diff --git "a/ced-tiny/ced_tiny.mlmodelc/model.mil" "b/ced-tiny/ced_tiny.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/ced-tiny/ced_tiny.mlmodelc/model.mil" @@ -0,0 +1,655 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor mel) { + tensor var_11 = const()[name = tensor("op_11"), val = tensor(-1)]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([1])]; + tensor mel_to_fp16_dtype_0 = const()[name = tensor("mel_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor mel_to_fp16 = cast(dtype = mel_to_fp16_dtype_0, x = mel)[name = tensor("cast_27")]; + tensor x_1_cast_fp16 = expand_dims(axes = x_1_axes_0, x = mel_to_fp16)[name = tensor("x_1_cast_fp16")]; + tensor var_33 = const()[name = tensor("op_33"), val = tensor([0, 2, 1, 3])]; + tensor m_encoder_init_bn_running_mean_to_fp16 = const()[name = tensor("m_encoder_init_bn_running_mean_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor m_encoder_init_bn_running_var_to_fp16 = const()[name = tensor("m_encoder_init_bn_running_var_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256)))]; + tensor m_encoder_init_bn_weight_to_fp16 = const()[name = tensor("m_encoder_init_bn_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448)))]; + tensor m_encoder_init_bn_bias_to_fp16 = const()[name = tensor("m_encoder_init_bn_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(640)))]; + tensor var_20_to_fp16 = const()[name = tensor("op_20_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_1_cast_fp16 = transpose(perm = var_33, x = x_1_cast_fp16)[name = tensor("transpose_26")]; + tensor x_3_cast_fp16 = batch_norm(beta = m_encoder_init_bn_bias_to_fp16, epsilon = var_20_to_fp16, gamma = m_encoder_init_bn_weight_to_fp16, mean = m_encoder_init_bn_running_mean_to_fp16, variance = m_encoder_init_bn_running_var_to_fp16, x = input_1_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor var_40 = const()[name = tensor("op_40"), val = tensor([0, 2, 1, 3])]; + tensor x_7_pad_type_0 = const()[name = tensor("x_7_pad_type_0"), val = tensor("valid")]; + tensor x_7_strides_0 = const()[name = tensor("x_7_strides_0"), val = tensor([16, 16])]; + tensor x_7_pad_0 = const()[name = tensor("x_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_7_dilations_0 = const()[name = tensor("x_7_dilations_0"), val = tensor([1, 1])]; + tensor x_7_groups_0 = const()[name = tensor("x_7_groups_0"), val = tensor(1)]; + tensor m_encoder_patch_embed_proj_weight_to_fp16 = const()[name = tensor("m_encoder_patch_embed_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(832)))]; + tensor m_encoder_patch_embed_proj_bias_to_fp16 = const()[name = tensor("m_encoder_patch_embed_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99200)))]; + tensor x_5_cast_fp16 = transpose(perm = var_40, x = x_3_cast_fp16)[name = tensor("transpose_25")]; + tensor x_7_cast_fp16 = conv(bias = m_encoder_patch_embed_proj_bias_to_fp16, dilations = x_7_dilations_0, groups = x_7_groups_0, pad = x_7_pad_0, pad_type = x_7_pad_type_0, strides = x_7_strides_0, weight = m_encoder_patch_embed_proj_weight_to_fp16, x = x_5_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor var_54_to_fp16 = const()[name = tensor("op_54_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99648)))]; + tensor x_9_cast_fp16 = add(x = x_7_cast_fp16, y = var_54_to_fp16)[name = tensor("x_9_cast_fp16")]; + tensor m_encoder_freq_pos_embed_to_fp16 = const()[name = tensor("m_encoder_freq_pos_embed_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(123520)))]; + tensor x_11_cast_fp16 = add(x = x_9_cast_fp16, y = m_encoder_freq_pos_embed_to_fp16)[name = tensor("x_11_cast_fp16")]; + tensor concat_0 = const()[name = tensor("concat_0"), val = tensor([1, 192, 248])]; + tensor var_61_cast_fp16 = reshape(shape = concat_0, x = x_11_cast_fp16)[name = tensor("op_61_cast_fp16")]; + tensor var_62 = const()[name = tensor("op_62"), val = tensor([0, 2, 1])]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor m_encoder_blocks_0_norm1_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_0_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125120)))]; + tensor m_encoder_blocks_0_norm1_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_0_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125568)))]; + tensor var_14_to_fp16 = const()[name = tensor("op_14_to_fp16"), val = tensor(0x1.1p-20)]; + tensor input_3_cast_fp16 = transpose(perm = var_62, x = var_61_cast_fp16)[name = tensor("transpose_24")]; + tensor x_13_cast_fp16 = layer_norm(axes = x_13_axes_0, beta = m_encoder_blocks_0_norm1_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_blocks_0_norm1_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor m_encoder_blocks_0_attn_qkv_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_0_attn_qkv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126016)))]; + tensor m_encoder_blocks_0_attn_qkv_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_0_attn_qkv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(347264)))]; + tensor linear_0_cast_fp16 = linear(bias = m_encoder_blocks_0_attn_qkv_bias_to_fp16, weight = m_encoder_blocks_0_attn_qkv_weight_to_fp16, x = x_13_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_96 = const()[name = tensor("op_96"), val = tensor([1, 248, 3, 3, 64])]; + tensor var_97_cast_fp16 = reshape(shape = var_96, x = linear_0_cast_fp16)[name = tensor("op_97_cast_fp16")]; + tensor var_98 = const()[name = tensor("op_98"), val = tensor([2, 0, 3, 1, 4])]; + tensor var_100_split_sizes_0 = const()[name = tensor("op_100_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_100_axis_0 = const()[name = tensor("op_100_axis_0"), val = tensor(0)]; + tensor qkv_1_cast_fp16 = transpose(perm = var_98, x = var_97_cast_fp16)[name = tensor("transpose_23")]; + tensor var_100_cast_fp16_0, tensor var_100_cast_fp16_1, tensor var_100_cast_fp16_2 = split(axis = var_100_axis_0, split_sizes = var_100_split_sizes_0, x = qkv_1_cast_fp16)[name = tensor("op_100_cast_fp16")]; + tensor squeeze_0_axes_0 = const()[name = tensor("squeeze_0_axes_0"), val = tensor([0])]; + tensor squeeze_0_cast_fp16 = squeeze(axes = squeeze_0_axes_0, x = var_100_cast_fp16_0)[name = tensor("squeeze_0_cast_fp16")]; + tensor squeeze_1_axes_0 = const()[name = tensor("squeeze_1_axes_0"), val = tensor([0])]; + tensor squeeze_1_cast_fp16 = squeeze(axes = squeeze_1_axes_0, x = var_100_cast_fp16_1)[name = tensor("squeeze_1_cast_fp16")]; + tensor squeeze_2_axes_0 = const()[name = tensor("squeeze_2_axes_0"), val = tensor([0])]; + tensor squeeze_2_cast_fp16 = squeeze(axes = squeeze_2_axes_0, x = var_100_cast_fp16_2)[name = tensor("squeeze_2_cast_fp16")]; + tensor var_105_transpose_x_1 = const()[name = tensor("op_105_transpose_x_1"), val = tensor(false)]; + tensor var_105_transpose_y_1 = const()[name = tensor("op_105_transpose_y_1"), val = tensor(true)]; + tensor var_105_cast_fp16 = matmul(transpose_x = var_105_transpose_x_1, transpose_y = var_105_transpose_y_1, x = squeeze_0_cast_fp16, y = squeeze_1_cast_fp16)[name = tensor("op_105_cast_fp16")]; + tensor var_106_to_fp16 = const()[name = tensor("op_106_to_fp16"), val = tensor(0x1p-3)]; + tensor attn_1_cast_fp16 = mul(x = var_105_cast_fp16, y = var_106_to_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor input_7_cast_fp16 = softmax(axis = var_11, x = attn_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor var_110_transpose_x_0 = const()[name = tensor("op_110_transpose_x_0"), val = tensor(false)]; + tensor var_110_transpose_y_0 = const()[name = tensor("op_110_transpose_y_0"), val = tensor(false)]; + tensor var_110_cast_fp16 = matmul(transpose_x = var_110_transpose_x_0, transpose_y = var_110_transpose_y_0, x = input_7_cast_fp16, y = squeeze_2_cast_fp16)[name = tensor("op_110_cast_fp16")]; + tensor var_111_perm_0 = const()[name = tensor("op_111_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_112 = const()[name = tensor("op_112"), val = tensor([1, 248, 192])]; + tensor var_111_cast_fp16 = transpose(perm = var_111_perm_0, x = var_110_cast_fp16)[name = tensor("transpose_22")]; + tensor input_9_cast_fp16 = reshape(shape = var_112, x = var_111_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor m_encoder_blocks_0_attn_proj_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_0_attn_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(348480)))]; + tensor m_encoder_blocks_0_attn_proj_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_0_attn_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(422272)))]; + tensor linear_1_cast_fp16 = linear(bias = m_encoder_blocks_0_attn_proj_bias_to_fp16, weight = m_encoder_blocks_0_attn_proj_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_3_cast_fp16, y = linear_1_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_axes_0 = const()[name = tensor("input_15_axes_0"), val = tensor([-1])]; + tensor m_encoder_blocks_0_norm2_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_0_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(422720)))]; + tensor m_encoder_blocks_0_norm2_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_0_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(423168)))]; + tensor input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = m_encoder_blocks_0_norm2_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_blocks_0_norm2_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor m_encoder_blocks_0_mlp_fc1_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_0_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(423616)))]; + tensor m_encoder_blocks_0_mlp_fc1_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_0_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(718592)))]; + tensor linear_2_cast_fp16 = linear(bias = m_encoder_blocks_0_mlp_fc1_bias_to_fp16, weight = m_encoder_blocks_0_mlp_fc1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor input_19_mode_0 = const()[name = tensor("input_19_mode_0"), val = tensor("EXACT")]; + tensor input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = linear_2_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor m_encoder_blocks_0_mlp_fc2_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_0_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(720192)))]; + tensor m_encoder_blocks_0_mlp_fc2_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_0_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1015168)))]; + tensor linear_3_cast_fp16 = linear(bias = m_encoder_blocks_0_mlp_fc2_bias_to_fp16, weight = m_encoder_blocks_0_mlp_fc2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = input_13_cast_fp16, y = linear_3_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor