desert-ant-labs/voz

1program(1.3)2[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3600.16.1"}, {"coremlc-version", "3600.25.1"}})]3{4    func decoder_1<ios18>(tensor<fp16, [1, 1280, 1, 1]> c_in, tensor<fp16, [1, 640, 1, 1]> embed, tensor<fp16, [1, 640, 1, 8]> enc_step, tensor<fp16, [1, 1280, 1, 1]> h_in) {5            int32 var_14 = const()[name = string("op_14"), val = int32(1)];6            tensor<int32, [4]> input_1_begin_0 = const()[name = string("input_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];7            tensor<int32, [4]> input_1_end_0 = const()[name = string("input_1_end_0"), val = tensor<int32, [4]>([1, 640, 1, 1])];8            tensor<bool, [4]> input_1_end_mask_0 = const()[name = string("input_1_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];9            tensor<fp16, [1, 640, 1, 1]> input_1_cast_fp16 = slice_by_index(begin = input_1_begin_0, end = input_1_end_0, end_mask = input_1_end_mask_0, x = h_in)[name = string("input_1_cast_fp16")];10            tensor<int32, [4]> c_1_begin_0 = const()[name = string("c_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];11            tensor<int32, [4]> c_1_end_0 = const()[name = string("c_1_end_0"), val = tensor<int32, [4]>([1, 640, 1, 1])];12            tensor<bool, [4]> c_1_end_mask_0 = const()[name = string("c_1_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];13            tensor<fp16, [1, 640, 1, 1]> c_1_cast_fp16 = slice_by_index(begin = c_1_begin_0, end = c_1_end_0, end_mask = c_1_end_mask_0, x = c_in)[name = string("c_1_cast_fp16")];14            string var_34_pad_type_0 = const()[name = string("op_34_pad_type_0"), val = string("valid")];15            tensor<int32, [2]> var_34_strides_0 = const()[name = string("op_34_strides_0"), val = tensor<int32, [2]>([1, 1])];16            tensor<int32, [4]> var_34_pad_0 = const()[name = string("op_34_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];17            tensor<int32, [2]> var_34_dilations_0 = const()[name = string("op_34_dilations_0"), val = tensor<int32, [2]>([1, 1])];18            int32 var_34_groups_0 = const()[name = string("op_34_groups_0"), val = int32(1)];19            tensor<fp16, [2560, 640, 1, 1]> decoder_cells_0_ih_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2560, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [2560, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1638528))))[name = string("decoder_cells_0_ih_weight_to_fp16_quantized")];20            tensor<fp16, [2560]> decoder_cells_0_ih_bias_to_fp16 = const()[name = string("decoder_cells_0_ih_bias_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1643712)))];21            tensor<fp16, [1, 2560, 1, 1]> var_34_cast_fp16 = conv(bias = decoder_cells_0_ih_bias_to_fp16, dilations = var_34_dilations_0, groups = var_34_groups_0, pad = var_34_pad_0, pad_type = var_34_pad_type_0, strides = var_34_strides_0, weight = decoder_cells_0_ih_weight_to_fp16_quantized, x = embed)[name = string("op_34_cast_fp16")];22            string var_40_pad_type_0 = const()[name = string("op_40_pad_type_0"), val = string("valid")];23            tensor<int32, [2]> var_40_strides_0 = const()[name = string("op_40_strides_0"), val = tensor<int32, [2]>([1, 1])];24            tensor<int32, [4]> var_40_pad_0 = const()[name = string("op_40_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];25            tensor<int32, [2]> var_40_dilations_0 = const()[name = string("op_40_dilations_0"), val = tensor<int32, [2]>([1, 1])];26            int32 var_40_groups_0 = const()[name = string("op_40_groups_0"), val = int32(1)];27            tensor<fp16, [2560, 640, 1, 1]> decoder_cells_0_hh_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2560, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1648896))), scale = tensor<fp16, [2560, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3287360))))[name = string("decoder_cells_0_hh_weight_to_fp16_quantized")];28            tensor<fp16, [1, 2560, 1, 1]> var_40_cast_fp16 = conv(dilations = var_40_dilations_0, groups = var_40_groups_0, pad = var_40_pad_0, pad_type = var_40_pad_type_0, strides = var_40_strides_0, weight = decoder_cells_0_hh_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("op_40_cast_fp16")];29            tensor<fp16, [1, 2560, 1, 1]> gates_1_cast_fp16 = add(x = var_34_cast_fp16, y = var_40_cast_fp16)[name = string("gates_1_cast_fp16")];30            tensor<int32, [4]> var_43_begin_0 = const()[name = string("op_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];31            tensor<int32, [4]> var_43_end_0 = const()[name = string("op_43_end_0"), val = tensor<int32, [4]>([1, 640, 1, 1])];32            tensor<bool, [4]> var_43_end_mask_0 = const()[name = string("op_43_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];33            tensor<fp16, [1, 640, 1, 1]> var_43_cast_fp16 = slice_by_index(begin = var_43_begin_0, end = var_43_end_0, end_mask = var_43_end_mask_0, x = gates_1_cast_fp16)[name = string("op_43_cast_fp16")];34            tensor<fp16, [1, 640, 1, 1]> i_1_cast_fp16 = sigmoid(x = var_43_cast_fp16)[name = string("i_1_cast_fp16")];35            tensor<int32, [4]> var_46_begin_0 = const()[name = string("op_46_begin_0"), val = tensor<int32, [4]>([0, 640, 0, 0])];36            tensor<int32, [4]> var_46_end_0 = const()[name = string("op_46_end_0"), val = tensor<int32, [4]>([1, 1280, 1, 1])];37            tensor<bool, [4]> var_46_end_mask_0 = const()[name = string("op_46_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];38            tensor<fp16, [1, 640, 1, 1]> var_46_cast_fp16 = slice_by_index(begin = var_46_begin_0, end = var_46_end_0, end_mask = var_46_end_mask_0, x = gates_1_cast_fp16)[name = string("op_46_cast_fp16")];39            tensor<fp16, [1, 640, 1, 1]> f_1_cast_fp16 = sigmoid(x = var_46_cast_fp16)[name = string("f_1_cast_fp16")];40            tensor<int32, [4]> var_49_begin_0 = const()[name = string("op_49_begin_0"), val = tensor<int32, [4]>([0, 1280, 0, 0])];41            tensor<int32, [4]> var_49_end_0 = const()[name = string("op_49_end_0"), val = tensor<int32, [4]>([1, 1920, 1, 1])];42            tensor<bool, [4]> var_49_end_mask_0 = const()[name = string("op_49_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];43            tensor<fp16, [1, 640, 1, 1]> var_49_cast_fp16 = slice_by_index(begin = var_49_begin_0, end = var_49_end_0, end_mask = var_49_end_mask_0, x = gates_1_cast_fp16)[name = string("op_49_cast_fp16")];44            tensor<fp16, [1, 640, 1, 1]> g_1_cast_fp16 = tanh(x = var_49_cast_fp16)[name = string("g_1_cast_fp16")];45            tensor<int32, [4]> var_52_begin_0 = const()[name = string("op_52_begin_0"), val = tensor<int32, [4]>([0, 1920, 0, 0])];46            tensor<int32, [4]> var_52_end_0 = const()[name = string("op_52_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];47            tensor<bool, [4]> var_52_end_mask_0 = const()[name = string("op_52_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];48            tensor<fp16, [1, 640, 1, 1]> var_52_cast_fp16 = slice_by_index(begin = var_52_begin_0, end = var_52_end_0, end_mask = var_52_end_mask_0, x = gates_1_cast_fp16)[name = string("op_52_cast_fp16")];49            tensor<fp16, [1, 640, 1, 1]> o_1_cast_fp16 = sigmoid(x = var_52_cast_fp16)[name = string("o_1_cast_fp16")];50            tensor<fp16, [1, 640, 1, 1]> var_54_cast_fp16 = mul(x = f_1_cast_fp16, y = c_1_cast_fp16)[name = string("op_54_cast_fp16")];51            tensor<fp16, [1, 640, 1, 1]> var_55_cast_fp16 = mul(x = i_1_cast_fp16, y = g_1_cast_fp16)[name = string("op_55_cast_fp16")];52            tensor<fp16, [1, 640, 1, 1]> c_new_1_cast_fp16 = add(x = var_54_cast_fp16, y = var_55_cast_fp16)[name = string("c_new_1_cast_fp16")];53            tensor<fp16, [1, 640, 1, 1]> var_57_cast_fp16 = tanh(x = c_new_1_cast_fp16)[name = string("op_57_cast_fp16")];54            tensor<fp16, [1, 640, 1, 1]> input_3_cast_fp16 = mul(x = o_1_cast_fp16, y = var_57_cast_fp16)[name = string("input_3_cast_fp16")];55            tensor<int32, [4]> input_5_begin_0 = const()[name = string("input_5_begin_0"), val = tensor<int32, [4]>([0, 640, 0, 0])];56            tensor<int32, [4]> input_5_end_0 = const()[name = string("input_5_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];57            tensor<bool, [4]> input_5_end_mask_0 = const()[name = string("input_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];58            tensor<fp16, [1, 640, 1, 1]> input_5_cast_fp16 = slice_by_index(begin = input_5_begin_0, end = input_5_end_0, end_mask = input_5_end_mask_0, x = h_in)[name = string("input_5_cast_fp16")];59            tensor<int32, [4]> c_begin_0 = const()[name = string("c_begin_0"), val = tensor<int32, [4]>([0, 640, 0, 0])];60            tensor<int32, [4]> c_end_0 = const()[name = string("c_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];61            tensor<bool, [4]> c_end_mask_0 = const()[name = string("c_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];62            tensor<fp16, [1, 640, 1, 1]> c_cast_fp16 = slice_by_index(begin = c_begin_0, end = c_end_0, end_mask = c_end_mask_0, x = c_in)[name = string("c_cast_fp16")];63            string var_74_pad_type_0 = const()[name = string("op_74_pad_type_0"), val = string("valid")];64            tensor<int32, [2]> var_74_strides_0 = const()[name = string("op_74_strides_0"), val = tensor<int32, [2]>([1, 1])];65            tensor<int32, [4]> var_74_pad_0 = const()[name = string("op_74_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];66            tensor<int32, [2]> var_74_dilations_0 = const()[name = string("op_74_dilations_0"), val = tensor<int32, [2]>([1, 1])];67            int32 var_74_groups_0 = const()[name = string("op_74_groups_0"), val = int32(1)];68            tensor<fp16, [2560, 640, 1, 1]> decoder_cells_1_ih_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2560, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3292544))), scale = tensor<fp16, [2560, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4931008))))[name = string("decoder_cells_1_ih_weight_to_fp16_quantized")];69            tensor<fp16, [2560]> decoder_cells_1_ih_bias_to_fp16 = const()[name = string("decoder_cells_1_ih_bias_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4936192)))];70            tensor<fp16, [1, 2560, 1, 1]> var_74_cast_fp16 = conv(bias = decoder_cells_1_ih_bias_to_fp16, dilations = var_74_dilations_0, groups = var_74_groups_0, pad = var_74_pad_0, pad_type = var_74_pad_type_0, strides = var_74_strides_0, weight = decoder_cells_1_ih_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("op_74_cast_fp16")];71            string var_80_pad_type_0 = const()[name = string("op_80_pad_type_0"), val = string("valid")];72            tensor<int32, [2]> var_80_strides_0 = const()[name = string("op_80_strides_0"), val = tensor<int32, [2]>([1, 1])];73            tensor<int32, [4]> var_80_pad_0 = const()[name = string("op_80_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];74            tensor<int32, [2]> var_80_dilations_0 = const()[name = string("op_80_dilations_0"), val = tensor<int32, [2]>([1, 1])];75            int32 var_80_groups_0 = const()[name = string("op_80_groups_0"), val = int32(1)];76            tensor<fp16, [2560, 640, 1, 