers_20_fc2_bias_to_fp16, dilations = hidden_states_45_dilations_0, groups = hidden_states_45_groups_0, pad = hidden_states_45_pad_0, pad_type = hidden_states_45_pad_type_0, strides = hidden_states_45_strides_0, weight = layers_20_fc2_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("hidden_states_45_cast_fp16")]; + tensor inputs_85_cast_fp16 = add(x = inputs_83_cast_fp16, y = hidden_states_45_cast_fp16)[name = tensor("inputs_85_cast_fp16")]; + tensor var_2656 = const()[name = tensor("op_2656"), val = tensor(3)]; + tensor out_85_axes_0 = const()[name = tensor("out_85_axes_0"), val = tensor([1])]; + tensor var_2675_to_fp16 = const()[name = tensor("op_2675_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_85_cast_fp16 = layer_norm(axes = out_85_axes_0, epsilon = var_2675_to_fp16, x = inputs_85_cast_fp16)[name = tensor("out_85_cast_fp16")]; + tensor obj_85_gamma_0_to_fp16 = const()[name = tensor("obj_85_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(841083072)))]; + tensor obj_85_beta_0_to_fp16 = const()[name = tensor("obj_85_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(841085696)))]; + tensor obj_85_epsilon_0_to_fp16 = const()[name = tensor("obj_85_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_85_cast_fp16 = batch_norm(beta = obj_85_beta_0_to_fp16, epsilon = obj_85_epsilon_0_to_fp16, gamma = obj_85_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_85_cast_fp16)[name = tensor("obj_85_cast_fp16")]; + tensor query_43_pad_type_0 = const()[name = tensor("query_43_pad_type_0"), val = tensor("valid")]; + tensor query_43_strides_0 = const()[name = tensor("query_43_strides_0"), val = tensor([1, 1])]; + tensor query_43_pad_0 = const()[name = tensor("query_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_43_dilations_0 = const()[name = tensor("query_43_dilations_0"), val = tensor([1, 1])]; + tensor query_43_groups_0 = const()[name = tensor("query_43_groups_0"), val = tensor(1)]; + tensor layers_21_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(841088320)))]; + tensor layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(844365184)))]; + tensor query_43_cast_fp16 = conv(bias = layers_21_self_attn_q_proj_bias_to_fp16, dilations = query_43_dilations_0, groups = query_43_groups_0, pad = query_43_pad_0, pad_type = query_43_pad_type_0, strides = query_43_strides_0, weight = layers_21_self_attn_q_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("query_43_cast_fp16")]; + tensor key_43_pad_type_0 = const()[name = tensor("key_43_pad_type_0"), val = tensor("valid")]; + tensor key_43_strides_0 = const()[name = tensor("key_43_strides_0"), val = tensor([1, 1])]; + tensor key_43_pad_0 = const()[name = tensor("key_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_43_dilations_0 = const()[name = tensor("key_43_dilations_0"), val = tensor([1, 1])]; + tensor key_43_groups_0 = const()[name = tensor("key_43_groups_0"), val = tensor(1)]; + tensor layers_21_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(844367808)))]; + tensor key_43_cast_fp16 = conv(dilations = key_43_dilations_0, groups = key_43_groups_0, pad = key_43_pad_0, pad_type = key_43_pad_type_0, strides = key_43_strides_0, weight = layers_21_self_attn_k_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("key_43_cast_fp16")]; + tensor value_43_pad_type_0 = const()[name = tensor("value_43_pad_type_0"), val = tensor("valid")]; + tensor value_43_strides_0 = const()[name = tensor("value_43_strides_0"), val = tensor([1, 1])]; + tensor value_43_pad_0 = const()[name = tensor("value_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_43_dilations_0 = const()[name = tensor("value_43_dilations_0"), val = tensor([1, 1])]; + tensor value_43_groups_0 = const()[name = tensor("value_43_groups_0"), val = tensor(1)]; + tensor layers_21_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(847644672)))]; + tensor layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(850921536)))]; + tensor value_43_cast_fp16 = conv(bias = layers_21_self_attn_v_proj_bias_to_fp16, dilations = value_43_dilations_0, groups = value_43_groups_0, pad = value_43_pad_0, pad_type = value_43_pad_type_0, strides = value_43_strides_0, weight = layers_21_self_attn_v_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("value_43_cast_fp16")]; + tensor var_2710 = const()[name = tensor("op_2710"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_43_cast_fp16 = reshape(shape = var_2710, x = query_43_cast_fp16)[name = tensor("mh_q_43_cast_fp16")]; + tensor var_2712_to_fp16 = const()[name = tensor("op_2712_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2713_cast_fp16 = mul(x = mh_q_43_cast_fp16, y = var_2712_to_fp16)[name = tensor("op_2713_cast_fp16")]; + tensor var_2714 = const()[name = tensor("op_2714"), val = tensor([1, 20, 64, -1])]; + tensor var_2715_cast_fp16 = reshape(shape = var_2714, x = key_43_cast_fp16)[name = tensor("op_2715_cast_fp16")]; + tensor mh_w_43_transpose_x_0 = const()[name = tensor("mh_w_43_transpose_x_0"), val = tensor(true)]; + tensor mh_w_43_transpose_y_0 = const()[name = tensor("mh_w_43_transpose_y_0"), val = tensor(false)]; + tensor mh_w_43_cast_fp16 = matmul(transpose_x = mh_w_43_transpose_x_0, transpose_y = mh_w_43_transpose_y_0, x = var_2713_cast_fp16, y = var_2715_cast_fp16)[name = tensor("mh_w_43_cast_fp16")]; + tensor var_2718_cast_fp16 = softmax(axis = var_2656, x = mh_w_43_cast_fp16)[name = tensor("op_2718_cast_fp16")]; + tensor var_2719 = const()[name = tensor("op_2719"), val = tensor([1, 20, 64, -1])]; + tensor var_2720_cast_fp16 = reshape(shape = var_2719, x = value_43_cast_fp16)[name = tensor("op_2720_cast_fp16")]; + tensor attn_43_transpose_x_0 = const()[name = tensor("attn_43_transpose_x_0"), val = tensor(false)]; + tensor attn_43_transpose_y_0 = const()[name = tensor("attn_43_transpose_y_0"), val = tensor(true)]; + tensor attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = var_2720_cast_fp16, y = var_2718_cast_fp16)[name = tensor("attn_43_cast_fp16")]; + tensor var_2723 = const()[name = tensor("op_2723"), val = tensor([1, 1280, 1, -1])]; + tensor input_169_cast_fp16 = reshape(shape = var_2723, x = attn_43_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor obj_87_pad_type_0 = const()[name = tensor("obj_87_pad_type_0"), val = tensor("valid")]; + tensor obj_87_strides_0 = const()[name = tensor("obj_87_strides_0"), val = tensor([1, 1])]; + tensor obj_87_pad_0 = const()[name = tensor("obj_87_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_87_dilations_0 = const()[name = tensor("obj_87_dilations_0"), val = tensor([1, 1])]; + tensor obj_87_groups_0 = const()[name = tensor("obj_87_groups_0"), val = tensor(1)]; + tensor layers_21_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(850924160)))]; + tensor layers_21_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(854201024)))]; + tensor obj_87_cast_fp16 = conv(bias = layers_21_self_attn_o_proj_bias_to_fp16, dilations = obj_87_dilations_0, groups = obj_87_groups_0, pad = obj_87_pad_0, pad_type = obj_87_pad_type_0, strides = obj_87_strides_0, weight = layers_21_self_attn_o_proj_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("obj_87_cast_fp16")]; + tensor inputs_87_cast_fp16 = add(x = inputs_85_cast_fp16, y = obj_87_cast_fp16)[name = tensor("inputs_87_cast_fp16")]; + tensor out_87_axes_0 = const()[name = tensor("out_87_axes_0"), val = tensor([1])]; + tensor var_2741_to_fp16 = const()[name = tensor("op_2741_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_87_cast_fp16 = layer_norm(axes = out_87_axes_0, epsilon = var_2741_to_fp16, x = inputs_87_cast_fp16)[name = tensor("out_87_cast_fp16")]; + tensor input_171_gamma_0_to_fp16 = const()[name = tensor("input_171_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(854203648)))]; + tensor input_171_beta_0_to_fp16 = const()[name = tensor("input_171_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(854206272)))]; + tensor input_171_epsilon_0_to_fp16 = const()[name = tensor("input_171_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_171_cast_fp16 = batch_norm(beta = input_171_beta_0_to_fp16, epsilon = input_171_epsilon_0_to_fp16, gamma = input_171_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_87_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor input_173_pad_type_0 = const()[name = tensor("input_173_pad_type_0"), val = tensor("valid")]; + tensor input_173_strides_0 = const()[name = tensor("input_173_strides_0"), val = tensor([1, 1])]; + tensor input_173_pad_0 = const()[name = tensor("input_173_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_173_dilations_0 = const()[name = tensor("input_173_dilations_0"), val = tensor([1, 1])]; + tensor input_173_groups_0 = const()[name = tensor("input_173_groups_0"), val = tensor(1)]; + tensor layers_21_fc1_weight_to_fp16 = const()[name = tensor("layers_21_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(854208896)))]; + tensor layers_21_fc1_bias_to_fp16 = const()[name = tensor("layers_21_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(867316160)))]; + tensor input_173_cast_fp16 = conv(bias = layers_21_fc1_bias_to_fp16, dilations = input_173_dilations_0, groups = input_173_groups_0, pad = input_173_pad_0, pad_type = input_173_pad_type_0, strides = input_173_strides_0, weight = layers_21_fc1_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor input_175_mode_0 = const()[name = tensor("input_175_mode_0"), val = tensor("EXACT")]; + tensor input_175_cast_fp16 = gelu(mode = input_175_mode_0, x = input_173_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor hidden_states_47_pad_type_0 = const()[name = tensor("hidden_states_47_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_47_strides_0 = const()[name = tensor("hidden_states_47_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_47_pad_0 = const()[name = tensor("hidden_states_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_47_dilations_0 = const()[name = tensor("hidden_states_47_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_47_groups_0 = const()[name = tensor("hidden_states_47_groups_0"), val = tensor(1)]; + tensor layers_21_fc2_weight_to_fp16 = const()[name = tensor("layers_21_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(867326464)))]; + tensor layers_21_fc2_bias_to_fp16 = const()[name = tensor("layers_21_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880433728)))]; + tensor hidden_states_47_cast_fp16 = conv(bias = layers_21_fc2_bias_to_fp16, dilations = hidden_states_47_dilations_0, groups = hidden_states_47_groups_0, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = hidden_states_47_strides_0, weight = layers_21_fc2_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("hidden_states_47_cast_fp16")]; + tensor inputs_89_cast_fp16 = add(x = inputs_87_cast_fp16, y = hidden_states_47_cast_fp16)[name = tensor("inputs_89_cast_fp16")]; + tensor var_2774 = const()[name = tensor("op_2774"), val = tensor(3)]; + tensor out_89_axes_0 = const()[name = tensor("out_89_axes_0"), val = tensor([1])]; + tensor var_2793_to_fp16 = const()[name = tensor("op_2793_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_89_cast_fp16 = layer_norm(axes = out_89_axes_0, epsilon = var_2793_to_fp16, x = inputs_89_cast_fp16)[name = tensor("out_89_cast_fp16")]; + tensor obj_89_gamma_0_to_fp16 = const()[name = tensor("obj_89_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880436352)))]; + tensor obj_89_beta_0_to_fp16 = const()[name = tensor("obj_89_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880438976)))]; + tensor obj_89_epsilon_0_to_fp16 = const()[name = tensor("obj_89_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_89_cast_fp16 = batch_norm(beta = obj_89_beta_0_to_fp16, epsilon = obj_89_epsilon_0_to_fp16, gamma = obj_89_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_89_cast_fp16)[name = tensor("obj_89_cast_fp16")]; + tensor query_45_pad_type_0 = const()[name = tensor("query_45_pad_type_0"), val = tensor("valid")]; + tensor query_45_strides_0 = const()[name = tensor("query_45_strides_0"), val = tensor([1, 1])]; + tensor query_45_pad_0 = const()[name = tensor("query_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_45_dilations_0 = const()[name = tensor("query_45_dilations_0"), val = tensor([1, 1])]; + tensor query_45_groups_0 = const()[name = tensor("query_45_groups_0"), val = tensor(1)]; + tensor layers_22_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880441600)))]; + tensor layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(883718464)))]; + tensor query_45_cast_fp16 = conv(bias = layers_22_self_attn_q_proj_bias_to_fp16, dilations = query_45_dilations_0, groups = query_45_groups_0, pad = query_45_pad_0, pad_type = query_45_pad_type_0, strides = query_45_strides_0, weight = layers_22_self_attn_q_proj_weight_to_fp16, x = obj_89_cast_fp16)[name = tensor("query_45_cast_fp16")]; + tensor key_45_pad_type_0 = const()[name = tensor("key_45_pad_type_0"), val = tensor("valid")]; + tensor key_45_strides_0 = const()[name = tensor("key_45_strides_0"), val = tensor([1, 1])]; + tensor key_45_pad_0 = const()[name = tensor("key_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_45_dilations_0 = const()[name = tensor("key_45_dilations_0"), val = tensor([1, 1])]; + tensor key_45_groups_0 = const()[name = tensor("key_45_groups_0"), val = tensor(1)]; + tensor layers_22_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(883721088)))]; + tensor key_45_cast_fp16 = conv(dilations = key_45_dilations_0, groups = key_45_groups_0, pad = key_45_pad_0, pad_type = key_45_pad_type_0, strides = key_45_strides_0, weight = layers_22_self_attn_k_proj_weight_to_fp16, x = obj_89_cast_fp16)[name = tensor("key_45_cast_fp16")]; + tensor value_45_pad_type_0 = const()[name = tensor("value_45_pad_type_0"), val = tensor("valid")]; + tensor value_45_strides_0 = const()[name = tensor("value_45_strides_0"), val = tensor([1, 1])]; + tensor value_45_pad_0 = const()[name = tensor("value_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_45_dilations_0 = const()[name = tensor("value_45_dilations_0"), val = tensor([1, 1])]; + tensor value_45_groups_0 = const()[name = tensor("value_45_groups_0"), val = tensor(1)]; + tensor layers_22_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(886997952)))]; + tensor layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(890274816)))]; + tensor value_45_cast_fp16 = conv(bias = layers_22_self_attn_v_proj_bias_to_fp16, dilations = value_45_dilations_0, groups = value_45_groups_0, pad = value_45_pad_0, pad_type = value_45_pad_type_0, strides = value_45_strides_0, weight = layers_22_self_attn_v_proj_weight_to_fp16, x = obj_89_cast_fp16)[name = tensor("value_45_cast_fp16")]; + tensor var_2828 = const()[name = tensor("op_2828"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_45_cast_fp16 = reshape(shape = var_2828, x = query_45_cast_fp16)[name = tensor("mh_q_45_cast_fp16")]; + tensor var_2830_to_fp16 = const()[name = tensor("op_2830_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2831_cast_fp16 = mul(x = mh_q_45_cast_fp16, y = var_2830_to_fp16)[name = tensor("op_2831_cast_fp16")]; + tensor var_2832 = const()[name = tensor("op_2832"), val = tensor([1, 20, 64, -1])]; + tensor var_2833_cast_fp16 = reshape(shape = var_2832, x = key_45_cast_fp16)[name = tensor("op_2833_cast_fp16")]; + tensor mh_w_45_transpose_x_0 = const()[name = tensor("mh_w_45_transpose_x_0"), val = tensor(true)]; + tensor mh_w_45_transpose_y_0 = const()[name = tensor("mh_w_45_transpose_y_0"), val = tensor(false)]; + tensor mh_w_45_cast_fp16 = matmul(transpose_x = mh_w_45_transpose_x_0, transpose_y = mh_w_45_transpose_y_0, x = var_2831_cast_fp16, y = var_2833_cast_fp16)[name = tensor("mh_w_45_cast_fp16")]; + tensor var_2836_cast_fp16 = softmax(axis = var_2774, x = mh_w_45_cast_fp16)[name = tensor("op_2836_cast_fp16")]; + tensor var_2837 = const()[name = tensor("op_2837"), val = tensor([1, 20, 64, -1])]; + tensor var_2838_cast_fp16 = reshape(shape = var_2837, x = value_45_cast_fp16)[name = tensor("op_2838_cast_fp16")]; + tensor attn_45_transpose_x_0 = const()[name = tensor("attn_45_transpose_x_0"), val = tensor(false)]; + tensor attn_45_transpose_y_0 = const()[name = tensor("attn_45_transpose_y_0"), val = tensor(true)]; + tensor attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = var_2838_cast_fp16, y = var_2836_cast_fp16)[name = tensor("attn_45_cast_fp16")]; + tensor var_2841 = const()[name = tensor("op_2841"), val = tensor([1, 1280, 1, -1])]; + tensor input_177_cast_fp16 = reshape(shape = var_2841, x = attn_45_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor obj_91_pad_type_0 = const()[name = tensor("obj_91_pad_type_0"), val = tensor("valid")]; + tensor obj_91_strides_0 = const()[name = tensor("obj_91_strides_0"), val = tensor([1, 1])]; + tensor obj_91_pad_0 = const()[name = tensor("obj_91_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_91_dilations_0 = const()[name = tensor("obj_91_dilations_0"), val = tensor([1, 1])]; + tensor obj_91_groups_0 = const()[name = tensor("obj_91_groups_0"), val = tensor(1)]; + tensor layers_22_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(890277440)))]; + tensor layers_22_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893554304)))]; + tensor obj_91_cast_fp16 = conv(bias = layers_22_self_attn_o_proj_bias_to_fp16, dilations = obj_91_dilations_0, groups = obj_91_groups_0, pad = obj_91_pad_0, pad_type = obj_91_pad_type_0, strides = obj_91_strides_0, weight = layers_22_self_attn_o_proj_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("obj_91_cast_fp16")]; + tensor inputs_91_cast_fp16 = add(x = inputs_89_cast_fp16, y = obj_91_cast_fp16)[name = tensor("inputs_91_cast_fp16")]; + tensor out_91_axes_0 = const()[name = tensor("out_91_axes_0"), val = tensor([1])]; + tensor var_2859_to_fp16 = const()[name = tensor("op_2859_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_91_cast_fp16 = layer_norm(axes = out_91_axes_0, epsilon = var_2859_to_fp16, x = inputs_91_cast_fp16)[name = tensor("out_91_cast_fp16")]; + tensor input_179_gamma_0_to_fp16 = const()[name = tensor("input_179_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893556928)))]; + tensor input_179_beta_0_to_fp16 = const()[name = tensor("input_179_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893559552)))]; + tensor input_179_epsilon_0_to_fp16 = const()[name = tensor("input_179_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_179_cast_fp16 = batch_norm(beta = input_179_beta_0_to_fp16, epsilon = input_179_epsilon_0_to_fp16, gamma = input_179_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_91_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor input_181_pad_type_0 = const()[name = tensor("input_181_pad_type_0"), val = tensor("valid")]; + tensor input_181_strides_0 = const()[name = tensor("input_181_strides_0"), val = tensor([1, 1])]; + tensor input_181_pad_0 = const()[name = tensor("input_181_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_181_dilations_0 = const()[name = tensor("input_181_dilations_0"), val = tensor([1, 1])]; + tensor input_181_groups_0 = const()[name = tensor("input_181_groups_0"), val = tensor(1)]; + tensor layers_22_fc1_weight_to_fp16 = const()[name = tensor("layers_22_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893562176)))]; + tensor layers_22_fc1_bias_to_fp16 = const()[name = tensor("layers_22_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(906669440)))]; + tensor input_181_cast_fp16 = conv(bias = layers_22_fc1_bias_to_fp16, dilations = input_181_dilations_0, groups = input_181_groups_0, pad = input_181_pad_0, pad_type = input_181_pad_type_0, strides = input_181_strides_0, weight = layers_22_fc1_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor input_183_mode_0 = const()[name = tensor("input_183_mode_0"), val = tensor("EXACT")]; + tensor input_183_cast_fp16 = gelu(mode = input_183_mode_0, x = input_181_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor hidden_states_49_pad_type_0 = const()[name = tensor("hidden_states_49_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_49_strides_0 = const()[name = tensor("hidden_states_49_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_49_pad_0 = const()[name = tensor("hidden_states_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_49_dilations_0 = const()[name = tensor("hidden_states_49_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_49_groups_0 = const()[name = tensor("hidden_states_49_groups_0"), val = tensor(1)]; + tensor layers_22_fc2_weight_to_fp16 = const()[name = tensor("layers_22_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(906679744)))]; + tensor layers_22_fc2_bias_to_fp16 = const()[name = tensor("layers_22_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919787008)))]; + tensor hidden_states_49_cast_fp16 = conv(bias = layers_22_fc2_bias_to_fp16, dilations = hidden_states_49_dilations_0, groups = hidden_states_49_groups_0, pad = hidden_states_49_pad_0, pad_type = hidden_states_49_pad_type_0, strides = hidden_states_49_strides_0, weight = layers_22_fc2_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("hidden_states_49_cast_fp16")]; + tensor inputs_93_cast_fp16 = add(x = inputs_91_cast_fp16, y = hidden_states_49_cast_fp16)[name = tensor("inputs_93_cast_fp16")]; + tensor var_2892 = const()[name = tensor("op_2892"), val = tensor(3)]; + tensor out_93_axes_0 = const()[name = tensor("out_93_axes_0"), val = tensor([1])]; + tensor var_2911_to_fp16 = const()[name = tensor("op_2911_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_93_cast_fp16 = layer_norm(axes = out_93_axes_0, epsilon = var_2911_to_fp16, x = inputs_93_cast_fp16)[name = tensor("out_93_cast_fp16")]; + tensor obj_93_gamma_0_to_fp16 = const()[name = tensor("obj_93_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919789632)))]; + tensor obj_93_beta_0_to_fp16 = const()[name = tensor("obj_93_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919792256)))]; + tensor obj_93_epsilon_0_to_fp16 = const()[name = tensor("obj_93_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_93_cast_fp16 = batch_norm(beta = obj_93_beta_0_to_fp16, epsilon = obj_93_epsilon_0_to_fp16, gamma = obj_93_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_93_cast_fp16)[name = tensor("obj_93_cast_fp16")]; + tensor query_47_pad_type_0 = const()[name = tensor("query_47_pad_type_0"), val = tensor("valid")]; + tensor query_47_strides_0 = const()[name = tensor("query_47_strides_0"), val = tensor([1, 1])]; + tensor query_47_pad_0 = const()[name = tensor("query_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_47_dilations_0 = const()[name = tensor("query_47_dilations_0"), val = tensor([1, 1])]; + tensor query_47_groups_0 = const()[name = tensor("query_47_groups_0"), val = tensor(1)]; + tensor layers_23_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919794880)))]; + tensor layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(923071744)))]; + tensor query_47_cast_fp16 = conv(bias = layers_23_self_attn_q_proj_bias_to_fp16, dilations = query_47_dilations_0, groups = query_47_groups_0, pad = query_47_pad_0, pad_type = query_47_pad_type_0, strides = query_47_strides_0, weight = layers_23_self_attn_q_proj_weight_to_fp16, x = obj_93_cast_fp16)[name = tensor("query_47_cast_fp16")]; + tensor key_47_pad_type_0 = const()[name = tensor("key_47_pad_type_0"), val = tensor("valid")]; + tensor key_47_strides_0 = const()[name = tensor("key_47_strides_0"), val = tensor([1, 1])]; + tensor key_47_pad_0 = const()[name = tensor("key_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_47_dilations_0 = const()[name = tensor("key_47_dilations_0"), val = tensor([1, 1])]; + tensor key_47_groups_0 = const()[name = tensor("key_47_groups_0"), val = tensor(1)]; + tensor layers_23_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(923074368)))]; + tensor key_47_cast_fp16 = conv(dilations = key_47_dilations_0, groups = key_47_groups_0, pad = key_47_pad_0, pad_type = key_47_pad_type_0, strides = key_47_strides_0, weight = layers_23_self_attn_k_proj_weight_to_fp16, x = obj_93_cast_fp16)[name = tensor("key_47_cast_fp16")]; + tensor value_47_pad_type_0 = const()[name = tensor("value_47_pad_type_0"), val = tensor("valid")]; + tensor value_47_strides_0 = const()[name = tensor("value_47_strides_0"), val = tensor([1, 1])]; + tensor value_47_pad_0 = const()[name = tensor("value_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_47_dilations_0 = const()[name = tensor("value_47_dilations_0"), val = tensor([1, 1])]; + tensor value_47_groups_0 = const()[name = tensor("value_47_groups_0"), val = tensor(1)]; + tensor layers_23_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(926351232)))]; + tensor layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(929628096)))]; + tensor value_47_cast_fp16 = conv(bias = layers_23_self_attn_v_proj_bias_to_fp16, dilations = value_47_dilations_0, groups = value_47_groups_0, pad = value_47_pad_0, pad_type = value_47_pad_type_0, strides = value_47_strides_0, weight = layers_23_self_attn_v_proj_weight_to_fp16, x = obj_93_cast_fp16)[name = tensor("value_47_cast_fp16")]; + tensor var_2946 = const()[name = tensor("op_2946"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_47_cast_fp16 = reshape(shape = var_2946, x = query_47_cast_fp16)[name = tensor("mh_q_47_cast_fp16")]; + tensor var_2948_to_fp16 = const()[name = tensor("op_2948_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2949_cast_fp16 = mul(x = mh_q_47_cast_fp16, y = var_2948_to_fp16)[name = tensor("op_2949_cast_fp16")]; + tensor var_2950 = const()[name = tensor("op_2950"), val = tensor([1, 20, 64, -1])]; + tensor var_2951_cast_fp16 = reshape(shape = var_2950, x = key_47_cast_fp16)[name = tensor("op_2951_cast_fp16")]; + tensor mh_w_47_transpose_x_0 = const()[name = tensor("mh_w_47_transpose_x_0"), val = tensor(true)]; + tensor mh_w_47_transpose_y_0 = const()[name = tensor("mh_w_47_transpose_y_0"), val = tensor(false)]; + tensor mh_w_47_cast_fp16 = matmul(transpose_x = mh_w_47_transpose_x_0, transpose_y = mh_w_47_transpose_y_0, x = var_2949_cast_fp16, y = var_2951_cast_fp16)[name = tensor("mh_w_47_cast_fp16")]; + tensor var_2954_cast_fp16 = softmax(axis = var_2892, x = mh_w_47_cast_fp16)[name = tensor("op_2954_cast_fp16")]; + tensor var_2955 = const()[name = tensor("op_2955"), val = tensor([1, 20, 64, -1])]; + tensor var_2956_cast_fp16 = reshape(shape = var_2955, x = value_47_cast_fp16)[name = tensor("op_2956_cast_fp16")]; + tensor attn_47_transpose_x_0 = const()[name = tensor("attn_47_transpose_x_0"), val = tensor(false)]; + tensor attn_47_transpose_y_0 = const()[name = tensor("attn_47_transpose_y_0"), val = tensor(true)]; + tensor attn_47_cast_fp16 = matmul(transpose_x = attn_47_transpose_x_0, transpose_y = attn_47_transpose_y_0, x = var_2956_cast_fp16, y = var_2954_cast_fp16)[name = tensor("attn_47_cast_fp16")]; + tensor var_2959 = const()[name = tensor("op_2959"), val = tensor([1, 1280, 1, -1])]; + tensor input_185_cast_fp16 = reshape(shape = var_2959, x = attn_47_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor obj_95_pad_type_0 = const()[name = tensor("obj_95_pad_type_0"), val = tensor("valid")]; + tensor obj_95_strides_0 = const()[name = tensor("obj_95_strides_0"), val = tensor([1, 1])]; + tensor obj_95_pad_0 = const()[name = tensor("obj_95_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_95_dilations_0 = const()[name = tensor("obj_95_dilations_0"), val = tensor([1, 1])]; + tensor obj_95_groups_0 = const()[name = tensor("obj_95_groups_0"), val = tensor(1)]; + tensor layers_23_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(929630720)))]; + tensor layers_23_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932907584)))]; + tensor obj_95_cast_fp16 = conv(bias = layers_23_self_attn_o_proj_bias_to_fp16, dilations = obj_95_dilations_0, groups = obj_95_groups_0, pad = obj_95_pad_0, pad_type = obj_95_pad_type_0, strides = obj_95_strides_0, weight = layers_23_self_attn_o_proj_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("obj_95_cast_fp16")]; + tensor inputs_95_cast_fp16 = add(x = inputs_93_cast_fp16, y = obj_95_cast_fp16)[name = tensor("inputs_95_cast_fp16")]; + tensor out_95_axes_0 = const()[name = tensor("out_95_axes_0"), val = tensor([1])]; + tensor var_2977_to_fp16 = const()[name = tensor("op_2977_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_95_cast_fp16 = layer_norm(axes = out_95_axes_0, epsilon = var_2977_to_fp16, x = inputs_95_cast_fp16)[name = tensor("out_95_cast_fp16")]; + tensor input_187_gamma_0_to_fp16 = const()[name = tensor("input_187_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932910208)))]; + tensor input_187_beta_0_to_fp16 = const()[name = tensor("input_187_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932912832)))]; + tensor input_187_epsilon_0_to_fp16 = const()[name = tensor("input_187_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_187_cast_fp16 = batch_norm(beta = input_187_beta_0_to_fp16, epsilon = input_187_epsilon_0_to_fp16, gamma = input_187_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_95_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_pad_type_0 = const()[name = tensor("input_189_pad_type_0"), val = tensor("valid")]; + tensor input_189_strides_0 = const()[name = tensor("input_189_strides_0"), val = tensor([1, 1])]; + tensor input_189_pad_0 = const()[name = tensor("input_189_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_189_dilations_0 = const()[name = tensor("input_189_dilations_0"), val = tensor([1, 1])]; + tensor input_189_groups_0 = const()[name = tensor("input_189_groups_0"), val = tensor(1)]; + tensor layers_23_fc1_weight_to_fp16 = const()[name = tensor("layers_23_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932915456)))]; + tensor layers_23_fc1_bias_to_fp16 = const()[name = tensor("layers_23_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(946022720)))]; + tensor input_189_cast_fp16 = conv(bias = layers_23_fc1_bias_to_fp16, dilations = input_189_dilations_0, groups = input_189_groups_0, pad = input_189_pad_0, pad_type = input_189_pad_type_0, strides = input_189_strides_0, weight = layers_23_fc1_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor input_191_mode_0 = const()[name = tensor("input_191_mode_0"), val = tensor("EXACT")]; + tensor input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = input_189_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor hidden_states_51_pad_type_0 = const()[name = tensor("hidden_states_51_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_51_strides_0 = const()[name = tensor("hidden_states_51_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_51_pad_0 = const()[name = tensor("hidden_states_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_51_dilations_0 = const()[name = tensor("hidden_states_51_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_51_groups_0 = const()[name = tensor("hidden_states_51_groups_0"), val = tensor(1)]; + tensor layers_23_fc2_weight_to_fp16 = const()[name = tensor("layers_23_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(946033024)))]; + tensor layers_23_fc2_bias_to_fp16 = const()[name = tensor("layers_23_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(959140288)))]; + tensor hidden_states_51_cast_fp16 = conv(bias = layers_23_fc2_bias_to_fp16, dilations = hidden_states_51_dilations_0, groups = hidden_states_51_groups_0, pad = hidden_states_51_pad_0, pad_type = hidden_states_51_pad_type_0, strides = hidden_states_51_strides_0, weight = layers_23_fc2_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("hidden_states_51_cast_fp16")]; + tensor inputs_97_cast_fp16 = add(x = inputs_95_cast_fp16, y = hidden_states_51_cast_fp16)[name = tensor("inputs_97_cast_fp16")]; + tensor var_3010 = const()[name = tensor("op_3010"), val = tensor(3)]; + tensor out_97_axes_0 = const()[name = tensor("out_97_axes_0"), val = tensor([1])]; + tensor var_3029_to_fp16 = const()[name = tensor("op_3029_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_97_cast_fp16 = layer_norm(axes = out_97_axes_0, epsilon = var_3029_to_fp16, x = inputs_97_cast_fp16)[name = tensor("out_97_cast_fp16")]; + tensor obj_97_gamma_0_to_fp16 = const()[name = tensor("obj_97_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(959142912)))]; + tensor obj_97_beta_0_to_fp16 = const()[name = tensor("obj_97_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(959145536)))]; + tensor obj_97_epsilon_0_to_fp16 = const()[name = tensor("obj_97_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_97_cast_fp16 = batch_norm(beta = obj_97_beta_0_to_fp16, epsilon = obj_97_epsilon_0_to_fp16, gamma = obj_97_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_97_cast_fp16)[name = tensor("obj_97_cast_fp16")]; + tensor query_49_pad_type_0 = const()[name = tensor("query_49_pad_type_0"), val = tensor("valid")]; + tensor query_49_strides_0 = const()[name = tensor("query_49_strides_0"), val = tensor([1, 1])]; + tensor query_49_pad_0 = const()[name = tensor("query_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_49_dilations_0 = const()[name = tensor("query_49_dilations_0"), val = tensor([1, 1])]; + tensor query_49_groups_0 = const()[name = tensor("query_49_groups_0"), val = tensor(1)]; + tensor layers_24_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_24_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(959148160)))]; + tensor layers_24_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_24_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(962425024)))]; + tensor query_49_cast_fp16 = conv(bias = layers_24_self_attn_q_proj_bias_to_fp16, dilations = query_49_dilations_0, groups = query_49_groups_0, pad = query_49_pad_0, pad_type = query_49_pad_type_0, strides = query_49_strides_0, weight = layers_24_self_attn_q_proj_weight_to_fp16, x = obj_97_cast_fp16)[name = tensor("query_49_cast_fp16")]; + tensor key_49_pad_type_0 = const()[name = tensor("key_49_pad_type_0"), val = tensor("valid")]; + tensor key_49_strides_0 = const()[name = tensor("key_49_strides_0"), val = tensor([1, 1])]; + tensor key_49_pad_0 = const()[name = tensor("key_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_49_dilations_0 = const()[name = tensor("key_49_dilations_0"), val = tensor([1, 1])]; + tensor key_49_groups_0 = const()[name = tensor("key_49_groups_0"), val = tensor(1)]; + tensor layers_24_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_24_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(962427648)))]; + tensor key_49_cast_fp16 = conv(dilations = key_49_dilations_0, groups = key_49_groups_0, pad = key_49_pad_0, pad_type = key_49_pad_type_0, strides = key_49_strides_0, weight = layers_24_self_attn_k_proj_weight_to_fp16, x = obj_97_cast_fp16)[name = tensor("key_49_cast_fp16")]; + tensor value_49_pad_type_0 = const()[name = tensor("value_49_pad_type_0"), val = tensor("valid")]; + tensor value_49_strides_0 = const()[name = tensor("value_49_strides_0"), val = tensor([1, 1])]; + tensor value_49_pad_0 = const()[name = tensor("value_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_49_dilations_0 = const()[name = tensor("value_49_dilations_0"), val = tensor([1, 1])]; + tensor value_49_groups_0 = const()[name = tensor("value_49_groups_0"), val = tensor(1)]; + tensor layers_24_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_24_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(965704512)))]; + tensor layers_24_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_24_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(968981376)))]; + tensor value_49_cast_fp16 = conv(bias = layers_24_self_attn_v_proj_bias_to_fp16, dilations = value_49_dilations_0, groups = value_49_groups_0, pad = value_49_pad_0, pad_type = value_49_pad_type_0, strides = value_49_strides_0, weight = layers_24_self_attn_v_proj_weight_to_fp16, x = obj_97_cast_fp16)[name = tensor("value_49_cast_fp16")]; + tensor var_3064 = const()[name = tensor("op_3064"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_49_cast_fp16 = reshape(shape = var_3064, x = query_49_cast_fp16)[name = tensor("mh_q_49_cast_fp16")]; + tensor var_3066_to_fp16 = const()[name = tensor("op_3066_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3067_cast_fp16 = mul(x = mh_q_49_cast_fp16, y = var_3066_to_fp16)[name = tensor("op_3067_cast_fp16")]; + tensor var_3068 = const()[name = tensor("op_3068"), val = tensor([1, 20, 64, -1])]; + tensor var_3069_cast_fp16 = reshape(shape = var_3068, x = key_49_cast_fp16)[name = tensor("op_3069_cast_fp16")]; + tensor mh_w_49_transpose_x_0 = const()[name = tensor("mh_w_49_transpose_x_0"), val = tensor(true)]; + tensor mh_w_49_transpose_y_0 = const()[name = tensor("mh_w_49_transpose_y_0"), val = tensor(false)]; + tensor mh_w_49_cast_fp16 = matmul(transpose_x = mh_w_49_transpose_x_0, transpose_y = mh_w_49_transpose_y_0, x = var_3067_cast_fp16, y = var_3069_cast_fp16)[name = tensor("mh_w_49_cast_fp16")]; + tensor var_3072_cast_fp16 = softmax(axis = var_3010, x = mh_w_49_cast_fp16)[name = tensor("op_3072_cast_fp16")]; + tensor var_3073 = const()[name = tensor("op_3073"), val = tensor([1, 20, 64, -1])]; + tensor var_3074_cast_fp16 = reshape(shape = var_3073, x = value_49_cast_fp16)[name = tensor("op_3074_cast_fp16")]; + tensor attn_49_transpose_x_0 = const()[name = tensor("attn_49_transpose_x_0"), val = tensor(false)]; + tensor attn_49_transpose_y_0 = const()[name = tensor("attn_49_transpose_y_0"), val = tensor(true)]; + tensor attn_49_cast_fp16 = matmul(transpose_x = attn_49_transpose_x_0, transpose_y = attn_49_transpose_y_0, x = var_3074_cast_fp16, y = var_3072_cast_fp16)[name = tensor("attn_49_cast_fp16")]; + tensor var_3077 = const()[name = tensor("op_3077"), val = tensor([1, 1280, 1, -1])]; + tensor input_193_cast_fp16 = reshape(shape = var_3077, x = attn_49_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor obj_99_pad_type_0 = const()[name = tensor("obj_99_pad_type_0"), val = tensor("valid")]; + tensor obj_99_strides_0 = const()[name = tensor("obj_99_strides_0"), val = tensor([1, 1])]; + tensor obj_99_pad_0 = const()[name = tensor("obj_99_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_99_dilations_0 = const()[name = tensor("obj_99_dilations_0"), val = tensor([1, 1])]; + tensor obj_99_groups_0 = const()[name = tensor("obj_99_groups_0"), val = tensor(1)]; + tensor layers_24_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_24_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(968984000)))]; + tensor layers_24_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_24_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(972260864)))]; + tensor obj_99_cast_fp16 = conv(bias = layers_24_self_attn_o_proj_bias_to_fp16, dilations = obj_99_dilations_0, groups = obj_99_groups_0, pad = obj_99_pad_0, pad_type = obj_99_pad_type_0, strides = obj_99_strides_0, weight = layers_24_self_attn_o_proj_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("obj_99_cast_fp16")]; + tensor inputs_99_cast_fp16 = add(x = inputs_97_cast_fp16, y = obj_99_cast_fp16)[name = tensor("inputs_99_cast_fp16")]; + tensor out_99_axes_0 = const()[name = tensor("out_99_axes_0"), val = tensor([1])]; + tensor var_3095_to_fp16 = const()[name = tensor("op_3095_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_99_cast_fp16 = layer_norm(axes = out_99_axes_0, epsilon = var_3095_to_fp16, x = inputs_99_cast_fp16)[name = tensor("out_99_cast_fp16")]; + tensor input_195_gamma_0_to_fp16 = const()[name = tensor("input_195_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(972263488)))]; + tensor input_195_beta_0_to_fp16 = const()[name = tensor("input_195_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(972266112)))]; + tensor input_195_epsilon_0_to_fp16 = const()[name = tensor("input_195_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_195_cast_fp16 = batch_norm(beta = input_195_beta_0_to_fp16, epsilon = input_195_epsilon_0_to_fp16, gamma = input_195_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_99_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor input_197_pad_type_0 = const()[name = tensor("input_197_pad_type_0"), val = tensor("valid")]; + tensor input_197_strides_0 = const()[name = tensor("input_197_strides_0"), val = tensor([1, 1])]; + tensor input_197_pad_0 = const()[name = tensor("input_197_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_197_dilations_0 = const()[name = tensor("input_197_dilations_0"), val = tensor([1, 1])]; + tensor input_197_groups_0 = const()[name = tensor("input_197_groups_0"), val = tensor(1)]; + tensor layers_24_fc1_weight_to_fp16 = const()[name = tensor("layers_24_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(972268736)))]; + tensor layers_24_fc1_bias_to_fp16 = const()[name = tensor("layers_24_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(985376000)))]; + tensor input_197_cast_fp16 = conv(bias = layers_24_fc1_bias_to_fp16, dilations = input_197_dilations_0, groups = input_197_groups_0, pad = input_197_pad_0, pad_type = input_197_pad_type_0, strides = input_197_strides_0, weight = layers_24_fc1_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor input_199_mode_0 = const()[name = tensor("input_199_mode_0"), val = tensor("EXACT")]; + tensor input_199_cast_fp16 = gelu(mode = input_199_mode_0, x = input_197_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor hidden_states_53_pad_type_0 = const()[name = tensor("hidden_states_53_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_53_strides_0 = const()[name = tensor("hidden_states_53_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_53_pad_0 = const()[name = tensor("hidden_states_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_53_dilations_0 = const()[name = tensor("hidden_states_53_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_53_groups_0 = const()[name = tensor("hidden_states_53_groups_0"), val = tensor(1)]; + tensor layers_24_fc2_weight_to_fp16 = const()[name = tensor("layers_24_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(985386304)))]; + tensor layers_24_fc2_bias_to_fp16 = const()[name = tensor("layers_24_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998493568)))]; + tensor hidden_states_53_cast_fp16 = conv(bias = layers_24_fc2_bias_to_fp16, dilations = hidden_states_53_dilations_0, groups = hidden_states_53_groups_0, pad = hidden_states_53_pad_0, pad_type = hidden_states_53_pad_type_0, strides = hidden_states_53_strides_0, weight = layers_24_fc2_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("hidden_states_53_cast_fp16")]; + tensor inputs_101_cast_fp16 = add(x = inputs_99_cast_fp16, y = hidden_states_53_cast_fp16)[name = tensor("inputs_101_cast_fp16")]; + tensor var_3128 = const()[name = tensor("op_3128"), val = tensor(3)]; + tensor out_101_axes_0 = const()[name = tensor("out_101_axes_0"), val = tensor([1])]; + tensor var_3147_to_fp16 = const()[name = tensor("op_3147_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_101_cast_fp16 = layer_norm(axes = out_101_axes_0, epsilon = var_3147_to_fp16, x = inputs_101_cast_fp16)[name = tensor("out_101_cast_fp16")]; + tensor obj_101_gamma_0_to_fp16 = const()[name = tensor("obj_101_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998496192)))]; + tensor obj_101_beta_0_to_fp16 = const()[name = tensor("obj_101_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998498816)))]; + tensor obj_101_epsilon_0_to_fp16 = const()[name = tensor("obj_101_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_101_cast_fp16 = batch_norm(beta = obj_101_beta_0_to_fp16, epsilon = obj_101_epsilon_0_to_fp16, gamma = obj_101_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_101_cast_fp16)[name = tensor("obj_101_cast_fp16")]; + tensor query_51_pad_type_0 = const()[name = tensor("query_51_pad_type_0"), val = tensor("valid")]; + tensor query_51_strides_0 = const()[name = tensor("query_51_strides_0"), val = tensor([1, 1])]; + tensor query_51_pad_0 = const()[name = tensor("query_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_51_dilations_0 = const()[name = tensor("query_51_dilations_0"), val = tensor([1, 1])]; + tensor query_51_groups_0 = const()[name = tensor("query_51_groups_0"), val = tensor(1)]; + tensor layers_25_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_25_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998501440)))]; + tensor layers_25_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_25_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1001778304)))]; + tensor query_51_cast_fp16 = conv(bias = layers_25_self_attn_q_proj_bias_to_fp16, dilations = query_51_dilations_0, groups = query_51_groups_0, pad = query_51_pad_0, pad_type = query_51_pad_type_0, strides = query_51_strides_0, weight = layers_25_self_attn_q_proj_weight_to_fp16, x = obj_101_cast_fp16)[name = tensor("query_51_cast_fp16")]; + tensor key_51_pad_type_0 = const()[name = tensor("key_51_pad_type_0"), val = tensor("valid")]; + tensor key_51_strides_0 = const()[name = tensor("key_51_strides_0"), val = tensor([1, 1])]; + tensor key_51_pad_0 = const()[name = tensor("key_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_51_dilations_0 = const()[name = tensor("key_51_dilations_0"), val = tensor([1, 1])]; + tensor key_51_groups_0 = const()[name = tensor("key_51_groups_0"), val = tensor(1)]; + tensor layers_25_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_25_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1001780928)))]; + tensor key_51_cast_fp16 = conv(dilations = key_51_dilations_0, groups = key_51_groups_0, pad = key_51_pad_0, pad_type = key_51_pad_type_0, strides = key_51_strides_0, weight = layers_25_self_attn_k_proj_weight_to_fp16, x = obj_101_cast_fp16)[name = tensor("key_51_cast_fp16")]; + tensor value_51_pad_type_0 = const()[name = tensor("value_51_pad_type_0"), val = tensor("valid")]; + tensor value_51_strides_0 = const()[name = tensor("value_51_strides_0"), val = tensor([1, 1])]; + tensor value_51_pad_0 = const()[name = tensor("value_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_51_dilations_0 = const()[name = tensor("value_51_dilations_0"), val = tensor([1, 1])]; + tensor value_51_groups_0 = const()[name = tensor("value_51_groups_0"), val = tensor(1)]; + tensor layers_25_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_25_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1005057792)))]; + tensor layers_25_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_25_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1008334656)))]; + tensor value_51_cast_fp16 = conv(bias = layers_25_self_attn_v_proj_bias_to_fp16, dilations = value_51_dilations_0, groups = value_51_groups_0, pad = value_51_pad_0, pad_type = value_51_pad_type_0, strides = value_51_strides_0, weight = layers_25_self_attn_v_proj_weight_to_fp16, x = obj_101_cast_fp16)[name = tensor("value_51_cast_fp16")]; + tensor var_3182 = const()[name = tensor("op_3182"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_51_cast_fp16 = reshape(shape = var_3182, x = query_51_cast_fp16)[name = tensor("mh_q_51_cast_fp16")]; + tensor var_3184_to_fp16 = const()[name = tensor("op_3184_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3185_cast_fp16 = mul(x = mh_q_51_cast_fp16, y = var_3184_to_fp16)[name = tensor("op_3185_cast_fp16")]; + tensor var_3186 = const()[name = tensor("op_3186"), val = tensor([1, 20, 64, -1])]; + tensor var_3187_cast_fp16 = reshape(shape = var_3186, x = key_51_cast_fp16)[name = tensor("op_3187_cast_fp16")]; + tensor mh_w_51_transpose_x_0 = const()[name = tensor("mh_w_51_transpose_x_0"), val = tensor(true)]; + tensor mh_w_51_transpose_y_0 = const()[name = tensor("mh_w_51_transpose_y_0"), val = tensor(false)]; + tensor mh_w_51_cast_fp16 = matmul(transpose_x = mh_w_51_transpose_x_0, transpose_y = mh_w_51_transpose_y_0, x = var_3185_cast_fp16, y = var_3187_cast_fp16)[name = tensor("mh_w_51_cast_fp16")]; + tensor var_3190_cast_fp16 = softmax(axis = var_3128, x = mh_w_51_cast_fp16)[name = tensor("op_3190_cast_fp16")]; + tensor var_3191 = const()[name = tensor("op_3191"), val = tensor([1, 20, 64, -1])]; + tensor var_3192_cast_fp16 = reshape(shape = var_3191, x = value_51_cast_fp16)[name = tensor("op_3192_cast_fp16")]; + tensor attn_51_transpose_x_0 = const()[name = tensor("attn_51_transpose_x_0"), val = tensor(false)]; + tensor attn_51_transpose_y_0 = const()[name = tensor("attn_51_transpose_y_0"), val = tensor(true)]; + tensor attn_51_cast_fp16 = matmul(transpose_x = attn_51_transpose_x_0, transpose_y = attn_51_transpose_y_0, x = var_3192_cast_fp16, y = var_3190_cast_fp16)[name = tensor("attn_51_cast_fp16")]; + tensor var_3195 = const()[name = tensor("op_3195"), val = tensor([1, 1280, 1, -1])]; + tensor input_201_cast_fp16 = reshape(shape = var_3195, x = attn_51_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor obj_103_pad_type_0 = const()[name = tensor("obj_103_pad_type_0"), val = tensor("valid")]; + tensor obj_103_strides_0 = const()[name = tensor("obj_103_strides_0"), val = tensor([1, 1])]; + tensor obj_103_pad_0 = const()[name = tensor("obj_103_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_103_dilations_0 = const()[name = tensor("obj_103_dilations_0"), val = tensor([1, 1])]; + tensor obj_103_groups_0 = const()[name = tensor("obj_103_groups_0"), val = tensor(1)]; + tensor layers_25_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_25_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1008337280)))]; + tensor layers_25_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_25_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011614144)))]; + tensor obj_103_cast_fp16 = conv(bias = layers_25_self_attn_o_proj_bias_to_fp16, dilations = obj_103_dilations_0, groups = obj_103_groups_0, pad = obj_103_pad_0, pad_type = obj_103_pad_type_0, strides = obj_103_strides_0, weight = layers_25_self_attn_o_proj_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("obj_103_cast_fp16")]; + tensor inputs_103_cast_fp16 = add(x = inputs_101_cast_fp16, y = obj_103_cast_fp16)[name = tensor("inputs_103_cast_fp16")]; + tensor out_103_axes_0 = const()[name = tensor("out_103_axes_0"), val = tensor([1])]; + tensor var_3213_to_fp16 = const()[name = tensor("op_3213_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_103_cast_fp16 = layer_norm(axes = out_103_axes_0, epsilon = var_3213_to_fp16, x = inputs_103_cast_fp16)[name = tensor("out_103_cast_fp16")]; + tensor input_203_gamma_0_to_fp16 = const()[name = tensor("input_203_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011616768)))]; + tensor input_203_beta_0_to_fp16 = const()[name = tensor("input_203_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011619392)))]; + tensor input_203_epsilon_0_to_fp16 = const()[name = tensor("input_203_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_203_cast_fp16 = batch_norm(beta = input_203_beta_0_to_fp16, epsilon = input_203_epsilon_0_to_fp16, gamma = input_203_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_103_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor input_205_pad_type_0 = const()[name = tensor("input_205_pad_type_0"), val = tensor("valid")]; + tensor input_205_strides_0 = const()[name = tensor("input_205_strides_0"), val = tensor([1, 1])]; + tensor input_205_pad_0 = const()[name = tensor("input_205_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_205_dilations_0 = const()[name = tensor("input_205_dilations_0"), val = tensor([1, 1])]; + tensor input_205_groups_0 = const()[name = tensor("input_205_groups_0"), val = tensor(1)]; + tensor layers_25_fc1_weight_to_fp16 = const()[name = tensor("layers_25_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011622016)))]; + tensor layers_25_fc1_bias_to_fp16 = const()[name = tensor("layers_25_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1024729280)))]; + tensor input_205_cast_fp16 = conv(bias = layers_25_fc1_bias_to_fp16, dilations = input_205_dilations_0, groups = input_205_groups_0, pad = input_205_pad_0, pad_type = input_205_pad_type_0, strides = input_205_strides_0, weight = layers_25_fc1_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor input_207_mode_0 = const()[name = tensor("input_207_mode_0"), val = tensor("EXACT")]; + tensor input_207_cast_fp16 = gelu(mode = input_207_mode_0, x = input_205_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor hidden_states_55_pad_type_0 = const()[name = tensor("hidden_states_55_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_55_strides_0 = const()[name = tensor("hidden_states_55_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_55_pad_0 = const()[name = tensor("hidden_states_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_55_dilations_0 = const()[name = tensor("hidden_states_55_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_55_groups_0 = const()[name = tensor("hidden_states_55_groups_0"), val = tensor(1)]; + tensor layers_25_fc2_weight_to_fp16 = const()[name = tensor("layers_25_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1024739584)))]; + tensor layers_25_fc2_bias_to_fp16 = const()[name = tensor("layers_25_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037846848)))]; + tensor hidden_states_55_cast_fp16 = conv(bias = layers_25_fc2_bias_to_fp16, dilations = hidden_states_55_dilations_0, groups = hidden_states_55_groups_0, pad = hidden_states_55_pad_0, pad_type = hidden_states_55_pad_type_0, strides = hidden_states_55_strides_0, weight = layers_25_fc2_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("hidden_states_55_cast_fp16")]; + tensor inputs_105_cast_fp16 = add(x = inputs_103_cast_fp16, y = hidden_states_55_cast_fp16)[name = tensor("inputs_105_cast_fp16")]; + tensor var_3246 = const()[name = tensor("op_3246"), val = tensor(3)]; + tensor out_105_axes_0 = const()[name = tensor("out_105_axes_0"), val = tensor([1])]; + tensor var_3265_to_fp16 = const()[name = tensor("op_3265_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_105_cast_fp16 = layer_norm(axes = out_105_axes_0, epsilon = var_3265_to_fp16, x = inputs_105_cast_fp16)[name = tensor("out_105_cast_fp16")]; + tensor obj_105_gamma_0_to_fp16 = const()[name = tensor("obj_105_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037849472)))]; + tensor obj_105_beta_0_to_fp16 = const()[name = tensor("obj_105_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037852096)))]; + tensor obj_105_epsilon_0_to_fp16 = const()[name = tensor("obj_105_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_105_cast_fp16 = batch_norm(beta = obj_105_beta_0_to_fp16, epsilon = obj_105_epsilon_0_to_fp16, gamma = obj_105_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_105_cast_fp16)[name = tensor("obj_105_cast_fp16")]; + tensor query_53_pad_type_0 = const()[name = tensor("query_53_pad_type_0"), val = tensor("valid")]; + tensor query_53_strides_0 = const()[name = tensor("query_53_strides_0"), val = tensor([1, 1])]; + tensor query_53_pad_0 = const()[name = tensor("query_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_53_dilations_0 = const()[name = tensor("query_53_dilations_0"), val = tensor([1, 1])]; + tensor query_53_groups_0 = const()[name = tensor("query_53_groups_0"), val = tensor(1)]; + tensor layers_26_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_26_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037854720)))]; + tensor layers_26_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_26_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1041131584)))]; + tensor query_53_cast_fp16 = conv(bias = layers_26_self_attn_q_proj_bias_to_fp16, dilations = query_53_dilations_0, groups = query_53_groups_0, pad = query_53_pad_0, pad_type = query_53_pad_type_0, strides = query_53_strides_0, weight = layers_26_self_attn_q_proj_weight_to_fp16, x = obj_105_cast_fp16)[name = tensor("query_53_cast_fp16")]; + tensor key_53_pad_type_0 = const()[name = tensor("key_53_pad_type_0"), val = tensor("valid")]; + tensor key_53_strides_0 = const()[name = tensor("key_53_strides_0"), val = tensor([1, 1])]; + tensor key_53_pad_0 = const()[name = tensor("key_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_53_dilations_0 = const()[name = tensor("key_53_dilations_0"), val = tensor([1, 1])]; + tensor key_53_groups_0 = const()[name = tensor("key_53_groups_0"), val = tensor(1)]; + tensor layers_26_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_26_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1041134208)))]; + tensor key_53_cast_fp16 = conv(dilations = key_53_dilations_0, groups = key_53_groups_0, pad = key_53_pad_0, pad_type = key_53_pad_type_0, strides = key_53_strides_0, weight = layers_26_self_attn_k_proj_weight_to_fp16, x = obj_105_cast_fp16)[name = tensor("key_53_cast_fp16")]; + tensor value_53_pad_type_0 = const()[name = tensor("value_53_pad_type_0"), val = tensor("valid")]; + tensor value_53_strides_0 = const()[name = tensor("value_53_strides_0"), val = tensor([1, 1])]; + tensor value_53_pad_0 = const()[name = tensor("value_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_53_dilations_0 = const()[name = tensor("value_53_dilations_0"), val = tensor([1, 1])]; + tensor value_53_groups_0 = const()[name = tensor("value_53_groups_0"), val = tensor(1)]; + tensor layers_26_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_26_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1044411072)))]; + tensor layers_26_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_26_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1047687936)))]; + tensor value_53_cast_fp16 = conv(bias = layers_26_self_attn_v_proj_bias_to_fp16, dilations = value_53_dilations_0, groups = value_53_groups_0, pad = value_53_pad_0, pad_type = value_53_pad_type_0, strides = value_53_strides_0, weight = layers_26_self_attn_v_proj_weight_to_fp16, x = obj_105_cast_fp16)[name = tensor("value_53_cast_fp16")]; + tensor var_3300 = const()[name = tensor("op_3300"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_53_cast_fp16 = reshape(shape = var_3300, x = query_53_cast_fp16)[name = tensor("mh_q_53_cast_fp16")]; + tensor var_3302_to_fp16 = const()[name = tensor("op_3302_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3303_cast_fp16 = mul(x = mh_q_53_cast_fp16, y = var_3302_to_fp16)[name = tensor("op_3303_cast_fp16")]; + tensor var_3304 = const()[name = tensor("op_3304"), val = tensor([1, 20, 64, -1])]; + tensor var_3305_cast_fp16 = reshape(shape = var_3304, x = key_53_cast_fp16)[name = tensor("op_3305_cast_fp16")]; + tensor mh_w_53_transpose_x_0 = const()[name = tensor("mh_w_53_transpose_x_0"), val = tensor(true)]; + tensor mh_w_53_transpose_y_0 = const()[name = tensor("mh_w_53_transpose_y_0"), val = tensor(false)]; + tensor mh_w_53_cast_fp16 = matmul(transpose_x = mh_w_53_transpose_x_0, transpose_y = mh_w_53_transpose_y_0, x = var_3303_cast_fp16, y = var_3305_cast_fp16)[name = tensor("mh_w_53_cast_fp16")]; + tensor var_3308_cast_fp16 = softmax(axis = var_3246, x = mh_w_53_cast_fp16)[name = tensor("op_3308_cast_fp16")]; + tensor var_3309 = const()[name = tensor("op_3309"), val = tensor([1, 20, 64, -1])]; + tensor var_3310_cast_fp16 = reshape(shape = var_3309, x = value_53_cast_fp16)[name = tensor("op_3310_cast_fp16")]; + tensor attn_53_transpose_x_0 = const()[name = tensor("attn_53_transpose_x_0"), val = tensor(false)]; + tensor attn_53_transpose_y_0 = const()[name = tensor("attn_53_transpose_y_0"), val = tensor(true)]; + tensor attn_53_cast_fp16 = matmul(transpose_x = attn_53_transpose_x_0, transpose_y = attn_53_transpose_y_0, x = var_3310_cast_fp16, y = var_3308_cast_fp16)[name = tensor("attn_53_cast_fp16")]; + tensor var_3313 = const()[name = tensor("op_3313"), val = tensor([1, 1280, 1, -1])]; + tensor input_209_cast_fp16 = reshape(shape = var_3313, x = attn_53_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor obj_107_pad_type_0 = const()[name = tensor("obj_107_pad_type_0"), val = tensor("valid")]; + tensor obj_107_strides_0 = const()[name = tensor("obj_107_strides_0"), val = tensor([1, 1])]; + tensor obj_107_pad_0 = const()[name = tensor("obj_107_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_107_dilations_0 = const()[name = tensor("obj_107_dilations_0"), val = tensor([1, 1])]; + tensor obj_107_groups_0 = const()[name = tensor("obj_107_groups_0"), val = tensor(1)]; + tensor layers_26_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_26_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1047690560)))]; + tensor layers_26_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_26_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050967424)))]; + tensor obj_107_cast_fp16 = conv(bias = layers_26_self_attn_o_proj_bias_to_fp16, dilations = obj_107_dilations_0, groups = obj_107_groups_0, pad = obj_107_pad_0, pad_type = obj_107_pad_type_0, strides = obj_107_strides_0, weight = layers_26_self_attn_o_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("obj_107_cast_fp16")]; + tensor inputs_107_cast_fp16 = add(x = inputs_105_cast_fp16, y = obj_107_cast_fp16)[name = tensor("inputs_107_cast_fp16")]; + tensor out_107_axes_0 = const()[name = tensor("out_107_axes_0"), val = tensor([1])]; + tensor var_3331_to_fp16 = const()[name = tensor("op_3331_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_107_cast_fp16 = layer_norm(axes = out_107_axes_0, epsilon = var_3331_to_fp16, x = inputs_107_cast_fp16)[name = tensor("out_107_cast_fp16")]; + tensor input_211_gamma_0_to_fp16 = const()[name = tensor("input_211_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050970048)))]; + tensor input_211_beta_0_to_fp16 = const()[name = tensor("input_211_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050972672)))]; + tensor input_211_epsilon_0_to_fp16 = const()[name = tensor("input_211_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_211_cast_fp16 = batch_norm(beta = input_211_beta_0_to_fp16, epsilon = input_211_epsilon_0_to_fp16, gamma = input_211_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_107_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor input_213_pad_type_0 = const()[name = tensor("input_213_pad_type_0"), val = tensor("valid")]; + tensor input_213_strides_0 = const()[name = tensor("input_213_strides_0"), val = tensor([1, 1])]; + tensor input_213_pad_0 = const()[name = tensor("input_213_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_213_dilations_0 = const()[name = tensor("input_213_dilations_0"), val = tensor([1, 1])]; + tensor input_213_groups_0 = const()[name = tensor("input_213_groups_0"), val = tensor(1)]; + tensor layers_26_fc1_weight_to_fp16 = const()[name = tensor("layers_26_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050975296)))]; + tensor layers_26_fc1_bias_to_fp16 = const()[name = tensor("layers_26_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1064082560)))]; + tensor input_213_cast_fp16 = conv(bias = layers_26_fc1_bias_to_fp16, dilations = input_213_dilations_0, groups = input_213_groups_0, pad = input_213_pad_0, pad_type = input_213_pad_type_0, strides = input_213_strides_0, weight = layers_26_fc1_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor input_215_mode_0 = const()[name = tensor("input_215_mode_0"), val = tensor("EXACT")]; + tensor input_215_cast_fp16 = gelu(mode = input_215_mode_0, x = input_213_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor hidden_states_57_pad_type_0 = const()[name = tensor("hidden_states_57_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_57_strides_0 = const()[name = tensor("hidden_states_57_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_57_pad_0 = const()[name = tensor("hidden_states_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_57_dilations_0 = const()[name = tensor("hidden_states_57_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_57_groups_0 = const()[name = tensor("hidden_states_57_groups_0"), val = tensor(1)]; + tensor layers_26_fc2_weight_to_fp16 = const()[name = tensor("layers_26_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1064092864)))]; + tensor layers_26_fc2_bias_to_fp16 = const()[name = tensor("layers_26_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1077200128)))]; + tensor hidden_states_57_cast_fp16 = conv(bias = layers_26_fc2_bias_to_fp16, dilations = hidden_states_57_dilations_0, groups = hidden_states_57_groups_0, pad = hidden_states_57_pad_0, pad_type = hidden_states_57_pad_type_0, strides = hidden_states_57_strides_0, weight = layers_26_fc2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("hidden_states_57_cast_fp16")]; + tensor inputs_109_cast_fp16 = add(x = inputs_107_cast_fp16, y = hidden_states_57_cast_fp16)[name = tensor("inputs_109_cast_fp16")]; + tensor var_3364 = const()[name = tensor("op_3364"), val = tensor(3)]; + tensor out_109_axes_0 = const()[name = tensor("out_109_axes_0"), val = tensor([1])]; + tensor var_3383_to_fp16 = const()[name = tensor("op_3383_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_109_cast_fp16 = layer_norm(axes = out_109_axes_0, epsilon = var_3383_to_fp16, x = inputs_109_cast_fp16)[name = tensor("out_109_cast_fp16")]; + tensor obj_109_gamma_0_to_fp16 = const()[name = tensor("obj_109_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1077202752)))]; + tensor obj_109_beta_0_to_fp16 = const()[name = tensor("obj_109_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1077205376)))]; + tensor obj_109_epsilon_0_to_fp16 = const()[name = tensor("obj_109_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_109_cast_fp16 = batch_norm(beta = obj_109_beta_0_to_fp16, epsilon = obj_109_epsilon_0_to_fp16, gamma = obj_109_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_109_cast_fp16)[name = tensor("obj_109_cast_fp16")]; + tensor query_55_pad_type_0 = const()[name = tensor("query_55_pad_type_0"), val = tensor("valid")]; + tensor query_55_strides_0 = const()[name = tensor("query_55_strides_0"), val = tensor([1, 1])]; + tensor query_55_pad_0 = const()[name = tensor("query_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_55_dilations_0 = const()[name = tensor("query_55_dilations_0"), val = tensor([1, 1])]; + tensor query_55_groups_0 = const()[name = tensor("query_55_groups_0"), val = tensor(1)]; + tensor layers_27_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_27_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1077208000)))]; + tensor layers_27_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_27_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1080484864)))]; + tensor query_55_cast_fp16 = conv(bias = layers_27_self_attn_q_proj_bias_to_fp16, dilations = query_55_dilations_0, groups = query_55_groups_0, pad = query_55_pad_0, pad_type = query_55_pad_type_0, strides = query_55_strides_0, weight = layers_27_self_attn_q_proj_weight_to_fp16, x = obj_109_cast_fp16)[name = tensor("query_55_cast_fp16")]; + tensor key_55_pad_type_0 = const()[name = tensor("key_55_pad_type_0"), val = tensor("valid")]; + tensor key_55_strides_0 = const()[name = tensor("key_55_strides_0"), val = tensor([1, 1])]; + tensor key_55_pad_0 = const()[name = tensor("key_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_55_dilations_0 = const()[name = tensor("key_55_dilations_0"), val = tensor([1, 1])]; + tensor key_55_groups_0 = const()[name = tensor("key_55_groups_0"), val = tensor(1)]; + tensor layers_27_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_27_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1080487488)))]; + tensor key_55_cast_fp16 = conv(dilations = key_55_dilations_0, groups = key_55_groups_0, pad = key_55_pad_0, pad_type = key_55_pad_type_0, strides = key_55_strides_0, weight = layers_27_self_attn_k_proj_weight_to_fp16, x = obj_109_cast_fp16)[name = tensor("key_55_cast_fp16")]; + tensor value_55_pad_type_0 = const()[name = tensor("value_55_pad_type_0"), val = tensor("valid")]; + tensor value_55_strides_0 = const()[name = tensor("value_55_strides_0"), val = tensor([1, 1])]; + tensor value_55_pad_0 = const()[name = tensor("value_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_55_dilations_0 = const()[name = tensor("value_55_dilations_0"), val = tensor([1, 1])]; + tensor value_55_groups_0 = const()[name = tensor("value_55_groups_0"), val = tensor(1)]; + tensor layers_27_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_27_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1083764352)))]; + tensor layers_27_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_27_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1087041216)))]; + tensor value_55_cast_fp16 = conv(bias = layers_27_self_attn_v_proj_bias_to_fp16, dilations = value_55_dilations_0, groups = value_55_groups_0, pad = value_55_pad_0, pad_type = value_55_pad_type_0, strides = value_55_strides_0, weight = layers_27_self_attn_v_proj_weight_to_fp16, x = obj_109_cast_fp16)[name = tensor("value_55_cast_fp16")]; + tensor var_3418 = const()[name = tensor("op_3418"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_55_cast_fp16 = reshape(shape = var_3418, x = query_55_cast_fp16)[name = tensor("mh_q_55_cast_fp16")]; + tensor var_3420_to_fp16 = const()[name = tensor("op_3420_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3421_cast_fp16 = mul(x = mh_q_55_cast_fp16, y = var_3420_to_fp16)[name = tensor("op_3421_cast_fp16")]; + tensor var_3422 = const()[name = tensor("op_3422"), val = tensor([1, 20, 64, -1])]; + tensor var_3423_cast_fp16 = reshape(shape = var_3422, x = key_55_cast_fp16)[name = tensor("op_3423_cast_fp16")]; + tensor mh_w_55_transpose_x_0 = const()[name = tensor("mh_w_55_transpose_x_0"), val = tensor(true)]; + tensor mh_w_55_transpose_y_0 = const()[name = tensor("mh_w_55_transpose_y_0"), val = tensor(false)]; + tensor mh_w_55_cast_fp16 = matmul(transpose_x = mh_w_55_transpose_x_0, transpose_y = mh_w_55_transpose_y_0, x = var_3421_cast_fp16, y = var_3423_cast_fp16)[name = tensor("mh_w_55_cast_fp16")]; + tensor var_3426_cast_fp16 = softmax(axis = var_3364, x = mh_w_55_cast_fp16)[name = tensor("op_3426_cast_fp16")]; + tensor var_3427 = const()[name = tensor("op_3427"), val = tensor([1, 20, 64, -1])]; + tensor var_3428_cast_fp16 = reshape(shape = var_3427, x = value_55_cast_fp16)[name = tensor("op_3428_cast_fp16")]; + tensor attn_55_transpose_x_0 = const()[name = tensor("attn_55_transpose_x_0"), val = tensor(false)]; + tensor attn_55_transpose_y_0 = const()[name = tensor("attn_55_transpose_y_0"), val = tensor(true)]; + tensor attn_55_cast_fp16 = matmul(transpose_x = attn_55_transpose_x_0, transpose_y = attn_55_transpose_y_0, x = var_3428_cast_fp16, y = var_3426_cast_fp16)[name = tensor("attn_55_cast_fp16")]; + tensor var_3431 = const()[name = tensor("op_3431"), val = tensor([1, 1280, 1, -1])]; + tensor input_217_cast_fp16 = reshape(shape = var_3431, x = attn_55_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor obj_111_pad_type_0 = const()[name = tensor("obj_111_pad_type_0"), val = tensor("valid")]; + tensor obj_111_strides_0 = const()[name = tensor("obj_111_strides_0"), val = tensor([1, 1])]; + tensor obj_111_pad_0 = const()[name = tensor("obj_111_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_111_dilations_0 = const()[name = tensor("obj_111_dilations_0"), val = tensor([1, 1])]; + tensor obj_111_groups_0 = const()[name = tensor("obj_111_groups_0"), val = tensor(1)]; + tensor layers_27_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_27_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1087043840)))]; + tensor layers_27_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_27_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1090320704)))]; + tensor obj_111_cast_fp16 = conv(bias = layers_27_self_attn_o_proj_bias_to_fp16, dilations = obj_111_dilations_0, groups = obj_111_groups_0, pad = obj_111_pad_0, pad_type = obj_111_pad_type_0, strides = obj_111_strides_0, weight = layers_27_self_attn_o_proj_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("obj_111_cast_fp16")]; + tensor inputs_111_cast_fp16 = add(x = inputs_109_cast_fp16, y = obj_111_cast_fp16)[name = tensor("inputs_111_cast_fp16")]; + tensor out_111_axes_0 = const()[name = tensor("out_111_axes_0"), val = tensor([1])]; + tensor var_3449_to_fp16 = const()[name = tensor("op_3449_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_111_cast_fp16 = layer_norm(axes = out_111_axes_0, epsilon = var_3449_to_fp16, x = inputs_111_cast_fp16)[name = tensor("out_111_cast_fp16")]; + tensor input_219_gamma_0_to_fp16 = const()[name = tensor("input_219_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1090323328)))]; + tensor input_219_beta_0_to_fp16 = const()[name = tensor("input_219_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1090325952)))]; + tensor input_219_epsilon_0_to_fp16 = const()[name = tensor("input_219_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_219_cast_fp16 = batch_norm(beta = input_219_beta_0_to_fp16, epsilon = input_219_epsilon_0_to_fp16, gamma = input_219_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_111_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor input_221_pad_type_0 = const()[name = tensor("input_221_pad_type_0"), val = tensor("valid")]; + tensor input_221_strides_0 = const()[name = tensor("input_221_strides_0"), val = tensor([1, 1])]; + tensor input_221_pad_0 = const()[name = tensor("input_221_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_221_dilations_0 = const()[name = tensor("input_221_dilations_0"), val = tensor([1, 1])]; + tensor input_221_groups_0 = const()[name = tensor("input_221_groups_0"), val = tensor(1)]; + tensor layers_27_fc1_weight_to_fp16 = const()[name = tensor("layers_27_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1090328576)))]; + tensor layers_27_fc1_bias_to_fp16 = const()[name = tensor("layers_27_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1103435840)))]; + tensor input_221_cast_fp16 = conv(bias = layers_27_fc1_bias_to_fp16, dilations = input_221_dilations_0, groups = input_221_groups_0, pad = input_221_pad_0, pad_type = input_221_pad_type_0, strides = input_221_strides_0, weight = layers_27_fc1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor input_223_mode_0 = const()[name = tensor("input_223_mode_0"), val = tensor("EXACT")]; + tensor input_223_cast_fp16 = gelu(mode = input_223_mode_0, x = input_221_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor hidden_states_59_pad_type_0 = const()[name = tensor("hidden_states_59_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_59_strides_0 = const()[name = tensor("hidden_states_59_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_59_pad_0 = const()[name = tensor("hidden_states_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_59_dilations_0 = const()[name = tensor("hidden_states_59_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_59_groups_0 = const()[name = tensor("hidden_states_59_groups_0"), val = tensor(1)]; + tensor layers_27_fc2_weight_to_fp16 = const()[name = tensor("layers_27_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1103446144)))]; + tensor layers_27_fc2_bias_to_fp16 = const()[name = tensor("layers_27_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116553408)))]; + tensor hidden_states_59_cast_fp16 = conv(bias = layers_27_fc2_bias_to_fp16, dilations = hidden_states_59_dilations_0, groups = hidden_states_59_groups_0, pad = hidden_states_59_pad_0, pad_type = hidden_states_59_pad_type_0, strides = hidden_states_59_strides_0, weight = layers_27_fc2_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("hidden_states_59_cast_fp16")]; + tensor inputs_113_cast_fp16 = add(x = inputs_111_cast_fp16, y = hidden_states_59_cast_fp16)[name = tensor("inputs_113_cast_fp16")]; + tensor var_3482 = const()[name = tensor("op_3482"), val = tensor(3)]; + tensor out_113_axes_0 = const()[name = tensor("out_113_axes_0"), val = tensor([1])]; + tensor var_3501_to_fp16 = const()[name = tensor("op_3501_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_113_cast_fp16 = layer_norm(axes = out_113_axes_0, epsilon = var_3501_to_fp16, x = inputs_113_cast_fp16)[name = tensor("out_113_cast_fp16")]; + tensor obj_113_gamma_0_to_fp16 = const()[name = tensor("obj_113_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116556032)))]; + tensor obj_113_beta_0_to_fp16 = const()[name = tensor("obj_113_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116558656)))]; + tensor obj_113_epsilon_0_to_fp16 = const()[name = tensor("obj_113_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_113_cast_fp16 = batch_norm(beta = obj_113_beta_0_to_fp16, epsilon = obj_113_epsilon_0_to_fp16, gamma = obj_113_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_113_cast_fp16)[name = tensor("obj_113_cast_fp16")]; + tensor query_57_pad_type_0 = const()[name = tensor("query_57_pad_type_0"), val = tensor("valid")]; + tensor query_57_strides_0 = const()[name = tensor("query_57_strides_0"), val = tensor([1, 1])]; + tensor query_57_pad_0 = const()[name = tensor("query_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_57_dilations_0 = const()[name = tensor("query_57_dilations_0"), val = tensor([1, 1])]; + tensor query_57_groups_0 = const()[name = tensor("query_57_groups_0"), val = tensor(1)]; + tensor layers_28_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_28_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116561280)))]; + tensor layers_28_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_28_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1119838144)))]; + tensor query_57_cast_fp16 = conv(bias = layers_28_self_attn_q_proj_bias_to_fp16, dilations = query_57_dilations_0, groups = query_57_groups_0, pad = query_57_pad_0, pad_type = query_57_pad_type_0, strides = query_57_strides_0, weight = layers_28_self_attn_q_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("query_57_cast_fp16")]; + tensor key_57_pad_type_0 = const()[name = tensor("key_57_pad_type_0"), val = tensor("valid")]; + tensor key_57_strides_0 = const()[name = tensor("key_57_strides_0"), val = tensor([1, 1])]; + tensor key_57_pad_0 = const()[name = tensor("key_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_57_dilations_0 = const()[name = tensor("key_57_dilations_0"), val = tensor([1, 1])]; + tensor key_57_groups_0 = const()[name = tensor("key_57_groups_0"), val = tensor(1)]; + tensor layers_28_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_28_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1119840768)))]; + tensor key_57_cast_fp16 = conv(dilations = key_57_dilations_0, groups = key_57_groups_0, pad = key_57_pad_0, pad_type = key_57_pad_type_0, strides = key_57_strides_0, weight = layers_28_self_attn_k_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("key_57_cast_fp16")]; + tensor value_57_pad_type_0 = const()[name = tensor("value_57_pad_type_0"), val = tensor("valid")]; + tensor value_57_strides_0 = const()[name = tensor("value_57_strides_0"), val = tensor([1, 1])]; + tensor value_57_pad_0 = const()[name = tensor("value_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_57_dilations_0 = const()[name = tensor("value_57_dilations_0"), val = tensor([1, 1])]; + tensor value_57_groups_0 = const()[name = tensor("value_57_groups_0"), val = tensor(1)]; + tensor layers_28_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_28_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1123117632)))]; + tensor layers_28_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_28_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1126394496)))]; + tensor value_57_cast_fp16 = conv(bias = layers_28_self_attn_v_proj_bias_to_fp16, dilations = value_57_dilations_0, groups = value_57_groups_0, pad = value_57_pad_0, pad_type = value_57_pad_type_0, strides = value_57_strides_0, weight = layers_28_self_attn_v_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("value_57_cast_fp16")]; + tensor var_3536 = const()[name = tensor("op_3536"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_57_cast_fp16 = reshape(shape = var_3536, x = query_57_cast_fp16)[name = tensor("mh_q_57_cast_fp16")]; + tensor var_3538_to_fp16 = const()[name = tensor("op_3538_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3539_cast_fp16 = mul(x = mh_q_57_cast_fp16, y = var_3538_to_fp16)[name = tensor("op_3539_cast_fp16")]; + tensor var_3540 = const()[name = tensor("op_3540"), val = tensor([1, 20, 64, -1])]; + tensor var_3541_cast_fp16 = reshape(shape = var_3540, x = key_57_cast_fp16)[name = tensor("op_3541_cast_fp16")]; + tensor mh_w_57_transpose_x_0 = const()[name = tensor("mh_w_57_transpose_x_0"), val = tensor(true)]; + tensor mh_w_57_transpose_y_0 = const()[name = tensor("mh_w_57_transpose_y_0"), val = tensor(false)]; + tensor mh_w_57_cast_fp16 = matmul(transpose_x = mh_w_57_transpose_x_0, transpose_y = mh_w_57_transpose_y_0, x = var_3539_cast_fp16, y = var_3541_cast_fp16)[name = tensor("mh_w_57_cast_fp16")]; + tensor var_3544_cast_fp16 = softmax(axis = var_3482, x = mh_w_57_cast_fp16)[name = tensor("op_3544_cast_fp16")]; + tensor var_3545 = const()[name = tensor("op_3545"), val = tensor([1, 20, 64, -1])]; + tensor var_3546_cast_fp16 = reshape(shape = var_3545, x = value_57_cast_fp16)[name = tensor("op_3546_cast_fp16")]; + tensor attn_57_transpose_x_0 = const()[name = tensor("attn_57_transpose_x_0"), val = tensor(false)]; + tensor attn_57_transpose_y_0 = const()[name = tensor("attn_57_transpose_y_0"), val = tensor(true)]; + tensor attn_57_cast_fp16 = matmul(transpose_x = attn_57_transpose_x_0, transpose_y = attn_57_transpose_y_0, x = var_3546_cast_fp16, y = var_3544_cast_fp16)[name = tensor("attn_57_cast_fp16")]; + tensor var_3549 = const()[name = tensor("op_3549"), val = tensor([1, 1280, 1, -1])]; + tensor input_225_cast_fp16 = reshape(shape = var_3549, x = attn_57_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor obj_115_pad_type_0 = const()[name = tensor("obj_115_pad_type_0"), val = tensor("valid")]; + tensor obj_115_strides_0 = const()[name = tensor("obj_115_strides_0"), val = tensor([1, 1])]; + tensor obj_115_pad_0 = const()[name = tensor("obj_115_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_115_dilations_0 = const()[name = tensor("obj_115_dilations_0"), val = tensor([1, 1])]; + tensor obj_115_groups_0 = const()[name = tensor("obj_115_groups_0"), val = tensor(1)]; + tensor layers_28_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_28_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1126397120)))]; + tensor layers_28_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_28_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129673984)))]; + tensor obj_115_cast_fp16 = conv(bias = layers_28_self_attn_o_proj_bias_to_fp16, dilations = obj_115_dilations_0, groups = obj_115_groups_0, pad = obj_115_pad_0, pad_type = obj_115_pad_type_0, strides = obj_115_strides_0, weight = layers_28_self_attn_o_proj_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("obj_115_cast_fp16")]; + tensor inputs_115_cast_fp16 = add(x = inputs_113_cast_fp16, y = obj_115_cast_fp16)[name = tensor("inputs_115_cast_fp16")]; + tensor out_115_axes_0 = const()[name = tensor("out_115_axes_0"), val = tensor([1])]; + tensor var_3567_to_fp16 = const()[name = tensor("op_3567_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_115_cast_fp16 = layer_norm(axes = out_115_axes_0, epsilon = var_3567_to_fp16, x = inputs_115_cast_fp16)[name = tensor("out_115_cast_fp16")]; + tensor input_227_gamma_0_to_fp16 = const()[name = tensor("input_227_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129676608)))]; + tensor input_227_beta_0_to_fp16 = const()[name = tensor("input_227_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129679232)))]; + tensor input_227_epsilon_0_to_fp16 = const()[name = tensor("input_227_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_227_cast_fp16 = batch_norm(beta = input_227_beta_0_to_fp16, epsilon = input_227_epsilon_0_to_fp16, gamma = input_227_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_115_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor input_229_pad_type_0 = const()[name = tensor("input_229_pad_type_0"), val = tensor("valid")]; + tensor input_229_strides_0 = const()[name = tensor("input_229_strides_0"), val = tensor([1, 1])]; + tensor input_229_pad_0 = const()[name = tensor("input_229_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_229_dilations_0 = const()[name = tensor("input_229_dilations_0"), val = tensor([1, 1])]; + tensor input_229_groups_0 = const()[name = tensor("input_229_groups_0"), val = tensor(1)]; + tensor layers_28_fc1_weight_to_fp16 = const()[name = tensor("layers_28_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129681856)))]; + tensor layers_28_fc1_bias_to_fp16 = const()[name = tensor("layers_28_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1142789120)))]; + tensor input_229_cast_fp16 = conv(bias = layers_28_fc1_bias_to_fp16, dilations = input_229_dilations_0, groups = input_229_groups_0, pad = input_229_pad_0, pad_type = input_229_pad_type_0, strides = input_229_strides_0, weight = layers_28_fc1_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor input_231_mode_0 = const()[name = tensor("input_231_mode_0"), val = tensor("EXACT")]; + tensor input_231_cast_fp16 = gelu(mode = input_231_mode_0, x = input_229_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor hidden_states_61_pad_type_0 = const()[name = tensor("hidden_states_61_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_61_strides_0 = const()[name = tensor("hidden_states_61_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_61_pad_0 = const()[name = tensor("hidden_states_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_61_dilations_0 = const()[name = tensor("hidden_states_61_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_61_groups_0 = const()[name = tensor("hidden_states_61_groups_0"), val = tensor(1)]; + tensor layers_28_fc2_weight_to_fp16 = const()[name = tensor("layers_28_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1142799424)))]; + tensor layers_28_fc2_bias_to_fp16 = const()[name = tensor("layers_28_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155906688)))]; + tensor hidden_states_61_cast_fp16 = conv(bias = layers_28_fc2_bias_to_fp16, dilations = hidden_states_61_dilations_0, groups = hidden_states_61_groups_0, pad = hidden_states_61_pad_0, pad_type = hidden_states_61_pad_type_0, strides = hidden_states_61_strides_0, weight = layers_28_fc2_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("hidden_states_61_cast_fp16")]; + tensor inputs_117_cast_fp16 = add(x = inputs_115_cast_fp16, y = hidden_states_61_cast_fp16)[name = tensor("inputs_117_cast_fp16")]; + tensor var_3600 = const()[name = tensor("op_3600"), val = tensor(3)]; + tensor out_117_axes_0 = const()[name = tensor("out_117_axes_0"), val = tensor([1])]; + tensor var_3619_to_fp16 = const()[name = tensor("op_3619_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_117_cast_fp16 = layer_norm(axes = out_117_axes_0, epsilon = var_3619_to_fp16, x = inputs_117_cast_fp16)[name = tensor("out_117_cast_fp16")]; + tensor obj_117_gamma_0_to_fp16 = const()[name = tensor("obj_117_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155909312)))]; + tensor obj_117_beta_0_to_fp16 = const()[name = tensor("obj_117_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155911936)))]; + tensor obj_117_epsilon_0_to_fp16 = const()[name = tensor("obj_117_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_117_cast_fp16 = batch_norm(beta = obj_117_beta_0_to_fp16, epsilon = obj_117_epsilon_0_to_fp16, gamma = obj_117_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_117_cast_fp16)[name = tensor("obj_117_cast_fp16")]; + tensor query_59_pad_type_0 = const()[name = tensor("query_59_pad_type_0"), val = tensor("valid")]; + tensor query_59_strides_0 = const()[name = tensor("query_59_strides_0"), val = tensor([1, 1])]; + tensor query_59_pad_0 = const()[name = tensor("query_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_59_dilations_0 = const()[name = tensor("query_59_dilations_0"), val = tensor([1, 1])]; + tensor query_59_groups_0 = const()[name = tensor("query_59_groups_0"), val = tensor(1)]; + tensor layers_29_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_29_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155914560)))]; + tensor layers_29_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_29_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1159191424)))]; + tensor query_59_cast_fp16 = conv(bias = layers_29_self_attn_q_proj_bias_to_fp16, dilations = query_59_dilations_0, groups = query_59_groups_0, pad = query_59_pad_0, pad_type = query_59_pad_type_0, strides = query_59_strides_0, weight = layers_29_self_attn_q_proj_weight_to_fp16, x = obj_117_cast_fp16)[name = tensor("query_59_cast_fp16")]; + tensor key_59_pad_type_0 = const()[name = tensor("key_59_pad_type_0"), val = tensor("valid")]; + tensor key_59_strides_0 = const()[name = tensor("key_59_strides_0"), val = tensor([1, 1])]; + tensor key_59_pad_0 = const()[name = tensor("key_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_59_dilations_0 = const()[name = tensor("key_59_dilations_0"), val = tensor([1, 1])]; + tensor key_59_groups_0 = const()[name = tensor("key_59_groups_0"), val = tensor(1)]; + tensor layers_29_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_29_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1159194048)))]; + tensor key_59_cast_fp16 = conv(dilations = key_59_dilations_0, groups = key_59_groups_0, pad = key_59_pad_0, pad_type = key_59_pad_type_0, strides = key_59_strides_0, weight = layers_29_self_attn_k_proj_weight_to_fp16, x = obj_117_cast_fp16)[name = tensor("key_59_cast_fp16")]; + tensor value_59_pad_type_0 = const()[name = tensor("value_59_pad_type_0"), val = tensor("valid")]; + tensor value_59_strides_0 = const()[name = tensor("value_59_strides_0"), val = tensor([1, 1])]; + tensor value_59_pad_0 = const()[name = tensor("value_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_59_dilations_0 = const()[name = tensor("value_59_dilations_0"), val = tensor([1, 1])]; + tensor value_59_groups_0 = const()[name = tensor("value_59_groups_0"), val = tensor(1)]; + tensor layers_29_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_29_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1162470912)))]; + tensor layers_29_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_29_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1165747776)))]; + tensor value_59_cast_fp16 = conv(bias = layers_29_self_attn_v_proj_bias_to_fp16, dilations = value_59_dilations_0, groups = value_59_groups_0, pad = value_59_pad_0, pad_type = value_59_pad_type_0, strides = value_59_strides_0, weight = layers_29_self_attn_v_proj_weight_to_fp16, x = obj_117_cast_fp16)[name = tensor("value_59_cast_fp16")]; + tensor var_3654 = const()[name = tensor("op_3654"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_59_cast_fp16 = reshape(shape = var_3654, x = query_59_cast_fp16)[name = tensor("mh_q_59_cast_fp16")]; + tensor var_3656_to_fp16 = const()[name = tensor("op_3656_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3657_cast_fp16 = mul(x = mh_q_59_cast_fp16, y = var_3656_to_fp16)[name = tensor("op_3657_cast_fp16")]; + tensor var_3658 = const()[name = tensor("op_3658"), val = tensor([1, 20, 64, -1])]; + tensor var_3659_cast_fp16 = reshape(shape = var_3658, x = key_59_cast_fp16)[name = tensor("op_3659_cast_fp16")]; + tensor mh_w_59_transpose_x_0 = const()[name = tensor("mh_w_59_transpose_x_0"), val = tensor(true)]; + tensor mh_w_59_transpose_y_0 = const()[name = tensor("mh_w_59_transpose_y_0"), val = tensor(false)]; + tensor mh_w_59_cast_fp16 = matmul(transpose_x = mh_w_59_transpose_x_0, transpose_y = mh_w_59_transpose_y_0, x = var_3657_cast_fp16, y = var_3659_cast_fp16)[name = tensor("mh_w_59_cast_fp16")]; + tensor var_3662_cast_fp16 = softmax(axis = var_3600, x = mh_w_59_cast_fp16)[name = tensor("op_3662_cast_fp16")]; + tensor var_3663 = const()[name = tensor("op_3663"), val = tensor([1, 20, 64, -1])]; + tensor var_3664_cast_fp16 = reshape(shape = var_3663, x = value_59_cast_fp16)[name = tensor("op_3664_cast_fp16")]; + tensor attn_59_transpose_x_0 = const()[name = tensor("attn_59_transpose_x_0"), val = tensor(false)]; + tensor attn_59_transpose_y_0 = const()[name = tensor("attn_59_transpose_y_0"), val = tensor(true)]; + tensor attn_59_cast_fp16 = matmul(transpose_x = attn_59_transpose_x_0, transpose_y = attn_59_transpose_y_0, x = var_3664_cast_fp16, y = var_3662_cast_fp16)[name = tensor("attn_59_cast_fp16")]; + tensor var_3667 = const()[name = tensor("op_3667"), val = tensor([1, 1280, 1, -1])]; + tensor input_233_cast_fp16 = reshape(shape = var_3667, x = attn_59_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor obj_119_pad_type_0 = const()[name = tensor("obj_119_pad_type_0"), val = tensor("valid")]; + tensor obj_119_strides_0 = const()[name = tensor("obj_119_strides_0"), val = tensor([1, 1])]; + tensor obj_119_pad_0 = const()[name = tensor("obj_119_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_119_dilations_0 = const()[name = tensor("obj_119_dilations_0"), val = tensor([1, 1])]; + tensor obj_119_groups_0 = const()[name = tensor("obj_119_groups_0"), val = tensor(1)]; + tensor layers_29_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_29_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1165750400)))]; + tensor layers_29_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_29_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1169027264)))]; + tensor obj_119_cast_fp16 = conv(bias = layers_29_self_attn_o_proj_bias_to_fp16, dilations = obj_119_dilations_0, groups = obj_119_groups_0, pad = obj_119_pad_0, pad_type = obj_119_pad_type_0, strides = obj_119_strides_0, weight = layers_29_self_attn_o_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("obj_119_cast_fp16")]; + tensor inputs_119_cast_fp16 = add(x = inputs_117_cast_fp16, y = obj_119_cast_fp16)[name = tensor("inputs_119_cast_fp16")]; + tensor out_119_axes_0 = const()[name = tensor("out_119_axes_0"), val = tensor([1])]; + tensor var_3685_to_fp16 = const()[name = tensor("op_3685_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_119_cast_fp16 = layer_norm(axes = out_119_axes_0, epsilon = var_3685_to_fp16, x = inputs_119_cast_fp16)[name = tensor("out_119_cast_fp16")]; + tensor input_235_gamma_0_to_fp16 = const()[name = tensor("input_235_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1169029888)))]; + tensor input_235_beta_0_to_fp16 = const()[name = tensor("input_235_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1169032512)))]; + tensor input_235_epsilon_0_to_fp16 = const()[name = tensor("input_235_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_235_cast_fp16 = batch_norm(beta = input_235_beta_0_to_fp16, epsilon = input_235_epsilon_0_to_fp16, gamma = input_235_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_119_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor input_237_pad_type_0 = const()[name = tensor("input_237_pad_type_0"), val = tensor("valid")]; + tensor input_237_strides_0 = const()[name = tensor("input_237_strides_0"), val = tensor([1, 1])]; + tensor input_237_pad_0 = const()[name = tensor("input_237_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_237_dilations_0 = const()[name = tensor("input_237_dilations_0"), val = tensor([1, 1])]; + tensor input_237_groups_0 = const()[name = tensor("input_237_groups_0"), val = tensor(1)]; + tensor layers_29_fc1_weight_to_fp16 = const()[name = tensor("layers_29_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1169035136)))]; + tensor layers_29_fc1_bias_to_fp16 = const()[name = tensor("layers_29_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1182142400)))]; + tensor input_237_cast_fp16 = conv(bias = layers_29_fc1_bias_to_fp16, dilations = input_237_dilations_0, groups = input_237_groups_0, pad = input_237_pad_0, pad_type = input_237_pad_type_0, strides = input_237_strides_0, weight = layers_29_fc1_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_239_mode_0 = const()[name = tensor("input_239_mode_0"), val = tensor("EXACT")]; + tensor input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor hidden_states_63_pad_type_0 = const()[name = tensor("hidden_states_63_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_63_strides_0 = const()[name = tensor("hidden_states_63_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_63_pad_0 = const()[name = tensor("hidden_states_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_63_dilations_0 = const()[name = tensor("hidden_states_63_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_63_groups_0 = const()[name = tensor("hidden_states_63_groups_0"), val = tensor(1)]; + tensor layers_29_fc2_weight_to_fp16 = const()[name = tensor("layers_29_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1182152704)))]; + tensor layers_29_fc2_bias_to_fp16 = const()[name = tensor("layers_29_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1195259968)))]; + tensor hidden_states_63_cast_fp16 = conv(bias = layers_29_fc2_bias_to_fp16, dilations = hidden_states_63_dilations_0, groups = hidden_states_63_groups_0, pad = hidden_states_63_pad_0, pad_type = hidden_states_63_pad_type_0, strides = hidden_states_63_strides_0, weight = layers_29_fc2_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("hidden_states_63_cast_fp16")]; + tensor inputs_121_cast_fp16 = add(x = inputs_119_cast_fp16, y = hidden_states_63_cast_fp16)[name = tensor("inputs_121_cast_fp16")]; + tensor var_3718 = const()[name = tensor("op_3718"), val = tensor(3)]; + tensor out_121_axes_0 = const()[name = tensor("out_121_axes_0"), val = tensor([1])]; + tensor var_3737_to_fp16 = const()[name = tensor("op_3737_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_121_cast_fp16 = layer_norm(axes = out_121_axes_0, epsilon = var_3737_to_fp16, x = inputs_121_cast_fp16)[name = tensor("out_121_cast_fp16")]; + tensor obj_121_gamma_0_to_fp16 = const()[name = tensor("obj_121_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1195262592)))]; + tensor obj_121_beta_0_to_fp16 = const()[name = tensor("obj_121_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1195265216)))]; + tensor obj_121_epsilon_0_to_fp16 = const()[name = tensor("obj_121_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_121_cast_fp16 = batch_norm(beta = obj_121_beta_0_to_fp16, epsilon = obj_121_epsilon_0_to_fp16, gamma = obj_121_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_121_cast_fp16)[name = tensor("obj_121_cast_fp16")]; + tensor query_61_pad_type_0 = const()[name = tensor("query_61_pad_type_0"), val = tensor("valid")]; + tensor query_61_strides_0 = const()[name = tensor("query_61_strides_0"), val = tensor([1, 1])]; + tensor query_61_pad_0 = const()[name = tensor("query_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_61_dilations_0 = const()[name = tensor("query_61_dilations_0"), val = tensor([1, 1])]; + tensor query_61_groups_0 = const()[name = tensor("query_61_groups_0"), val = tensor(1)]; + tensor layers_30_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_30_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1195267840)))]; + tensor layers_30_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_30_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1198544704)))]; + tensor query_61_cast_fp16 = conv(bias = layers_30_self_attn_q_proj_bias_to_fp16, dilations = query_61_dilations_0, groups = query_61_groups_0, pad = query_61_pad_0, pad_type = query_61_pad_type_0, strides = query_61_strides_0, weight = layers_30_self_attn_q_proj_weight_to_fp16, x = obj_121_cast_fp16)[name = tensor("query_61_cast_fp16")]; + tensor key_61_pad_type_0 = const()[name = tensor("key_61_pad_type_0"), val = tensor("valid")]; + tensor key_61_strides_0 = const()[name = tensor("key_61_strides_0"), val = tensor([1, 1])]; + tensor key_61_pad_0 = const()[name = tensor("key_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_61_dilations_0 = const()[name = tensor("key_61_dilations_0"), val = tensor([1, 1])]; + tensor key_61_groups_0 = const()[name = tensor("key_61_groups_0"), val = tensor(1)]; + tensor layers_30_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_30_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1198547328)))]; + tensor key_61_cast_fp16 = conv(dilations = key_61_dilations_0, groups = key_61_groups_0, pad = key_61_pad_0, pad_type = key_61_pad_type_0, strides = key_61_strides_0, weight = layers_30_self_attn_k_proj_weight_to_fp16, x = obj_121_cast_fp16)[name = tensor("key_61_cast_fp16")]; + tensor value_61_pad_type_0 = const()[name = tensor("value_61_pad_type_0"), val = tensor("valid")]; + tensor value_61_strides_0 = const()[name = tensor("value_61_strides_0"), val = tensor([1, 1])]; + tensor value_61_pad_0 = const()[name = tensor("value_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_61_dilations_0 = const()[name = tensor("value_61_dilations_0"), val = tensor([1, 1])]; + tensor value_61_groups_0 = const()[name = tensor("value_61_groups_0"), val = tensor(1)]; + tensor layers_30_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_30_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1201824192)))]; + tensor layers_30_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_30_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1205101056)))]; + tensor value_61_cast_fp16 = conv(bias = layers_30_self_attn_v_proj_bias_to_fp16, dilations = value_61_dilations_0, groups = value_61_groups_0, pad = value_61_pad_0, pad_type = value_61_pad_type_0, strides = value_61_strides_0, weight = layers_30_self_attn_v_proj_weight_to_fp16, x = obj_121_cast_fp16)[name = tensor("value_61_cast_fp16")]; + tensor var_3772 = const()[name = tensor("op_3772"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_61_cast_fp16 = reshape(shape = var_3772, x = query_61_cast_fp16)[name = tensor("mh_q_61_cast_fp16")]; + tensor var_3774_to_fp16 = const()[name = tensor("op_3774_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3775_cast_fp16 = mul(x = mh_q_61_cast_fp16, y = var_3774_to_fp16)[name = tensor("op_3775_cast_fp16")]; + tensor var_3776 = const()[name = tensor("op_3776"), val = tensor([1, 20, 64, -1])]; + tensor var_3777_cast_fp16 = reshape(shape = var_3776, x = key_61_cast_fp16)[name = tensor("op_3777_cast_fp16")]; + tensor mh_w_61_transpose_x_0 = const()[name = tensor("mh_w_61_transpose_x_0"), val = tensor(true)]; + tensor mh_w_61_transpose_y_0 = const()[name = tensor("mh_w_61_transpose_y_0"), val = tensor(false)]; + tensor mh_w_61_cast_fp16 = matmul(transpose_x = mh_w_61_transpose_x_0, transpose_y = mh_w_61_transpose_y_0, x = var_3775_cast_fp16, y = var_3777_cast_fp16)[name = tensor("mh_w_61_cast_fp16")]; + tensor var_3780_cast_fp16 = softmax(axis = var_3718, x = mh_w_61_cast_fp16)[name = tensor("op_3780_cast_fp16")]; + tensor var_3781 = const()[name = tensor("op_3781"), val = tensor([1, 20, 64, -1])]; + tensor var_3782_cast_fp16 = reshape(shape = var_3781, x = value_61_cast_fp16)[name = tensor("op_3782_cast_fp16")]; + tensor attn_61_transpose_x_0 = const()[name = tensor("attn_61_transpose_x_0"), val = tensor(false)]; + tensor attn_61_transpose_y_0 = const()[name = tensor("attn_61_transpose_y_0"), val = tensor(true)]; + tensor attn_61_cast_fp16 = matmul(transpose_x = attn_61_transpose_x_0, transpose_y = attn_61_transpose_y_0, x = var_3782_cast_fp16, y = var_3780_cast_fp16)[name = tensor("attn_61_cast_fp16")]; + tensor var_3785 = const()[name = tensor("op_3785"), val = tensor([1, 1280, 1, -1])]; + tensor input_241_cast_fp16 = reshape(shape = var_3785, x = attn_61_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor obj_123_pad_type_0 = const()[name = tensor("obj_123_pad_type_0"), val = tensor("valid")]; + tensor obj_123_strides_0 = const()[name = tensor("obj_123_strides_0"), val = tensor([1, 1])]; + tensor obj_123_pad_0 = const()[name = tensor("obj_123_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_123_dilations_0 = const()[name = tensor("obj_123_dilations_0"), val = tensor([1, 1])]; + tensor obj_123_groups_0 = const()[name = tensor("obj_123_groups_0"), val = tensor(1)]; + tensor layers_30_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_30_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1205103680)))]; + tensor layers_30_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_30_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208380544)))]; + tensor obj_123_cast_fp16 = conv(bias = layers_30_self_attn_o_proj_bias_to_fp16, dilations = obj_123_dilations_0, groups = obj_123_groups_0, pad = obj_123_pad_0, pad_type = obj_123_pad_type_0, strides = obj_123_strides_0, weight = layers_30_self_attn_o_proj_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("obj_123_cast_fp16")]; + tensor inputs_123_cast_fp16 = add(x = inputs_121_cast_fp16, y = obj_123_cast_fp16)[name = tensor("inputs_123_cast_fp16")]; + tensor out_123_axes_0 = const()[name = tensor("out_123_axes_0"), val = tensor([1])]; + tensor var_3803_to_fp16 = const()[name = tensor("op_3803_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_123_cast_fp16 = layer_norm(axes = out_123_axes_0, epsilon = var_3803_to_fp16, x = inputs_123_cast_fp16)[name = tensor("out_123_cast_fp16")]; + tensor input_243_gamma_0_to_fp16 = const()[name = tensor("input_243_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208383168)))]; + tensor input_243_beta_0_to_fp16 = const()[name = tensor("input_243_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208385792)))]; + tensor input_243_epsilon_0_to_fp16 = const()[name = tensor("input_243_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_243_cast_fp16 = batch_norm(beta = input_243_beta_0_to_fp16, epsilon = input_243_epsilon_0_to_fp16, gamma = input_243_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_123_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor input_245_pad_type_0 = const()[name = tensor("input_245_pad_type_0"), val = tensor("valid")]; + tensor input_245_strides_0 = const()[name = tensor("input_245_strides_0"), val = tensor([1, 1])]; + tensor input_245_pad_0 = const()[name = tensor("input_245_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_245_dilations_0 = const()[name = tensor("input_245_dilations_0"), val = tensor([1, 1])]; + tensor input_245_groups_0 = const()[name = tensor("input_245_groups_0"), val = tensor(1)]; + tensor layers_30_fc1_weight_to_fp16 = const()[name = tensor("layers_30_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208388416)))]; + tensor layers_30_fc1_bias_to_fp16 = const()[name = tensor("layers_30_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1221495680)))]; + tensor input_245_cast_fp16 = conv(bias = layers_30_fc1_bias_to_fp16, dilations = input_245_dilations_0, groups = input_245_groups_0, pad = input_245_pad_0, pad_type = input_245_pad_type_0, strides = input_245_strides_0, weight = layers_30_fc1_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor input_247_mode_0 = const()[name = tensor("input_247_mode_0"), val = tensor("EXACT")]; + tensor input_247_cast_fp16 = gelu(mode = input_247_mode_0, x = input_245_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor hidden_states_65_pad_type_0 = const()[name = tensor("hidden_states_65_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_65_strides_0 = const()[name = tensor("hidden_states_65_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_65_pad_0 = const()[name = tensor("hidden_states_65_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_65_dilations_0 = const()[name = tensor("hidden_states_65_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_65_groups_0 = const()[name = tensor("hidden_states_65_groups_0"), val = tensor(1)]; + tensor layers_30_fc2_weight_to_fp16 = const()[name = tensor("layers_30_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1221505984)))]; + tensor layers_30_fc2_bias_to_fp16 = const()[name = tensor("layers_30_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234613248)))]; + tensor hidden_states_65_cast_fp16 = conv(bias = layers_30_fc2_bias_to_fp16, dilations = hidden_states_65_dilations_0, groups = hidden_states_65_groups_0, pad = hidden_states_65_pad_0, pad_type = hidden_states_65_pad_type_0, strides = hidden_states_65_strides_0, weight = layers_30_fc2_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("hidden_states_65_cast_fp16")]; + tensor inputs_125_cast_fp16 = add(x = inputs_123_cast_fp16, y = hidden_states_65_cast_fp16)[name = tensor("inputs_125_cast_fp16")]; + tensor var_3836 = const()[name = tensor("op_3836"), val = tensor(3)]; + tensor out_125_axes_0 = const()[name = tensor("out_125_axes_0"), val = tensor([1])]; + tensor var_3855_to_fp16 = const()[name = tensor("op_3855_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_125_cast_fp16 = layer_norm(axes = out_125_axes_0, epsilon = var_3855_to_fp16, x = inputs_125_cast_fp16)[name = tensor("out_125_cast_fp16")]; + tensor obj_125_gamma_0_to_fp16 = const()[name = tensor("obj_125_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234615872)))]; + tensor obj_125_beta_0_to_fp16 = const()[name = tensor("obj_125_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234618496)))]; + tensor obj_125_epsilon_0_to_fp16 = const()[name = tensor("obj_125_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_125_cast_fp16 = batch_norm(beta = obj_125_beta_0_to_fp16, epsilon = obj_125_epsilon_0_to_fp16, gamma = obj_125_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_125_cast_fp16)[name = tensor("obj_125_cast_fp16")]; + tensor query_pad_type_0 = const()[name = tensor("query_pad_type_0"), val = tensor("valid")]; + tensor query_strides_0 = const()[name = tensor("query_strides_0"), val = tensor([1, 1])]; + tensor query_pad_0 = const()[name = tensor("query_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_dilations_0 = const()[name = tensor("query_dilations_0"), val = tensor([1, 1])]; + tensor query_groups_0 = const()[name = tensor("query_groups_0"), val = tensor(1)]; + tensor layers_31_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_31_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234621120)))]; + tensor layers_31_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_31_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1237897984)))]; + tensor query_cast_fp16 = conv(bias = layers_31_self_attn_q_proj_bias_to_fp16, dilations = query_dilations_0, groups = query_groups_0, pad = query_pad_0, pad_type = query_pad_type_0, strides = query_strides_0, weight = layers_31_self_attn_q_proj_weight_to_fp16, x = obj_125_cast_fp16)[name = tensor("query_cast_fp16")]; + tensor key_pad_type_0 = const()[name = tensor("key_pad_type_0"), val = tensor("valid")]; + tensor key_strides_0 = const()[name = tensor("key_strides_0"), val = tensor([1, 1])]; + tensor key_pad_0 = const()[name = tensor("key_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_dilations_0 = const()[name = tensor("key_dilations_0"), val = tensor([1, 1])]; + tensor key_groups_0 = const()[name = tensor("key_groups_0"), val = tensor(1)]; + tensor layers_31_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_31_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1237900608)))]; + tensor key_cast_fp16 = conv(dilations = key_dilations_0, groups = key_groups_0, pad = key_pad_0, pad_type = key_pad_type_0, strides = key_strides_0, weight = layers_31_self_attn_k_proj_weight_to_fp16, x = obj_125_cast_fp16)[name = tensor("key_cast_fp16")]; + tensor value_pad_type_0 = const()[name = tensor("value_pad_type_0"), val = tensor("valid")]; + tensor value_strides_0 = const()[name = tensor("value_strides_0"), val = tensor([1, 1])]; + tensor value_pad_0 = const()[name = tensor("value_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_dilations_0 = const()[name = tensor("value_dilations_0"), val = tensor([1, 1])]; + tensor value_groups_0 = const()[name = tensor("value_groups_0"), val = tensor(1)]; + tensor layers_31_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_31_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1241177472)))]; + tensor layers_31_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_31_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1244454336)))]; + tensor value_cast_fp16 = conv(bias = layers_31_self_attn_v_proj_bias_to_fp16, dilations = value_dilations_0, groups = value_groups_0, pad = value_pad_0, pad_type = value_pad_type_0, strides = value_strides_0, weight = layers_31_self_attn_v_proj_weight_to_fp16, x = obj_125_cast_fp16)[name = tensor("value_cast_fp16")]; + tensor var_3890 = const()[name = tensor("op_3890"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_cast_fp16 = reshape(shape = var_3890, x = query_cast_fp16)[name = tensor("mh_q_cast_fp16")]; + tensor var_3892_to_fp16 = const()[name = tensor("op_3892_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3893_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_3892_to_fp16)[name = tensor("op_3893_cast_fp16")]; + tensor var_3894 = const()[name = tensor("op_3894"), val = tensor([1, 20, 64, -1])]; + tensor var_3895_cast_fp16 = reshape(shape = var_3894, x = key_cast_fp16)[name = tensor("op_3895_cast_fp16")]; + tensor mh_w_transpose_x_0 = const()[name = tensor("mh_w_transpose_x_0"), val = tensor(true)]; + tensor mh_w_transpose_y_0 = const()[name = tensor("mh_w_transpose_y_0"), val = tensor(false)]; + tensor mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_3893_cast_fp16, y = var_3895_cast_fp16)[name = tensor("mh_w_cast_fp16")]; + tensor var_3898_cast_fp16 = softmax(axis = var_3836, x = mh_w_cast_fp16)[name = tensor("op_3898_cast_fp16")]; + tensor var_3899 = const()[name = tensor("op_3899"), val = tensor([1, 20, 64, -1])]; + tensor var_3900_cast_fp16 = reshape(shape = var_3899, x = value_cast_fp16)[name = tensor("op_3900_cast_fp16")]; + tensor attn_transpose_x_0 = const()[name = tensor("attn_transpose_x_0"), val = tensor(false)]; + tensor attn_transpose_y_0 = const()[name = tensor("attn_transpose_y_0"), val = tensor(true)]; + tensor attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_3900_cast_fp16, y = var_3898_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_3903 = const()[name = tensor("op_3903"), val = tensor([1, 1280, 1, -1])]; + tensor input_249_cast_fp16 = reshape(shape = var_3903, x = attn_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor obj_pad_type_0 = const()[name = tensor("obj_pad_type_0"), val = tensor("valid")]; + tensor obj_strides_0 = const()[name = tensor("obj_strides_0"), val = tensor([1, 1])]; + tensor obj_pad_0 = const()[name = tensor("obj_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_dilations_0 = const()[name = tensor("obj_dilations_0"), val = tensor([1, 1])]; + tensor obj_groups_0 = const()[name = tensor("obj_groups_0"), val = tensor(1)]; + tensor layers_31_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_31_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1244456960)))]; + tensor layers_31_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_31_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247733824)))]; + tensor obj_cast_fp16 = conv(bias = layers_31_self_attn_o_proj_bias_to_fp16, dilations = obj_dilations_0, groups = obj_groups_0, pad = obj_pad_0, pad_type = obj_pad_type_0, strides = obj_strides_0, weight = layers_31_self_attn_o_proj_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("obj_cast_fp16")]; + tensor inputs_127_cast_fp16 = add(x = inputs_125_cast_fp16, y = obj_cast_fp16)[name = tensor("inputs_127_cast_fp16")]; + tensor out_127_axes_0 = const()[name = tensor("out_127_axes_0"), val = tensor([1])]; + tensor var_3921_to_fp16 = const()[name = tensor("op_3921_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_127_cast_fp16 = layer_norm(axes = out_127_axes_0, epsilon = var_3921_to_fp16, x = inputs_127_cast_fp16)[name = tensor("out_127_cast_fp16")]; + tensor input_251_gamma_0_to_fp16 = const()[name = tensor("input_251_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247736448)))]; + tensor input_251_beta_0_to_fp16 = const()[name = tensor("input_251_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247739072)))]; + tensor input_251_epsilon_0_to_fp16 = const()[name = tensor("input_251_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_251_cast_fp16 = batch_norm(beta = input_251_beta_0_to_fp16, epsilon = input_251_epsilon_0_to_fp16, gamma = input_251_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_127_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor input_253_pad_type_0 = const()[name = tensor("input_253_pad_type_0"), val = tensor("valid")]; + tensor input_253_strides_0 = const()[name = tensor("input_253_strides_0"), val = tensor([1, 1])]; + tensor input_253_pad_0 = const()[name = tensor("input_253_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_253_dilations_0 = const()[name = tensor("input_253_dilations_0"), val = tensor([1, 1])]; + tensor input_253_groups_0 = const()[name = tensor("input_253_groups_0"), val = tensor(1)]; + tensor layers_31_fc1_weight_to_fp16 = const()[name = tensor("layers_31_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247741696)))]; + tensor layers_31_fc1_bias_to_fp16 = const()[name = tensor("layers_31_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1260848960)))]; + tensor input_253_cast_fp16 = conv(bias = layers_31_fc1_bias_to_fp16, dilations = input_253_dilations_0, groups = input_253_groups_0, pad = input_253_pad_0, pad_type = input_253_pad_type_0, strides = input_253_strides_0, weight = layers_31_fc1_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor input_mode_0 = const()[name = tensor("input_mode_0"), val = tensor("EXACT")]; + tensor input_cast_fp16 = gelu(mode = input_mode_0, x = input_253_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor hidden_states_pad_type_0 = const()[name = tensor("hidden_states_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_strides_0 = const()[name = tensor("hidden_states_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_pad_0 = const()[name = tensor("hidden_states_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_dilations_0 = const()[name = tensor("hidden_states_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_groups_0 = const()[name = tensor("hidden_states_groups_0"), val = tensor(1)]; + tensor layers_31_fc2_weight_to_fp16 = const()[name = tensor("layers_31_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1260859264)))]; + tensor layers_31_fc2_bias_to_fp16 = const()[name = tensor("layers_31_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1273966528)))]; + tensor hidden_states_cast_fp16 = conv(bias = layers_31_fc2_bias_to_fp16, dilations = hidden_states_dilations_0, groups = hidden_states_groups_0, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = hidden_states_strides_0, weight = layers_31_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor("hidden_states_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_127_cast_fp16, y = hidden_states_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor out_axes_0 = const()[name = tensor("out_axes_0"), val = tensor([1])]; + tensor var_3959_to_fp16 = const()[name = tensor("op_3959_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_3959_to_fp16, x = inputs_cast_fp16)[name = tensor("out_cast_fp16")]; + tensor encoder_output_embeds_type_fp32_gamma_0_to_fp16 = const()[name = tensor("encoder_output_embeds_type_fp32_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1273969152)))]; + tensor encoder_output_embeds_type_fp32_beta_0_to_fp16 = const()[name = tensor("encoder_output_embeds_type_fp32_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1273971776)))]; + tensor encoder_output_embeds_type_fp32_epsilon_0_to_fp16 = const()[name = tensor("encoder_output_embeds_type_fp32_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor encoder_output_embeds = batch_norm(beta = encoder_output_embeds_type_fp32_beta_0_to_fp16, epsilon = encoder_output_embeds_type_fp32_epsilon_0_to_fp16, gamma = encoder_output_embeds_type_fp32_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = tensor("encoder_output_embeds_type_fp32_cast_fp16")]; + } -> (encoder_output_embeds); +} \ No newline at end of file