x_15_axes_0 = const()[name = tensor("x_15_axes_0"), val = tensor([-1])]; + tensor m_encoder_blocks_1_norm1_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_1_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1015616)))]; + tensor m_encoder_blocks_1_norm1_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_1_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1016064)))]; + tensor x_15_cast_fp16 = layer_norm(axes = x_15_axes_0, beta = m_encoder_blocks_1_norm1_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_blocks_1_norm1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor m_encoder_blocks_1_attn_qkv_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_1_attn_qkv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1016512)))]; + tensor m_encoder_blocks_1_attn_qkv_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_1_attn_qkv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1237760)))]; + tensor linear_4_cast_fp16 = linear(bias = m_encoder_blocks_1_attn_qkv_bias_to_fp16, weight = m_encoder_blocks_1_attn_qkv_weight_to_fp16, x = x_15_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor var_154 = const()[name = tensor("op_154"), val = tensor([1, 248, 3, 3, 64])]; + tensor var_155_cast_fp16 = reshape(shape = var_154, x = linear_4_cast_fp16)[name = tensor("op_155_cast_fp16")]; + tensor var_156 = const()[name = tensor("op_156"), val = tensor([2, 0, 3, 1, 4])]; + tensor var_158_split_sizes_0 = const()[name = tensor("op_158_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_158_axis_0 = const()[name = tensor("op_158_axis_0"), val = tensor(0)]; + tensor qkv_3_cast_fp16 = transpose(perm = var_156, x = var_155_cast_fp16)[name = tensor("transpose_21")]; + tensor var_158_cast_fp16_0, tensor var_158_cast_fp16_1, tensor var_158_cast_fp16_2 = split(axis = var_158_axis_0, split_sizes = var_158_split_sizes_0, x = qkv_3_cast_fp16)[name = tensor("op_158_cast_fp16")]; + tensor squeeze_3_axes_0 = const()[name = tensor("squeeze_3_axes_0"), val = tensor([0])]; + tensor squeeze_3_cast_fp16 = squeeze(axes = squeeze_3_axes_0, x = var_158_cast_fp16_0)[name = tensor("squeeze_3_cast_fp16")]; + tensor squeeze_4_axes_0 = const()[name = tensor("squeeze_4_axes_0"), val = tensor([0])]; + tensor squeeze_4_cast_fp16 = squeeze(axes = squeeze_4_axes_0, x = var_158_cast_fp16_1)[name = tensor("squeeze_4_cast_fp16")]; + tensor squeeze_5_axes_0 = const()[name = tensor("squeeze_5_axes_0"), val = tensor([0])]; + tensor squeeze_5_cast_fp16 = squeeze(axes = squeeze_5_axes_0, x = var_158_cast_fp16_2)[name = tensor("squeeze_5_cast_fp16")]; + tensor var_163_transpose_x_1 = const()[name = tensor("op_163_transpose_x_1"), val = tensor(false)]; + tensor var_163_transpose_y_1 = const()[name = tensor("op_163_transpose_y_1"), val = tensor(true)]; + tensor var_163_cast_fp16 = matmul(transpose_x = var_163_transpose_x_1, transpose_y = var_163_transpose_y_1, x = squeeze_3_cast_fp16, y = squeeze_4_cast_fp16)[name = tensor("op_163_cast_fp16")]; + tensor var_164_to_fp16 = const()[name = tensor("op_164_to_fp16"), val = tensor(0x1p-3)]; + tensor attn_5_cast_fp16 = mul(x = var_163_cast_fp16, y = var_164_to_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor input_27_cast_fp16 = softmax(axis = var_11, x = attn_5_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor var_168_transpose_x_0 = const()[name = tensor("op_168_transpose_x_0"), val = tensor(false)]; + tensor var_168_transpose_y_0 = const()[name = tensor("op_168_transpose_y_0"), val = tensor(false)]; + tensor var_168_cast_fp16 = matmul(transpose_x = var_168_transpose_x_0, transpose_y = var_168_transpose_y_0, x = input_27_cast_fp16, y = squeeze_5_cast_fp16)[name = tensor("op_168_cast_fp16")]; + tensor var_169_perm_0 = const()[name = tensor("op_169_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_170 = const()[name = tensor("op_170"), val = tensor([1, 248, 192])]; + tensor var_169_cast_fp16 = transpose(perm = var_169_perm_0, x = var_168_cast_fp16)[name = tensor("transpose_20")]; + tensor input_29_cast_fp16 = reshape(shape = var_170, x = var_169_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor m_encoder_blocks_1_attn_proj_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_1_attn_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1238976)))]; + tensor m_encoder_blocks_1_attn_proj_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_1_attn_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1312768)))]; + tensor linear_5_cast_fp16 = linear(bias = m_encoder_blocks_1_attn_proj_bias_to_fp16, weight = m_encoder_blocks_1_attn_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = input_25_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor m_encoder_blocks_1_norm2_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_1_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1313216)))]; + tensor m_encoder_blocks_1_norm2_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_1_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1313664)))]; + tensor input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = m_encoder_blocks_1_norm2_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_blocks_1_norm2_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor m_encoder_blocks_1_mlp_fc1_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_1_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1314112)))]; + tensor m_encoder_blocks_1_mlp_fc1_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_1_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1609088)))]; + tensor linear_6_cast_fp16 = linear(bias = m_encoder_blocks_1_mlp_fc1_bias_to_fp16, weight = m_encoder_blocks_1_mlp_fc1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_6_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 = linear_6_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor m_encoder_blocks_1_mlp_fc2_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_1_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1610688)))]; + tensor m_encoder_blocks_1_mlp_fc2_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_1_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1905664)))]; + tensor linear_7_cast_fp16 = linear(bias = m_encoder_blocks_1_mlp_fc2_bias_to_fp16, weight = m_encoder_blocks_1_mlp_fc2_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_45_cast_fp16 = add(x = input_33_cast_fp16, y = linear_7_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor m_encoder_blocks_2_norm1_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_2_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1906112)))]; + tensor m_encoder_blocks_2_norm1_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_2_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1906560)))]; + tensor x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = m_encoder_blocks_2_norm1_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_blocks_2_norm1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor m_encoder_blocks_2_attn_qkv_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_2_attn_qkv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1907008)))]; + tensor m_encoder_blocks_2_attn_qkv_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_2_attn_qkv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2128256)))]; + tensor linear_8_cast_fp16 = linear(bias = m_encoder_blocks_2_attn_qkv_bias_to_fp16, weight = m_encoder_blocks_2_attn_qkv_weight_to_fp16, x = x_17_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_212 = const()[name = tensor("op_212"), val = tensor([1, 248, 3, 3, 64])]; + tensor var_213_cast_fp16 = reshape(shape = var_212, x = linear_8_cast_fp16)[name = tensor("op_213_cast_fp16")]; + tensor var_214 = const()[name = tensor("op_214"), val = tensor([2, 0, 3, 1, 4])]; + tensor var_216_split_sizes_0 = const()[name = tensor("op_216_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_216_axis_0 = const()[name = tensor("op_216_axis_0"), val = tensor(0)]; + tensor qkv_5_cast_fp16 = transpose(perm = var_214, x = var_213_cast_fp16)[name = tensor("transpose_19")]; + tensor var_216_cast_fp16_0, tensor var_216_cast_fp16_1, tensor var_216_cast_fp16_2 = split(axis = var_216_axis_0, split_sizes = var_216_split_sizes_0, x = qkv_5_cast_fp16)[name = tensor("op_216_cast_fp16")]; + tensor squeeze_6_axes_0 = const()[name = tensor("squeeze_6_axes_0"), val = tensor([0])]; + tensor squeeze_6_cast_fp16 = squeeze(axes = squeeze_6_axes_0, x = var_216_cast_fp16_0)[name = tensor("squeeze_6_cast_fp16")]; + tensor squeeze_7_axes_0 = const()[name = tensor("squeeze_7_axes_0"), val = tensor([0])]; + tensor squeeze_7_cast_fp16 = squeeze(axes = squeeze_7_axes_0, x = var_216_cast_fp16_1)[name = tensor("squeeze_7_cast_fp16")]; + tensor squeeze_8_axes_0 = const()[name = tensor("squeeze_8_axes_0"), val = tensor([0])]; + tensor squeeze_8_cast_fp16 = squeeze(axes = squeeze_8_axes_0, x = var_216_cast_fp16_2)[name = tensor("squeeze_8_cast_fp16")]; + tensor var_221_transpose_x_1 = const()[name = tensor("op_221_transpose_x_1"), val = tensor(false)]; + tensor var_221_transpose_y_1 = const()[name = tensor("op_221_transpose_y_1"), val = tensor(true)]; + tensor var_221_cast_fp16 = matmul(transpose_x = var_221_transpose_x_1, transpose_y = var_221_transpose_y_1, x = squeeze_6_cast_fp16, y = squeeze_7_cast_fp16)[name = tensor("op_221_cast_fp16")]; + tensor var_222_to_fp16 = const()[name = tensor("op_222_to_fp16"), val = tensor(0x1p-3)]; + tensor attn_9_cast_fp16 = mul(x = var_221_cast_fp16, y = var_222_to_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor input_47_cast_fp16 = softmax(axis = var_11, x = attn_9_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor var_226_transpose_x_0 = const()[name = tensor("op_226_transpose_x_0"), val = tensor(false)]; + tensor var_226_transpose_y_0 = const()[name = tensor("op_226_transpose_y_0"), val = tensor(false)]; + tensor var_226_cast_fp16 = matmul(transpose_x = var_226_transpose_x_0, transpose_y = var_226_transpose_y_0, x = input_47_cast_fp16, y = squeeze_8_cast_fp16)[name = tensor("op_226_cast_fp16")]; + tensor var_227_perm_0 = const()[name = tensor("op_227_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_228 = const()[name = tensor("op_228"), val = tensor([1, 248, 192])]; + tensor var_227_cast_fp16 = transpose(perm = var_227_perm_0, x = var_226_cast_fp16)[name = tensor("transpose_18")]; + tensor input_49_cast_fp16 = reshape(shape = var_228, x = var_227_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor m_encoder_blocks_2_attn_proj_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_2_attn_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2129472)))]; + tensor