1]> decoder_cells_1_hh_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2560, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4941376))), scale = tensor<fp16, [2560, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6579840))))[name = string("decoder_cells_1_hh_weight_to_fp16_quantized")];77            tensor<fp16, [1, 2560, 1, 1]> var_80_cast_fp16 = conv(dilations = var_80_dilations_0, groups = var_80_groups_0, pad = var_80_pad_0, pad_type = var_80_pad_type_0, strides = var_80_strides_0, weight = decoder_cells_1_hh_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("op_80_cast_fp16")];78            tensor<fp16, [1, 2560, 1, 1]> gates_cast_fp16 = add(x = var_74_cast_fp16, y = var_80_cast_fp16)[name = string("gates_cast_fp16")];79            tensor<int32, [4]> var_83_begin_0 = const()[name = string("op_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];80            tensor<int32, [4]> var_83_end_0 = const()[name = string("op_83_end_0"), val = tensor<int32, [4]>([1, 640, 1, 1])];81            tensor<bool, [4]> var_83_end_mask_0 = const()[name = string("op_83_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];82            tensor<fp16, [1, 640, 1, 1]> var_83_cast_fp16 = slice_by_index(begin = var_83_begin_0, end = var_83_end_0, end_mask = var_83_end_mask_0, x = gates_cast_fp16)[name = string("op_83_cast_fp16")];83            tensor<fp16, [1, 640, 1, 1]> i_cast_fp16 = sigmoid(x = var_83_cast_fp16)[name = string("i_cast_fp16")];84            tensor<int32, [4]> var_86_begin_0 = const()[name = string("op_86_begin_0"), val = tensor<int32, [4]>([0, 640, 0, 0])];85            tensor<int32, [4]> var_86_end_0 = const()[name = string("op_86_end_0"), val = tensor<int32, [4]>([1, 1280, 1, 1])];86            tensor<bool, [4]> var_86_end_mask_0 = const()[name = string("op_86_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];87            tensor<fp16, [1, 640, 1, 1]> var_86_cast_fp16 = slice_by_index(begin = var_86_begin_0, end = var_86_end_0, end_mask = var_86_end_mask_0, x = gates_cast_fp16)[name = string("op_86_cast_fp16")];88            tensor<fp16, [1, 640, 1, 1]> f_cast_fp16 = sigmoid(x = var_86_cast_fp16)[name = string("f_cast_fp16")];89            tensor<int32, [4]> var_89_begin_0 = const()[name = string("op_89_begin_0"), val = tensor<int32, [4]>([0, 1280, 0, 0])];90            tensor<int32, [4]> var_89_end_0 = const()[name = string("op_89_end_0"), val = tensor<int32, [4]>([1, 1920, 1, 1])];91            tensor<bool, [4]> var_89_end_mask_0 = const()[name = string("op_89_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];92            tensor<fp16, [1, 640, 1, 1]> var_89_cast_fp16 = slice_by_index(begin = var_89_begin_0, end = var_89_end_0, end_mask = var_89_end_mask_0, x = gates_cast_fp16)[name = string("op_89_cast_fp16")];93            tensor<fp16, [1, 640, 1, 1]> g_cast_fp16 = tanh(x = var_89_cast_fp16)[name = string("g_cast_fp16")];94            tensor<int32, [4]> var_92_begin_0 = const()[name = string("op_92_begin_0"), val = tensor<int32, [4]>([0, 1920, 0, 0])];95            tensor<int32, [4]> var_92_end_0 = const()[name = string("op_92_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];96            tensor<bool, [4]> var_92_end_mask_0 = const()[name = string("op_92_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];97            tensor<fp16, [1, 640, 1, 1]> var_92_cast_fp16 = slice_by_index(begin = var_92_begin_0, end = var_92_end_0, end_mask = var_92_end_mask_0, x = gates_cast_fp16)[name = string("op_92_cast_fp16")];98            tensor<fp16, [1, 640, 1, 1]> o_cast_fp16 = sigmoid(x = var_92_cast_fp16)[name = string("o_cast_fp16")];99            tensor<fp16, [1, 640, 1, 1]> var_94_cast_fp16 = mul(x = f_cast_fp16, y = c_cast_fp16)[name = string("op_94_cast_fp16")];100            tensor<fp16, [1, 640, 1, 1]> var_95_cast_fp16 = mul(x = i_cast_fp16, y = g_cast_fp16)[name = string("op_95_cast_fp16")];101            tensor<fp16, [1, 640, 1, 1]> c_new_cast_fp16 = add(x = var_94_cast_fp16, y = var_95_cast_fp16)[name = string("c_new_cast_fp16")];102            tensor<fp16, [1, 640, 1, 1]> var_97_cast_fp16 = tanh(x = c_new_cast_fp16)[name = string("op_97_cast_fp16")];103            tensor<fp16, [1, 640, 1, 1]> input_7_cast_fp16 = mul(x = o_cast_fp16, y = var_97_cast_fp16)[name = string("input_7_cast_fp16")];104            string pred_pad_type_0 = const()[name = string("pred_pad_type_0"), val = string("valid")];105            tensor<int32, [2]> pred_strides_0 = const()[name = string("pred_strides_0"), val = tensor<int32, [2]>([1, 1])];106            tensor<int32, [4]> pred_pad_0 = const()[name = string("pred_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];107            tensor<int32, [2]> pred_dilations_0 = const()[name = string("pred_dilations_0"), val = tensor<int32, [2]>([1, 1])];108            int32 pred_groups_0 = const()[name = string("pred_groups_0"), val = int32(1)];109            tensor<fp16, [640, 640, 1, 1]> decoder_joint_pred_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [640, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6585024))), scale = tensor<fp16, [640, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6994688))))[name = string("decoder_joint_pred_weight_to_fp16_quantized")];110            tensor<fp16, [640]> decoder_joint_pred_bias_to_fp16 = const()[name = string("decoder_joint_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6996032)))];111            tensor<fp16, [1, 640, 1, 1]> pred_cast_fp16 = conv(bias = decoder_joint_pred_bias_to_fp16, dilations = pred_dilations_0, groups = pred_groups_0, pad = pred_pad_0, pad_type = pred_pad_type_0, strides = pred_strides_0, weight = decoder_joint_pred_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("pred_cast_fp16")];112            bool var_110_interleave_0 = const()[name = string("op_110_interleave_0"), val = bool(false)];113            tensor<fp16, [1, 1280, 1, 1]> h_out = concat(axis = var_14, interleave = var_110_interleave_0, values = (input_3_cast_fp16, input_7_cast_fp16))[name = string("op_110_cast_fp16")];114            bool var_112_interleave_0 = const()[name = string("op_112_interleave_0"), val = bool(false)];115            tensor<fp16, [1, 1280, 1, 1]> c_out = concat(axis = var_14, interleave = var_112_interleave_0, values = (c_new_1_cast_fp16, c_new_cast_fp16))[name = string("op_112_cast_fp16")];116            tensor<fp16, [1, 640, 1, 8]> var_122_cast_fp16 = add(x = enc_step, y = pred_cast_fp16)[name = string("op_122_cast_fp16")];117            tensor<fp16, [1, 640, 1, 8]> input_cast_fp16 = relu(x = var_122_cast_fp16)[name = string("input_cast_fp16")];118            string x_pad_type_0 = const()[name = string("x_pad_type_0"), val = string("valid")];119            tensor<int32, [2]> x_strides_0 = const()[name = string("x_strides_0"), val = tensor<int32, [2]>([1, 1])];120            tensor<int32, [4]> x_pad_0 = const()[name = string("x_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];121            tensor<int32, [2]> x_dilations_0 = const()[name = string("x_dilations_0"), val = tensor<int32, [2]>([1, 1])];122            int32 x_groups_0 = const()[name = string("x_groups_0"), val = int32(1)];123            tensor<fp16, [8198, 640, 1, 1]> joint_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [8198, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6997376))), scale = tensor<fp16, [8198, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12244160))))[name = string("joint_out_weight_to_fp16_quantized")];124            