m_encoder_blocks_2_attn_proj_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_2_attn_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2203264)))]; + tensor linear_9_cast_fp16 = linear(bias = m_encoder_blocks_2_attn_proj_bias_to_fp16, weight = m_encoder_blocks_2_attn_proj_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor input_53_cast_fp16 = add(x = input_45_cast_fp16, y = linear_9_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor input_55_axes_0 = const()[name = tensor("input_55_axes_0"), val = tensor([-1])]; + tensor m_encoder_blocks_2_norm2_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_2_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2203712)))]; + tensor m_encoder_blocks_2_norm2_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_2_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2204160)))]; + tensor input_55_cast_fp16 = layer_norm(axes = input_55_axes_0, beta = m_encoder_blocks_2_norm2_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_blocks_2_norm2_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor m_encoder_blocks_2_mlp_fc1_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_2_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2204608)))]; + tensor m_encoder_blocks_2_mlp_fc1_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_2_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2499584)))]; + tensor linear_10_cast_fp16 = linear(bias = m_encoder_blocks_2_mlp_fc1_bias_to_fp16, weight = m_encoder_blocks_2_mlp_fc1_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = linear_10_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor m_encoder_blocks_2_mlp_fc2_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_2_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2501184)))]; + tensor m_encoder_blocks_2_mlp_fc2_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_2_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2796160)))]; + tensor linear_11_cast_fp16 = linear(bias = m_encoder_blocks_2_mlp_fc2_bias_to_fp16, weight = m_encoder_blocks_2_mlp_fc2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_65_cast_fp16 = add(x = input_53_cast_fp16, y = linear_11_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor m_encoder_blocks_3_norm1_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_3_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2796608)))]; + tensor m_encoder_blocks_3_norm1_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_3_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2797056)))]; + tensor x_19_cast_fp16 = layer_norm(axes = x_19_axes_0, beta = m_encoder_blocks_3_norm1_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_blocks_3_norm1_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor m_encoder_blocks_3_attn_qkv_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_3_attn_qkv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2797504)))]; + tensor m_encoder_blocks_3_attn_qkv_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_3_attn_qkv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3018752)))]; + tensor linear_12_cast_fp16 = linear(bias = m_encoder_blocks_3_attn_qkv_bias_to_fp16, weight = m_encoder_blocks_3_attn_qkv_weight_to_fp16, x = x_19_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_270 = const()[name = tensor("op_270"), val = tensor([1, 248, 3, 3, 64])]; + tensor var_271_cast_fp16 = reshape(shape = var_270, x = linear_12_cast_fp16)[name = tensor("op_271_cast_fp16")]; + tensor var_272 = const()[name = tensor("op_272"), val = tensor([2, 0, 3, 1, 4])]; + tensor var_274_split_sizes_0 = const()[name = tensor("op_274_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_274_axis_0 = const()[name = tensor("op_274_axis_0"), val = tensor(0)]; + tensor qkv_7_cast_fp16 = transpose(perm = var_272, x = var_271_cast_fp16)[name = tensor("transpose_17")]; + tensor var_274_cast_fp16_0, tensor var_274_cast_fp16_1, tensor var_274_cast_fp16_2 = split(axis = var_274_axis_0, split_sizes = var_274_split_sizes_0, x = qkv_7_cast_fp16)[name = tensor("op_274_cast_fp16")]; + tensor squeeze_9_axes_0 = const()[name = tensor("squeeze_9_axes_0"), val = tensor([0])]; + tensor squeeze_9_cast_fp16 = squeeze(axes = squeeze_9_axes_0, x = var_274_cast_fp16_0)[name = tensor("squeeze_9_cast_fp16")]; + tensor squeeze_10_axes_0 = const()[name = tensor("squeeze_10_axes_0"), val = tensor([0])]; + tensor squeeze_10_cast_fp16 = squeeze(axes = squeeze_10_axes_0, x = var_274_cast_fp16_1)[name = tensor("squeeze_10_cast_fp16")]; + tensor squeeze_11_axes_0 = const()[name = tensor("squeeze_11_axes_0"), val = tensor([0])]; + tensor squeeze_11_cast_fp16 = squeeze(axes = squeeze_11_axes_0, x = var_274_cast_fp16_2)[name = tensor("squeeze_11_cast_fp16")]; + tensor var_279_transpose_x_1 = const()[name = tensor("op_279_transpose_x_1"), val = tensor(false)]; + tensor var_279_transpose_y_1 = const()[name = tensor("op_279_transpose_y_1"), val = tensor(true)]; + tensor var_279_cast_fp16 = matmul(transpose_x = var_279_transpose_x_1, transpose_y = var_279_transpose_y_1, x = squeeze_9_cast_fp16, y = squeeze_10_cast_fp16)[name = tensor("op_279_cast_fp16")]; + tensor var_280_to_fp16 = const()[name = tensor("op_280_to_fp16"), val = tensor(0x1p-3)]; + tensor attn_13_cast_fp16 = mul(x = var_279_cast_fp16, y = var_280_to_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor input_67_cast_fp16 = softmax(axis = var_11, x = attn_13_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor var_284_transpose_x_0 = const()[name = tensor("op_284_transpose_x_0"), val = tensor(false)]; + tensor var_284_transpose_y_0 = const()[name = tensor("op_284_transpose_y_0"), val = tensor(false)]; + tensor var_284_cast_fp16 = matmul(transpose_x = var_284_transpose_x_0, transpose_y = var_284_transpose_y_0, x = input_67_cast_fp16, y = squeeze_11_cast_fp16)[name = tensor("op_284_cast_fp16")]; + tensor var_285_perm_0 = const()[name = tensor("op_285_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_286 = const()[name = tensor("op_286"), val = tensor([1, 248, 192])]; + tensor var_285_cast_fp16 = transpose(perm = var_285_perm_0, x = var_284_cast_fp16)[name = tensor("transpose_16")]; + tensor input_69_cast_fp16 = reshape(shape = var_286, x = var_285_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor m_encoder_blocks_3_attn_proj_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_3_attn_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3019968)))]; + tensor m_encoder_blocks_3_attn_proj_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_3_attn_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3093760)))]; + tensor linear_13_cast_fp16 = linear(bias = m_encoder_blocks_3_attn_proj_bias_to_fp16, weight = m_encoder_blocks_3_attn_proj_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor input_73_cast_fp16 = add(x = input_65_cast_fp16, y = linear_13_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor m_encoder_blocks_3_norm2_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_3_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3094208)))]; + tensor m_encoder_blocks_3_norm2_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_3_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3094656)))]; + tensor input_75_cast_fp16 = layer_norm(axes = input_75_axes_0, beta = m_encoder_blocks_3_norm2_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_blocks_3_norm2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor m_encoder_blocks_3_mlp_fc1_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_3_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3095104)))]; + tensor m_encoder_blocks_3_mlp_fc1_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_3_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3390080)))]; + tensor linear_14_cast_fp16 = linear(bias = m_encoder_blocks_3_mlp_fc1_bias_to_fp16, weight = m_encoder_blocks_3_mlp_fc1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("linear_14_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 = linear_14_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor m_encoder_blocks_3_mlp_fc2_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_3_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3391680)))]; + tensor m_encoder_blocks_3_mlp_fc2_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_3_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3686656)))]; + tensor linear_15_cast_fp16 = linear(bias = m_encoder_blocks_3_mlp_fc2_bias_to_fp16, weight = m_encoder_blocks_3_mlp_fc2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_85_cast_fp16 = add(x = input_73_cast_fp16, y = linear_15_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor x_21_axes_0 = const()[name = tensor("x_21_axes_0"), val = tensor([-1])]; + tensor m_encoder_blocks_4_norm1_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_4_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3687104)))]; + tensor m_encoder_blocks_4_norm1_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_4_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3687552)))]; + tensor x_21_cast_fp16 = layer_norm(axes = x_21_axes_0, beta = m_encoder_blocks_4_norm1_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_blocks_4_norm1_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("x_21_cast_fp16")]; + tensor m_encoder_blocks_4_attn_qkv_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_4_attn_qkv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3688000)))]; + tensor m_encoder_blocks_4_attn_qkv_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_4_attn_qkv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3909248)))]; + tensor linear_16_cast_fp16 = linear(bias = m_encoder_blocks_4_attn_qkv_bias_to_fp16, weight = m_encoder_blocks_4_attn_qkv_weight_to_fp16, x = x_21_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor var_328 = const()[name = tensor("op_328"), val = tensor([1, 248, 3, 3, 64])]; + tensor var_329_cast_fp16 = reshape(shape = var_328, x = linear_16_cast_fp16)[name = tensor("op_329_cast_fp16")]; + tensor var_330 = const()[name = tensor("op_330"), val = tensor([2, 0, 3, 1, 4])]; + tensor var_332_split_sizes_0 = const()[name = tensor("op_332_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_332_axis_0 = const()[name = tensor("op_332_axis_0"), val = tensor(0)]; + tensor qkv_9_cast_fp16 = transpose(perm = var_330, x = var_329_cast_fp16)[name = tensor("transpose_15")]; + tensor var_332_cast_fp16_0, tensor var_332_cast_fp16_1, tensor var_332_cast_fp16_2 = split(axis = var_332_axis_0, split_sizes = var_332_split_sizes_0, x = qkv_9_cast_fp16)[name = tensor("op_332_cast_fp16")]; + tensor squeeze_12_axes_0 = const()[name = tensor("squeeze_12_axes_0"), val = tensor([0])]; + tensor squeeze_12_cast_fp16 = squeeze(axes = squeeze_12_axes_0, x = var_332_cast_fp16_0)[name = tensor("squeeze_12_cast_fp16")]; + tensor squeeze_13_axes_0 = const()[name = tensor("squeeze_13_axes_0"), val = tensor([0])]; + tensor squeeze_13_cast_fp16 = squeeze(axes = squeeze_13_axes_0, x = var_332_cast_fp16_1)[name = tensor("squeeze_13_cast_fp16")]; + tensor squeeze_14_axes_0 = const()[name = tensor("squeeze_14_axes_0"), val = tensor([0])]; + tensor squeeze_14_cast_fp16 = squeeze(axes = squeeze_14_axes_0, x = var_332_cast_fp16_2)[name = tensor("squeeze_14_cast_fp16")]; + tensor var_337_transpose_x_1 = const()[name = tensor("op_337_transpose_x_1"), val = tensor(false)]; + tensor var_337_transpose_y_1 = const()[name = tensor("op_337_transpose_y_1"), val = tensor(true)]; + tensor var_337_cast_fp16 = matmul(transpose_x = var_337_transpose_x_1, transpose_y = var_337_transpose_y_1, x = squeeze_12_cast_fp16, y = squeeze_13_cast_fp16)[name = tensor("op_337_cast_fp16")]; + tensor var_338_to_fp16 = const()[name = tensor("op_338_to_fp16"), val = tensor(0x1p-3)]; + tensor attn_17_cast_fp16 = mul(x = var_337_cast_fp16, y = var_338_to_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor input_87_cast_fp16 = softmax(axis = var_11, x = attn_17_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor var_342_transpose_x_0 = const()[name = tensor("op_342_transpose_x_0"), val = tensor(false)]; + tensor var_342_transpose_y_0 = const()[name = tensor("op_342_transpose_y_0"), val = tensor(false)]; + tensor var_342_cast_fp16 = matmul(transpose_x = var_342_transpose_x_0, transpose_y = var_342_transpose_y_0, x = input_87_cast_fp16, y = squeeze_14_cast_fp16)[name = tensor("op_342_cast_fp16")]; + tensor var_343_perm_0 = const()[name = tensor("op_343_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_344 = const()[name = tensor("op_344"), val = tensor([1, 248, 192])]; + tensor var_343_cast_fp16 = transpose(perm = var_343_perm_0, x = var_342_cast_fp16)[name = tensor("transpose_14")]; + tensor input_89_cast_fp16 = reshape(shape = var_344, x = var_343_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor m_encoder_blocks_4_attn_proj_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_4_attn_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3910464)))]; + tensor m_encoder_blocks_4_attn_proj_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_4_attn_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3984256)))]; + tensor linear_17_cast_fp16 = linear(bias = m_encoder_blocks_4_attn_proj_bias_to_fp16, weight = m_encoder_blocks_4_attn_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor input_93_cast_fp16 = add(x = input_85_cast_fp16, y = linear_17_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor input_95_axes_0 = const()[name = tensor("input_95_axes_0"), val = tensor([-1])]; + tensor m_encoder_blocks_4_norm2_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_4_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3984704)))]; + tensor m_encoder_blocks_4_norm2_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_4_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3985152)))]; + tensor input_95_cast_fp16 = layer_norm(axes = input_95_axes_0, beta = m_encoder_blocks_4_norm2_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_blocks_4_norm2_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor m_encoder_blocks_4_mlp_fc1_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_4_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3985600)))]; + tensor m_encoder_blocks_4_mlp_fc1_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_4_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4280576)))]; + tensor linear_18_cast_fp16 = linear(bias = m_encoder_blocks_4_mlp_fc1_bias_to_fp16, weight = m_encoder_blocks_4_mlp_fc1_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor input_99_mode_0 = const()[name = tensor("input_99_mode_0"), val = tensor("EXACT")]; + tensor input_99_cast_fp16 = gelu(mode = input_99_mode_0, x = linear_18_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor m_encoder_blocks_4_mlp_fc2_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_4_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4282176)))]; + tensor m_encoder_blocks_4_mlp_fc2_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_4_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4577152)))]; + tensor linear_19_cast_fp16 = linear(bias = m_encoder_blocks_4_mlp_fc2_bias_to_fp16, weight = m_encoder_blocks_4_mlp_fc2_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_105_cast_fp16 = add(x = input_93_cast_fp16, y = linear_19_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor m_encoder_blocks_5_norm1_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_5_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4577600)))]; + tensor m_encoder_blocks_5_norm1_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_5_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4578048)))]; + tensor x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = m_encoder_blocks_5_norm1_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_blocks_5_norm1_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor m_encoder_blocks_5_attn_qkv_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_5_attn_qkv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4578496)))]; + tensor m_encoder_blocks_5_attn_qkv_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_5_attn_qkv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4799744)))]; + tensor linear_20_cast_fp16 = linear(bias = m_encoder_blocks_5_attn_qkv_bias_to_fp16, weight = m_encoder_blocks_5_attn_qkv_weight_to_fp16, x = x_23_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_386 = const()[name = tensor("op_386"), val = tensor([1, 248, 3, 3, 64])]; + tensor var_387_cast_fp16 = reshape(shape = var_386, x = linear_20_cast_fp16)[name = tensor("op_387_cast_fp16")]; + tensor var_388 = const()[name = tensor("op_388"), val = tensor([2, 0, 3, 1, 4])]; + tensor var_390_split_sizes_0 = const()[name = tensor("op_390_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_390_axis_0 = const()[name = tensor("op_390_axis_0"), val = tensor(0)]; + tensor qkv_11_cast_fp16 = transpose(perm = var_388, x = var_387_cast_fp16)[name = tensor("transpose_13")]; + tensor var_390_cast_fp16_0, tensor var_390_cast_fp16_1, tensor var_390_cast_fp16_2 = split(axis = var_390_axis_0, split_sizes = var_390_split_sizes_0, x = qkv_11_cast_fp16)[name = tensor("op_390_cast_fp16")]; + tensor squeeze_15_axes_0 = const()[name = tensor("squeeze_15_axes_0"), val = tensor([0])]; + tensor squeeze_15_cast_fp16 = squeeze(axes = squeeze_15_axes_0, x = var_390_cast_fp16_0)[name = tensor("squeeze_15_cast_fp16")]; + tensor squeeze_16_axes_0 = const()[name = tensor("squeeze_16_axes_0"), val = tensor([0])]; + tensor squeeze_16_cast_fp16 = squeeze(axes = squeeze_16_axes_0, x = var_390_cast_fp16_1)[name = tensor("squeeze_16_cast_fp16")]; + tensor squeeze_17_axes_0 = const()[name = tensor("squeeze_17_axes_0"), val = tensor([0])]; + tensor squeeze_17_cast_fp16 = squeeze(axes = squeeze_17_axes_0, x = var_390_cast_fp16_2)[name = tensor("squeeze_17_cast_fp16")]; + tensor var_395_transpose_x_1 = const()[name = tensor("op_395_transpose_x_1"), val = tensor(false)]; + tensor var_395_transpose_y_1 = const()[name = tensor("op_395_transpose_y_1"), val = tensor(true)]; + tensor var_395_cast_fp16 = matmul(transpose_x = var_395_transpose_x_1, transpose_y = var_395_transpose_y_1, x = squeeze_15_cast_fp16, y = squeeze_16_cast_fp16)[name = tensor("op_395_cast_fp16")]; + tensor var_396_to_fp16 = const()[name = tensor("op_396_to_fp16"), val = tensor(0x1p-3)]; + tensor attn_21_cast_fp16 = mul(x = var_395_cast_fp16, y = var_396_to_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor input_107_cast_fp16 = softmax(axis = var_11, x = attn_21_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor var_400_transpose_x_0 = const()[name = tensor("op_400_transpose_x_0"), val = tensor(false)]; + tensor var_400_transpose_y_0 = const()[name = tensor("op_400_transpose_y_0"), val = tensor(false)]; + tensor var_400_cast_fp16 = matmul(transpose_x = var_400_transpose_x_0, transpose_y = var_400_transpose_y_0, x = input_107_cast_fp16, y = squeeze_17_cast_fp16)[name = tensor("op_400_cast_fp16")]; + tensor var_401_perm_0 = const()[name = tensor("op_401_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_402 = const()[name = tensor("op_402"), val = tensor([1, 248, 192])]; + tensor var_401_cast_fp16 = transpose(perm = var_401_perm_0, x = var_400_cast_fp16)[name = tensor("transpose_12")]; + tensor input_109_cast_fp16 = reshape(shape = var_402, x = var_401_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor m_encoder_blocks_5_attn_proj_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_5_attn_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4800960)))]; + tensor m_encoder_blocks_5_attn_proj_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_5_attn_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4874752)))]; + tensor linear_21_cast_fp16 = linear(bias = m_encoder_blocks_5_attn_proj_bias_to_fp16, weight = m_encoder_blocks_5_attn_proj_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor input_113_cast_fp16 = add(x = input_105_cast_fp16, y = linear_21_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor input_115_axes_0 = const()[name = tensor("input_115_axes_0"), val = tensor([-1])]; + tensor m_encoder_blocks_5_norm2_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_5_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4875200)))]; + tensor m_encoder_blocks_5_norm2_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_5_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4875648)))]; + tensor input_115_cast_fp16 = layer_norm(axes = input_115_axes_0, beta = m_encoder_blocks_5_norm2_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_blocks_5_norm2_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor m_encoder_blocks_5_mlp_fc1_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_5_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4876096)))]; + tensor m_encoder_blocks_5_mlp_fc1_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_5_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5171072)))]; + tensor linear_22_cast_fp16 = linear(bias = m_encoder_blocks_5_mlp_fc1_bias_to_fp16, weight = m_encoder_blocks_5_mlp_fc1_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("linear_22_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 = linear_22_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor m_encoder_blocks_5_mlp_fc2_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_5_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5172672)))]; + tensor m_encoder_blocks_5_mlp_fc2_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_5_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5467648)))]; + tensor linear_23_cast_fp16 = linear(bias = m_encoder_blocks_5_mlp_fc2_bias_to_fp16, weight = m_encoder_blocks_5_mlp_fc2_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_125_cast_fp16 = add(x = input_113_cast_fp16, y = linear_23_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor m_encoder_blocks_6_norm1_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_6_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5468096)))]; + tensor m_encoder_blocks_6_norm1_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_6_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5468544)))]; + tensor x_25_cast_fp16 = layer_norm(axes = x_25_axes_0, beta = m_encoder_blocks_6_norm1_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_blocks_6_norm1_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor m_encoder_blocks_6_attn_qkv_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_6_attn_qkv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5468992)))]; + tensor m_encoder_blocks_6_attn_qkv_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_6_attn_qkv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5690240)))]; + tensor linear_24_cast_fp16 = linear(bias = m_encoder_blocks_6_attn_qkv_bias_to_fp16, weight = m_encoder_blocks_6_attn_qkv_weight_to_fp16, x = x_25_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor var_444 = const()[name = tensor("op_444"), val = tensor([1, 248, 3, 3, 64])]; + tensor var_445_cast_fp16 = reshape(shape = var_444, x = linear_24_cast_fp16)[name = tensor("op_445_cast_fp16")]; + tensor var_446 = const()[name = tensor("op_446"), val = tensor([2, 0, 3, 1, 4])]; + tensor var_448_split_sizes_0 = const()[name = tensor("op_448_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_448_axis_0 = const()[name = tensor("op_448_axis_0"), val = tensor(0)]; + tensor qkv_13_cast_fp16 = transpose(perm = var_446, x = var_445_cast_fp16)[name = tensor("transpose_11")]; + tensor var_448_cast_fp16_0, tensor var_448_cast_fp16_1, tensor var_448_cast_fp16_2 = split(axis = var_448_axis_0, split_sizes = var_448_split_sizes_0, x = qkv_13_cast_fp16)[name = tensor("op_448_cast_fp16")]; + tensor squeeze_18_axes_0 = const()[name = tensor("squeeze_18_axes_0"), val = tensor([0])]; + tensor squeeze_18_cast_fp16 = squeeze(axes = squeeze_18_axes_0, x = var_448_cast_fp16_0)[name = tensor("squeeze_18_cast_fp16")]; + tensor squeeze_19_axes_0 = const()[name = tensor("squeeze_19_axes_0"), val = tensor([0])]; + tensor squeeze_19_cast_fp16 = squeeze(axes = squeeze_19_axes_0, x = var_448_cast_fp16_1)[name = tensor("squeeze_19_cast_fp16")]; + tensor squeeze_20_axes_0 = const()[name = tensor("squeeze_20_axes_0"), val = tensor([0])]; + tensor squeeze_20_cast_fp16 = squeeze(axes = squeeze_20_axes_0, x = var_448_cast_fp16_2)[name = tensor("squeeze_20_cast_fp16")]; + tensor var_453_transpose_x_1 = const()[name = tensor("op_453_transpose_x_1"), val = tensor(false)]; + tensor var_453_transpose_y_1 = const()[name = tensor("op_453_transpose_y_1"), val = tensor(true)]; + tensor var_453_cast_fp16 = matmul(transpose_x = var_453_transpose_x_1, transpose_y = var_453_transpose_y_1, x = squeeze_18_cast_fp16, y = squeeze_19_cast_fp16)[name = tensor("op_453_cast_fp16")]; + tensor var_454_to_fp16 = const()[name = tensor("op_454_to_fp16"), val = tensor(0x1p-3)]; + tensor attn_25_cast_fp16 = mul(x = var_453_cast_fp16, y = var_454_to_fp16)[name = tensor("attn_25_cast_fp16")]; + tensor input_127_cast_fp16 = softmax(axis = var_11, x = attn_25_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor var_458_transpose_x_0 = const()[name = tensor("op_458_transpose_x_0"), val = tensor(false)]; + tensor var_458_transpose_y_0 = const()[name = tensor("op_458_transpose_y_0"), val = tensor(false)]; + tensor var_458_cast_fp16 = matmul(transpose_x = var_458_transpose_x_0, transpose_y = var_458_transpose_y_0, x = input_127_cast_fp16, y = squeeze_20_cast_fp16)[name = tensor("op_458_cast_fp16")]; + tensor var_459_perm_0 = const()[name = tensor("op_459_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_460 = const()[name = tensor("op_460"), val = tensor([1, 248, 192])]; + tensor var_459_cast_fp16 = transpose(perm = var_459_perm_0, x = var_458_cast_fp16)[name = tensor("transpose_10")]; + tensor input_129_cast_fp16 = reshape(shape = var_460, x = var_459_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor m_encoder_blocks_6_attn_proj_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_6_attn_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5691456)))]; + tensor m_encoder_blocks_6_attn_proj_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_6_attn_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5765248)))]; + tensor linear_25_cast_fp16 = linear(bias = m_encoder_blocks_6_attn_proj_bias_to_fp16, weight = m_encoder_blocks_6_attn_proj_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = input_125_cast_fp16, y = linear_25_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor input_135_axes_0 = const()[name = tensor("input_135_axes_0"), val = tensor([-1])]; + tensor m_encoder_blocks_6_norm2_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_6_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5765696)))]; + tensor m_encoder_blocks_6_norm2_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_6_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5766144)))]; + tensor input_135_cast_fp16 = layer_norm(axes = input_135_axes_0, beta = m_encoder_blocks_6_norm2_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_blocks_6_norm2_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor m_encoder_blocks_6_mlp_fc1_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_6_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5766592)))]; + tensor m_encoder_blocks_6_mlp_fc1_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_6_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6061568)))]; + tensor linear_26_cast_fp16 = linear(bias = m_encoder_blocks_6_mlp_fc1_bias_to_fp16, weight = m_encoder_blocks_6_mlp_fc1_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor input_139_mode_0 = const()[name = tensor("input_139_mode_0"), val = tensor("EXACT")]; + tensor input_139_cast_fp16 = gelu(mode = input_139_mode_0, x = linear_26_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor m_encoder_blocks_6_mlp_fc2_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_6_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6063168)))]; + tensor m_encoder_blocks_6_mlp_fc2_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_6_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6358144)))]; + tensor linear_27_cast_fp16 = linear(bias = m_encoder_blocks_6_mlp_fc2_bias_to_fp16, weight = m_encoder_blocks_6_mlp_fc2_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_145_cast_fp16 = add(x = input_133_cast_fp16, y = linear_27_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor x_27_axes_0 = const()[name = tensor("x_27_axes_0"), val = tensor([-1])]; + tensor m_encoder_blocks_7_norm1_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_7_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6358592)))]; + tensor m_encoder_blocks_7_norm1_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_7_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6359040)))]; + tensor x_27_cast_fp16 = layer_norm(axes = x_27_axes_0, beta = m_encoder_blocks_7_norm1_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_blocks_7_norm1_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("x_27_cast_fp16")]; + tensor m_encoder_blocks_7_attn_qkv_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_7_attn_qkv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6359488)))]; + tensor m_encoder_blocks_7_attn_qkv_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_7_attn_qkv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6580736)))]; + tensor linear_28_cast_fp16 = linear(bias = m_encoder_blocks_7_attn_qkv_bias_to_fp16, weight = m_encoder_blocks_7_attn_qkv_weight_to_fp16, x = x_27_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor var_502 = const()[name = tensor("op_502"), val = tensor([1, 248, 3, 3, 64])]; + tensor var_503_cast_fp16 = reshape(shape = var_502, x = linear_28_cast_fp16)[name = tensor("op_503_cast_fp16")]; + tensor var_504 = const()[name = tensor("op_504"), val = tensor([2, 0, 3, 1, 4])]; + tensor var_506_split_sizes_0 = const()[name = tensor("op_506_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_506_axis_0 = const()[name = tensor("op_506_axis_0"), val = tensor(0)]; + tensor qkv_15_cast_fp16 = transpose(perm = var_504, x = var_503_cast_fp16)[name = tensor("transpose_9")]; + tensor var_506_cast_fp16_0, tensor var_506_cast_fp16_1, tensor var_506_cast_fp16_2 = split(axis = var_506_axis_0, split_sizes = var_506_split_sizes_0, x = qkv_15_cast_fp16)[name = tensor("op_506_cast_fp16")]; + tensor squeeze_21_axes_0 = const()[name = tensor("squeeze_21_axes_0"), val = tensor([0])]; + tensor squeeze_21_cast_fp16 = squeeze(axes = squeeze_21_axes_0, x = var_506_cast_fp16_0)[name = tensor("squeeze_21_cast_fp16")]; + tensor squeeze_22_axes_0 = const()[name = tensor("squeeze_22_axes_0"), val = tensor([0])]; + tensor squeeze_22_cast_fp16 = squeeze(axes = squeeze_22_axes_0, x = var_506_cast_fp16_1)[name = tensor("squeeze_22_cast_fp16")]; + tensor squeeze_23_axes_0 = const()[name = tensor("squeeze_23_axes_0"), val = tensor([0])]; + tensor squeeze_23_cast_fp16 = squeeze(axes = squeeze_23_axes_0, x = var_506_cast_fp16_2)[name = tensor("squeeze_23_cast_fp16")]; + tensor var_511_transpose_x_1 = const()[name = tensor("op_511_transpose_x_1"), val = tensor(false)]; + tensor var_511_transpose_y_1 = const()[name = tensor("op_511_transpose_y_1"), val = tensor(true)]; + tensor var_511_cast_fp16 = matmul(transpose_x = var_511_transpose_x_1, transpose_y = var_511_transpose_y_1, x = squeeze_21_cast_fp16, y = squeeze_22_cast_fp16)[name = tensor("op_511_cast_fp16")]; + tensor var_512_to_fp16 = const()[name = tensor("op_512_to_fp16"), val = tensor(0x1p-3)]; + tensor attn_29_cast_fp16 = mul(x = var_511_cast_fp16, y = var_512_to_fp16)[name = tensor("attn_29_cast_fp16")]; + tensor input_147_cast_fp16 = softmax(axis = var_11, x = attn_29_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor var_516_transpose_x_0 = const()[name = tensor("op_516_transpose_x_0"), val = tensor(false)]; + tensor var_516_transpose_y_0 = const()[name = tensor("op_516_transpose_y_0"), val = tensor(false)]; + tensor var_516_cast_fp16 = matmul(transpose_x = var_516_transpose_x_0, transpose_y = var_516_transpose_y_0, x = input_147_cast_fp16, y = squeeze_23_cast_fp16)[name = tensor("op_516_cast_fp16")]; + tensor var_517_perm_0 = const()[name = tensor("op_517_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_518 = const()[name = tensor("op_518"), val = tensor([1, 248, 192])]; + tensor var_517_cast_fp16 = transpose(perm = var_517_perm_0, x = var_516_cast_fp16)[name = tensor("transpose_8")]; + tensor input_149_cast_fp16 = reshape(shape = var_518, x = var_517_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor m_encoder_blocks_7_attn_proj_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_7_attn_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6581952)))]; + tensor m_encoder_blocks_7_attn_proj_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_7_attn_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6655744)))]; + tensor linear_29_cast_fp16 = linear(bias = m_encoder_blocks_7_attn_proj_bias_to_fp16, weight = m_encoder_blocks_7_attn_proj_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor input_153_cast_fp16 = add(x = input_145_cast_fp16, y = linear_29_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor m_encoder_blocks_7_norm2_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_7_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6656192)))]; + tensor m_encoder_blocks_7_norm2_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_7_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6656640)))]; + tensor input_155_cast_fp16 = layer_norm(axes = input_155_axes_0, beta = m_encoder_blocks_7_norm2_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_blocks_7_norm2_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor m_encoder_blocks_7_mlp_fc1_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_7_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6657088)))]; + tensor m_encoder_blocks_7_mlp_fc1_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_7_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6952064)))]; + tensor linear_30_cast_fp16 = linear(bias = m_encoder_blocks_7_mlp_fc1_bias_to_fp16, weight = m_encoder_blocks_7_mlp_fc1_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("linear_30_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 = linear_30_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor m_encoder_blocks_7_mlp_fc2_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_7_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6953664)))]; + tensor m_encoder_blocks_7_mlp_fc2_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_7_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7248640)))]; + tensor linear_31_cast_fp16 = linear(bias = m_encoder_blocks_7_mlp_fc2_bias_to_fp16, weight = m_encoder_blocks_7_mlp_fc2_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_165_cast_fp16 = add(x = input_153_cast_fp16, y = linear_31_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor x_29_axes_0 = const()[name = tensor("x_29_axes_0"), val = tensor([-1])]; + tensor m_encoder_blocks_8_norm1_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_8_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7249088)))]; + tensor m_encoder_blocks_8_norm1_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_8_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7249536)))]; + tensor x_29_cast_fp16 = layer_norm(axes = x_29_axes_0, beta = m_encoder_blocks_8_norm1_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_blocks_8_norm1_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor m_encoder_blocks_8_attn_qkv_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_8_attn_qkv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7249984)))]; + tensor m_encoder_blocks_8_attn_qkv_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_8_attn_qkv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7471232)))]; + tensor linear_32_cast_fp16 = linear(bias = m_encoder_blocks_8_attn_qkv_bias_to_fp16, weight = m_encoder_blocks_8_attn_qkv_weight_to_fp16, x = x_29_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor var_560 = const()[name = tensor("op_560"), val = tensor([1, 248, 3, 3, 64])]; + tensor var_561_cast_fp16 = reshape(shape = var_560, x = linear_32_cast_fp16)[name = tensor("op_561_cast_fp16")]; + tensor var_562 = const()[name = tensor("op_562"), val = tensor([2, 0, 3, 1, 4])]; + tensor var_564_split_sizes_0 = const()[name = tensor("op_564_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_564_axis_0 = const()[name = tensor("op_564_axis_0"), val = tensor(0)]; + tensor qkv_17_cast_fp16 = transpose(perm = var_562, x = var_561_cast_fp16)[name = tensor("transpose_7")]; + tensor var_564_cast_fp16_0, tensor var_564_cast_fp16_1, tensor var_564_cast_fp16_2 = split(axis = var_564_axis_0, split_sizes = var_564_split_sizes_0, x = qkv_17_cast_fp16)[name = tensor("op_564_cast_fp16")]; + tensor squeeze_24_axes_0 = const()[name = tensor("squeeze_24_axes_0"), val = tensor([0])]; + tensor squeeze_24_cast_fp16 = squeeze(axes = squeeze_24_axes_0, x = var_564_cast_fp16_0)[name = tensor("squeeze_24_cast_fp16")]; + tensor squeeze_25_axes_0 = const()[name = tensor("squeeze_25_axes_0"), val = tensor([0])]; + tensor squeeze_25_cast_fp16 = squeeze(axes = squeeze_25_axes_0, x = var_564_cast_fp16_1)[name = tensor("squeeze_25_cast_fp16")]; + tensor squeeze_26_axes_0 = const()[name = tensor("squeeze_26_axes_0"), val = tensor([0])]; + tensor squeeze_26_cast_fp16 = squeeze(axes = squeeze_26_axes_0, x = var_564_cast_fp16_2)[name = tensor("squeeze_26_cast_fp16")]; + tensor var_569_transpose_x_1 = const()[name = tensor("op_569_transpose_x_1"), val = tensor(false)]; + tensor var_569_transpose_y_1 = const()[name = tensor("op_569_transpose_y_1"), val = tensor(true)]; + tensor var_569_cast_fp16 = matmul(transpose_x = var_569_transpose_x_1, transpose_y = var_569_transpose_y_1, x = squeeze_24_cast_fp16, y = squeeze_25_cast_fp16)[name = tensor("op_569_cast_fp16")]; + tensor var_570_to_fp16 = const()[name = tensor("op_570_to_fp16"), val = tensor(0x1p-3)]; + tensor attn_33_cast_fp16 = mul(x = var_569_cast_fp16, y = var_570_to_fp16)[name = tensor("attn_33_cast_fp16")]; + tensor input_167_cast_fp16 = softmax(axis = var_11, x = attn_33_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor var_574_transpose_x_0 = const()[name = tensor("op_574_transpose_x_0"), val = tensor(false)]; + tensor var_574_transpose_y_0 = const()[name = tensor("op_574_transpose_y_0"), val = tensor(false)]; + tensor var_574_cast_fp16 = matmul(transpose_x = var_574_transpose_x_0, transpose_y = var_574_transpose_y_0, x = input_167_cast_fp16, y = squeeze_26_cast_fp16)[name = tensor("op_574_cast_fp16")]; + tensor var_575_perm_0 = const()[name = tensor("op_575_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_576 = const()[name = tensor("op_576"), val = tensor([1, 248, 192])]; + tensor var_575_cast_fp16 = transpose(perm = var_575_perm_0, x = var_574_cast_fp16)[name = tensor("transpose_6")]; + tensor input_169_cast_fp16 = reshape(shape = var_576, x = var_575_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor m_encoder_blocks_8_attn_proj_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_8_attn_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7472448)))]; + tensor m_encoder_blocks_8_attn_proj_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_8_attn_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7546240)))]; + tensor linear_33_cast_fp16 = linear(bias = m_encoder_blocks_8_attn_proj_bias_to_fp16, weight = m_encoder_blocks_8_attn_proj_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor input_173_cast_fp16 = add(x = input_165_cast_fp16, y = linear_33_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor input_175_axes_0 = const()[name = tensor("input_175_axes_0"), val = tensor([-1])]; + tensor m_encoder_blocks_8_norm2_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_8_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7546688)))]; + tensor m_encoder_blocks_8_norm2_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_8_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7547136)))]; + tensor input_175_cast_fp16 = layer_norm(axes = input_175_axes_0, beta = m_encoder_blocks_8_norm2_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_blocks_8_norm2_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor m_encoder_blocks_8_mlp_fc1_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_8_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7547584)))]; + tensor m_encoder_blocks_8_mlp_fc1_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_8_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7842560)))]; + tensor linear_34_cast_fp16 = linear(bias = m_encoder_blocks_8_mlp_fc1_bias_to_fp16, weight = m_encoder_blocks_8_mlp_fc1_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor input_179_mode_0 = const()[name = tensor("input_179_mode_0"), val = tensor("EXACT")]; + tensor input_179_cast_fp16 = gelu(mode = input_179_mode_0, x = linear_34_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor m_encoder_blocks_8_mlp_fc2_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_8_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7844160)))]; + tensor m_encoder_blocks_8_mlp_fc2_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_8_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8139136)))]; + tensor linear_35_cast_fp16 = linear(bias = m_encoder_blocks_8_mlp_fc2_bias_to_fp16, weight = m_encoder_blocks_8_mlp_fc2_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_185_cast_fp16 = add(x = input_173_cast_fp16, y = linear_35_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor x_31_axes_0 = const()[name = tensor("x_31_axes_0"), val = tensor([-1])]; + tensor m_encoder_blocks_9_norm1_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_9_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8139584)))]; + tensor m_encoder_blocks_9_norm1_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_9_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8140032)))]; + tensor x_31_cast_fp16 = layer_norm(axes = x_31_axes_0, beta = m_encoder_blocks_9_norm1_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_blocks_9_norm1_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor m_encoder_blocks_9_attn_qkv_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_9_attn_qkv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8140480)))]; + tensor m_encoder_blocks_9_attn_qkv_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_9_attn_qkv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8361728)))]; + tensor linear_36_cast_fp16 = linear(bias = m_encoder_blocks_9_attn_qkv_bias_to_fp16, weight = m_encoder_blocks_9_attn_qkv_weight_to_fp16, x = x_31_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor var_618 = const()[name = tensor("op_618"), val = tensor([1, 248, 3, 3, 64])]; + tensor var_619_cast_fp16 = reshape(shape = var_618, x = linear_36_cast_fp16)[name = tensor("op_619_cast_fp16")]; + tensor var_620 = const()[name = tensor("op_620"), val = tensor([2, 0, 3, 1, 4])]; + tensor var_622_split_sizes_0 = const()[name = tensor("op_622_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_622_axis_0 = const()[name = tensor("op_622_axis_0"), val = tensor(0)]; + tensor qkv_19_cast_fp16 = transpose(perm = var_620, x = var_619_cast_fp16)[name = tensor("transpose_5")]; + tensor var_622_cast_fp16_0, tensor var_622_cast_fp16_1, tensor var_622_cast_fp16_2 = split(axis = var_622_axis_0, split_sizes = var_622_split_sizes_0, x = qkv_19_cast_fp16)[name = tensor("op_622_cast_fp16")]; + tensor squeeze_27_axes_0 = const()[name = tensor("squeeze_27_axes_0"), val = tensor([0])]; + tensor squeeze_27_cast_fp16 = squeeze(axes = squeeze_27_axes_0, x = var_622_cast_fp16_0)[name = tensor("squeeze_27_cast_fp16")]; + tensor squeeze_28_axes_0 = const()[name = tensor("squeeze_28_axes_0"), val = tensor([0])]; + tensor squeeze_28_cast_fp16 = squeeze(axes = squeeze_28_axes_0, x = var_622_cast_fp16_1)[name = tensor("squeeze_28_cast_fp16")]; + tensor squeeze_29_axes_0 = const()[name = tensor("squeeze_29_axes_0"), val = tensor([0])]; + tensor squeeze_29_cast_fp16 = squeeze(axes = squeeze_29_axes_0, x = var_622_cast_fp16_2)[name = tensor("squeeze_29_cast_fp16")]; + tensor var_627_transpose_x_1 = const()[name = tensor("op_627_transpose_x_1"), val = tensor(false)]; + tensor var_627_transpose_y_1 = const()[name = tensor("op_627_transpose_y_1"), val = tensor(true)]; + tensor var_627_cast_fp16 = matmul(transpose_x = var_627_transpose_x_1, transpose_y = var_627_transpose_y_1, x = squeeze_27_cast_fp16, y = squeeze_28_cast_fp16)[name = tensor("op_627_cast_fp16")]; + tensor var_628_to_fp16 = const()[name = tensor("op_628_to_fp16"), val = tensor(0x1p-3)]; + tensor attn_37_cast_fp16 = mul(x = var_627_cast_fp16, y = var_628_to_fp16)[name = tensor("attn_37_cast_fp16")]; + tensor input_187_cast_fp16 = softmax(axis = var_11, x = attn_37_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor var_632_transpose_x_0 = const()[name = tensor("op_632_transpose_x_0"), val = tensor(false)]; + tensor var_632_transpose_y_0 = const()[name = tensor("op_632_transpose_y_0"), val = tensor(false)]; + tensor var_632_cast_fp16 = matmul(transpose_x = var_632_transpose_x_0, transpose_y = var_632_transpose_y_0, x = input_187_cast_fp16, y = squeeze_29_cast_fp16)[name = tensor("op_632_cast_fp16")]; + tensor var_633_perm_0 = const()[name = tensor("op_633_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_634 = const()[name = tensor("op_634"), val = tensor([1, 248, 192])]; + tensor var_633_cast_fp16 = transpose(perm = var_633_perm_0, x = var_632_cast_fp16)[name = tensor("transpose_4")]; + tensor input_189_cast_fp16 = reshape(shape = var_634, x = var_633_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor m_encoder_blocks_9_attn_proj_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_9_attn_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8362944)))]; + tensor m_encoder_blocks_9_attn_proj_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_9_attn_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8436736)))]; + tensor linear_37_cast_fp16 = linear(bias = m_encoder_blocks_9_attn_proj_bias_to_fp16, weight = m_encoder_blocks_9_attn_proj_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_185_cast_fp16, y = linear_37_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor m_encoder_blocks_9_norm2_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_9_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8437184)))]; + tensor m_encoder_blocks_9_norm2_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_9_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8437632)))]; + tensor input_195_cast_fp16 = layer_norm(axes = input_195_axes_0, beta = m_encoder_blocks_9_norm2_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_blocks_9_norm2_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor m_encoder_blocks_9_mlp_fc1_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_9_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8438080)))]; + tensor m_encoder_blocks_9_mlp_fc1_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_9_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8733056)))]; + tensor linear_38_cast_fp16 = linear(bias = m_encoder_blocks_9_mlp_fc1_bias_to_fp16, weight = m_encoder_blocks_9_mlp_fc1_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("linear_38_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 = linear_38_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor m_encoder_blocks_9_mlp_fc2_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_9_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8734656)))]; + tensor m_encoder_blocks_9_mlp_fc2_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_9_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9029632)))]; + tensor linear_39_cast_fp16 = linear(bias = m_encoder_blocks_9_mlp_fc2_bias_to_fp16, weight = m_encoder_blocks_9_mlp_fc2_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_205_cast_fp16 = add(x = input_193_cast_fp16, y = linear_39_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor x_33_axes_0 = const()[name = tensor("x_33_axes_0"), val = tensor([-1])]; + tensor m_encoder_blocks_10_norm1_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_10_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9030080)))]; + tensor m_encoder_blocks_10_norm1_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_10_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9030528)))]; + tensor x_33_cast_fp16 = layer_norm(axes = x_33_axes_0, beta = m_encoder_blocks_10_norm1_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_blocks_10_norm1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("x_33_cast_fp16")]; + tensor m_encoder_blocks_10_attn_qkv_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_10_attn_qkv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9030976)))]; + tensor m_encoder_blocks_10_attn_qkv_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_10_attn_qkv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9252224)))]; + tensor linear_40_cast_fp16 = linear(bias = m_encoder_blocks_10_attn_qkv_bias_to_fp16, weight = m_encoder_blocks_10_attn_qkv_weight_to_fp16, x = x_33_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor var_676 = const()[name = tensor("op_676"), val = tensor([1, 248, 3, 3, 64])]; + tensor var_677_cast_fp16 = reshape(shape = var_676, x = linear_40_cast_fp16)[name = tensor("op_677_cast_fp16")]; + tensor var_678 = const()[name = tensor("op_678"), val = tensor([2, 0, 3, 1, 4])]; + tensor var_680_split_sizes_0 = const()[name = tensor("op_680_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_680_axis_0 = const()[name = tensor("op_680_axis_0"), val = tensor(0)]; + tensor qkv_21_cast_fp16 = transpose(perm = var_678, x = var_677_cast_fp16)[name = tensor("transpose_3")]; + tensor var_680_cast_fp16_0, tensor var_680_cast_fp16_1, tensor var_680_cast_fp16_2 = split(axis = var_680_axis_0, split_sizes = var_680_split_sizes_0, x = qkv_21_cast_fp16)[name = tensor("op_680_cast_fp16")]; + tensor squeeze_30_axes_0 = const()[name = tensor("squeeze_30_axes_0"), val = tensor([0])]; + tensor squeeze_30_cast_fp16 = squeeze(axes = squeeze_30_axes_0, x = var_680_cast_fp16_0)[name = tensor("squeeze_30_cast_fp16")]; + tensor squeeze_31_axes_0 = const()[name = tensor("squeeze_31_axes_0"), val = tensor([0])]; + tensor squeeze_31_cast_fp16 = squeeze(axes = squeeze_31_axes_0, x = var_680_cast_fp16_1)[name = tensor("squeeze_31_cast_fp16")]; + tensor squeeze_32_axes_0 = const()[name = tensor("squeeze_32_axes_0"), val = tensor([0])]; + tensor squeeze_32_cast_fp16 = squeeze(axes = squeeze_32_axes_0, x = var_680_cast_fp16_2)[name = tensor("squeeze_32_cast_fp16")]; + tensor var_685_transpose_x_1 = const()[name = tensor("op_685_transpose_x_1"), val = tensor(false)]; + tensor var_685_transpose_y_1 = const()[name = tensor("op_685_transpose_y_1"), val = tensor(true)]; + tensor var_685_cast_fp16 = matmul(transpose_x = var_685_transpose_x_1, transpose_y = var_685_transpose_y_1, x = squeeze_30_cast_fp16, y = squeeze_31_cast_fp16)[name = tensor("op_685_cast_fp16")]; + tensor var_686_to_fp16 = const()[name = tensor("op_686_to_fp16"), val = tensor(0x1p-3)]; + tensor attn_41_cast_fp16 = mul(x = var_685_cast_fp16, y = var_686_to_fp16)[name = tensor("attn_41_cast_fp16")]; + tensor input_207_cast_fp16 = softmax(axis = var_11, x = attn_41_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor var_690_transpose_x_0 = const()[name = tensor("op_690_transpose_x_0"), val = tensor(false)]; + tensor var_690_transpose_y_0 = const()[name = tensor("op_690_transpose_y_0"), val = tensor(false)]; + tensor var_690_cast_fp16 = matmul(transpose_x = var_690_transpose_x_0, transpose_y = var_690_transpose_y_0, x = input_207_cast_fp16, y = squeeze_32_cast_fp16)[name = tensor("op_690_cast_fp16")]; + tensor var_691_perm_0 = const()[name = tensor("op_691_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_692 = const()[name = tensor("op_692"), val = tensor([1, 248, 192])]; + tensor var_691_cast_fp16 = transpose(perm = var_691_perm_0, x = var_690_cast_fp16)[name = tensor("transpose_2")]; + tensor input_209_cast_fp16 = reshape(shape = var_692, x = var_691_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor m_encoder_blocks_10_attn_proj_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_10_attn_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9253440)))]; + tensor m_encoder_blocks_10_attn_proj_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_10_attn_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9327232)))]; + tensor linear_41_cast_fp16 = linear(bias = m_encoder_blocks_10_attn_proj_bias_to_fp16, weight = m_encoder_blocks_10_attn_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor input_213_cast_fp16 = add(x = input_205_cast_fp16, y = linear_41_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor input_215_axes_0 = const()[name = tensor("input_215_axes_0"), val = tensor([-1])]; + tensor m_encoder_blocks_10_norm2_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_10_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9327680)))]; + tensor m_encoder_blocks_10_norm2_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_10_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9328128)))]; + tensor input_215_cast_fp16 = layer_norm(axes = input_215_axes_0, beta = m_encoder_blocks_10_norm2_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_blocks_10_norm2_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor m_encoder_blocks_10_mlp_fc1_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_10_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9328576)))]; + tensor m_encoder_blocks_10_mlp_fc1_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_10_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9623552)))]; + tensor linear_42_cast_fp16 = linear(bias = m_encoder_blocks_10_mlp_fc1_bias_to_fp16, weight = m_encoder_blocks_10_mlp_fc1_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor input_219_mode_0 = const()[name = tensor("input_219_mode_0"), val = tensor("EXACT")]; + tensor input_219_cast_fp16 = gelu(mode = input_219_mode_0, x = linear_42_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor m_encoder_blocks_10_mlp_fc2_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_10_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9625152)))]; + tensor m_encoder_blocks_10_mlp_fc2_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_10_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9920128)))]; + tensor linear_43_cast_fp16 = linear(bias = m_encoder_blocks_10_mlp_fc2_bias_to_fp16, weight = m_encoder_blocks_10_mlp_fc2_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_225_cast_fp16 = add(x = input_213_cast_fp16, y = linear_43_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor x_35_axes_0 = const()[name = tensor("x_35_axes_0"), val = tensor([-1])]; + tensor m_encoder_blocks_11_norm1_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_11_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9920576)))]; + tensor m_encoder_blocks_11_norm1_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_11_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9921024)))]; + tensor x_35_cast_fp16 = layer_norm(axes = x_35_axes_0, beta = m_encoder_blocks_11_norm1_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_blocks_11_norm1_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor m_encoder_blocks_11_attn_qkv_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_11_attn_qkv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9921472)))]; + tensor m_encoder_blocks_11_attn_qkv_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_11_attn_qkv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10142720)))]; + tensor linear_44_cast_fp16 = linear(bias = m_encoder_blocks_11_attn_qkv_bias_to_fp16, weight = m_encoder_blocks_11_attn_qkv_weight_to_fp16, x = x_35_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor var_734 = const()[name = tensor("op_734"), val = tensor([1, 248, 3, 3, 64])]; + tensor var_735_cast_fp16 = reshape(shape = var_734, x = linear_44_cast_fp16)[name = tensor("op_735_cast_fp16")]; + tensor var_736 = const()[name = tensor("op_736"), val = tensor([2, 0, 3, 1, 4])]; + tensor var_738_split_sizes_0 = const()[name = tensor("op_738_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_738_axis_0 = const()[name = tensor("op_738_axis_0"), val = tensor(0)]; + tensor qkv_23_cast_fp16 = transpose(perm = var_736, x = var_735_cast_fp16)[name = tensor("transpose_1")]; + tensor var_738_cast_fp16_0, tensor var_738_cast_fp16_1, tensor var_738_cast_fp16_2 = split(axis = var_738_axis_0, split_sizes = var_738_split_sizes_0, x = qkv_23_cast_fp16)[name = tensor("op_738_cast_fp16")]; + tensor squeeze_33_axes_0 = const()[name = tensor("squeeze_33_axes_0"), val = tensor([0])]; + tensor squeeze_33_cast_fp16 = squeeze(axes = squeeze_33_axes_0, x = var_738_cast_fp16_0)[name = tensor("squeeze_33_cast_fp16")]; + tensor squeeze_34_axes_0 = const()[name = tensor("squeeze_34_axes_0"), val = tensor([0])]; + tensor squeeze_34_cast_fp16 = squeeze(axes = squeeze_34_axes_0, x = var_738_cast_fp16_1)[name = tensor("squeeze_34_cast_fp16")]; + tensor squeeze_35_axes_0 = const()[name = tensor("squeeze_35_axes_0"), val = tensor([0])]; + tensor squeeze_35_cast_fp16 = squeeze(axes = squeeze_35_axes_0, x = var_738_cast_fp16_2)[name = tensor("squeeze_35_cast_fp16")]; + tensor var_743_transpose_x_1 = const()[name = tensor("op_743_transpose_x_1"), val = tensor(false)]; + tensor var_743_transpose_y_1 = const()[name = tensor("op_743_transpose_y_1"), val = tensor(true)]; + tensor var_743_cast_fp16 = matmul(transpose_x = var_743_transpose_x_1, transpose_y = var_743_transpose_y_1, x = squeeze_33_cast_fp16, y = squeeze_34_cast_fp16)[name = tensor("op_743_cast_fp16")]; + tensor var_744_to_fp16 = const()[name = tensor("op_744_to_fp16"), val = tensor(0x1p-3)]; + tensor attn_45_cast_fp16 = mul(x = var_743_cast_fp16, y = var_744_to_fp16)[name = tensor("attn_45_cast_fp16")]; + tensor input_227_cast_fp16 = softmax(axis = var_11, x = attn_45_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor var_748_transpose_x_0 = const()[name = tensor("op_748_transpose_x_0"), val = tensor(false)]; + tensor var_748_transpose_y_0 = const()[name = tensor("op_748_transpose_y_0"), val = tensor(false)]; + tensor var_748_cast_fp16 = matmul(transpose_x = var_748_transpose_x_0, transpose_y = var_748_transpose_y_0, x = input_227_cast_fp16, y = squeeze_35_cast_fp16)[name = tensor("op_748_cast_fp16")]; + tensor var_749_perm_0 = const()[name = tensor("op_749_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_750 = const()[name = tensor("op_750"), val = tensor([1, 248, 192])]; + tensor var_749_cast_fp16 = transpose(perm = var_749_perm_0, x = var_748_cast_fp16)[name = tensor("transpose_0")]; + tensor input_229_cast_fp16 = reshape(shape = var_750, x = var_749_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor m_encoder_blocks_11_attn_proj_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_11_attn_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10143936)))]; + tensor m_encoder_blocks_11_attn_proj_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_11_attn_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10217728)))]; + tensor linear_45_cast_fp16 = linear(bias = m_encoder_blocks_11_attn_proj_bias_to_fp16, weight = m_encoder_blocks_11_attn_proj_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor input_233_cast_fp16 = add(x = input_225_cast_fp16, y = linear_45_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor input_235_axes_0 = const()[name = tensor("input_235_axes_0"), val = tensor([-1])]; + tensor m_encoder_blocks_11_norm2_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_11_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10218176)))]; + tensor m_encoder_blocks_11_norm2_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_11_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10218624)))]; + tensor input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = m_encoder_blocks_11_norm2_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_blocks_11_norm2_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor m_encoder_blocks_11_mlp_fc1_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_11_mlp_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10219072)))]; + tensor m_encoder_blocks_11_mlp_fc1_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_11_mlp_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10514048)))]; + tensor linear_46_cast_fp16 = linear(bias = m_encoder_blocks_11_mlp_fc1_bias_to_fp16, weight = m_encoder_blocks_11_mlp_fc1_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("linear_46_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 = linear_46_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor m_encoder_blocks_11_mlp_fc2_weight_to_fp16 = const()[name = tensor("m_encoder_blocks_11_mlp_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10515648)))]; + tensor m_encoder_blocks_11_mlp_fc2_bias_to_fp16 = const()[name = tensor("m_encoder_blocks_11_mlp_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10810624)))]; + tensor linear_47_cast_fp16 = linear(bias = m_encoder_blocks_11_mlp_fc2_bias_to_fp16, weight = m_encoder_blocks_11_mlp_fc2_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_245_cast_fp16 = add(x = input_233_cast_fp16, y = linear_47_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor x_axes_0 = const()[name = tensor("x_axes_0"), val = tensor([-1])]; + tensor m_encoder_norm_weight_to_fp16 = const()[name = tensor("m_encoder_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10811072)))]; + tensor m_encoder_norm_bias_to_fp16 = const()[name = tensor("m_encoder_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10811520)))]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = m_encoder_norm_bias_to_fp16, epsilon = var_14_to_fp16, gamma = m_encoder_norm_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor input_247_axes_0 = const()[name = tensor("input_247_axes_0"), val = tensor([1])]; + tensor input_247_keep_dims_0 = const()[name = tensor("input_247_keep_dims_0"), val = tensor(false)]; + tensor input_247_cast_fp16 = reduce_mean(axes = input_247_axes_0, keep_dims = input_247_keep_dims_0, x = x_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor m_outputlayer_0_weight_to_fp16 = const()[name = tensor("m_outputlayer_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10811968)))]; + tensor m_outputlayer_0_bias_to_fp16 = const()[name = tensor("m_outputlayer_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10812416)))]; + tensor var_790_to_fp16 = const()[name = tensor("op_790_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = m_outputlayer_0_bias_to_fp16, epsilon = var_790_to_fp16, gamma = m_outputlayer_0_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor m_outputlayer_1_weight_to_fp16 = const()[name = tensor("m_outputlayer_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10812864)))]; + tensor m_outputlayer_1_bias_to_fp16 = const()[name = tensor("m_outputlayer_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11015296)))]; + tensor linear_48_cast_fp16 = linear(bias = m_outputlayer_1_bias_to_fp16, weight = m_outputlayer_1_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor linear_48_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_48_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor logits = cast(dtype = linear_48_cast_fp16_to_fp32_dtype_0, x = linear_48_cast_fp16)[name = tensor("cast_26")]; + } -> (logits); +} \ No newline at end of file