tensor<fp16, [8198]> joint_out_bias_to_fp16 = const()[name = string("joint_out_bias_to_fp16"), val = tensor<fp16, [8198]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12260672)))];125            tensor<fp16, [1, 8198, 1, 8]> logits = conv(bias = joint_out_bias_to_fp16, dilations = x_dilations_0, groups = x_groups_0, pad = x_pad_0, pad_type = x_pad_type_0, strides = x_strides_0, weight = joint_out_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("x_cast_fp16")];126        } -> (logits, h_out, c_out);127    func decoder_2<ios18>(tensor<fp16, [2, 1280, 1, 1]> c_in, tensor<fp16, [2, 640, 1, 1]> embed, tensor<fp16, [2, 640, 1, 8]> enc_step, tensor<fp16, [2, 1280, 1, 1]> h_in) {128            int32 var_14 = const()[name = string("op_14"), val = int32(1)];129            tensor<int32, [4]> input_1_begin_0 = const()[name = string("input_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];130            tensor<int32, [4]> input_1_end_0 = const()[name = string("input_1_end_0"), val = tensor<int32, [4]>([2, 640, 1, 1])];131            tensor<bool, [4]> input_1_end_mask_0 = const()[name = string("input_1_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];132            tensor<fp16, [2, 640, 1, 1]> input_1_cast_fp16 = slice_by_index(begin = input_1_begin_0, end = input_1_end_0, end_mask = input_1_end_mask_0, x = h_in)[name = string("input_1_cast_fp16")];133            tensor<int32, [4]> c_1_begin_0 = const()[name = string("c_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];134            tensor<int32, [4]> c_1_end_0 = const()[name = string("c_1_end_0"), val = tensor<int32, [4]>([2, 640, 1, 1])];135            tensor<bool, [4]> c_1_end_mask_0 = const()[name = string("c_1_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];136            tensor<fp16, [2, 640, 1, 1]> c_1_cast_fp16 = slice_by_index(begin = c_1_begin_0, end = c_1_end_0, end_mask = c_1_end_mask_0, x = c_in)[name = string("c_1_cast_fp16")];137            string var_34_pad_type_0 = const()[name = string("op_34_pad_type_0"), val = string("valid")];138            tensor<int32, [2]> var_34_strides_0 = const()[name = string("op_34_strides_0"), val = tensor<int32, [2]>([1, 1])];139            tensor<int32, [4]> var_34_pad_0 = const()[name = string("op_34_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];140            tensor<int32, [2]> var_34_dilations_0 = const()[name = string("op_34_dilations_0"), val = tensor<int32, [2]>([1, 1])];141            int32 var_34_groups_0 = const()[name = string("op_34_groups_0"), val = int32(1)];142            tensor<fp16, [2560, 640, 1, 1]> decoder_cells_0_ih_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2560, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [2560, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1638528))))[name = string("decoder_cells_0_ih_weight_to_fp16_quantized")];143            tensor<fp16, [2560]> decoder_cells_0_ih_bias_to_fp16 = const()[name = string("decoder_cells_0_ih_bias_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1643712)))];144            tensor<fp16, [2, 2560, 1, 1]> var_34_cast_fp16 = conv(bias = decoder_cells_0_ih_bias_to_fp16, dilations = var_34_dilations_0, groups = var_34_groups_0, pad = var_34_pad_0, pad_type = var_34_pad_type_0, strides = var_34_strides_0, weight = decoder_cells_0_ih_weight_to_fp16_quantized, x = embed)[name = string("op_34_cast_fp16")];145            string var_40_pad_type_0 = const()[name = string("op_40_pad_type_0"), val = string("valid")];146            tensor<int32, [2]> var_40_strides_0 = const()[name = string("op_40_strides_0"), val = tensor<int32, [2]>([1, 1])];147            tensor<int32, [4]> var_40_pad_0 = const()[name = string("op_40_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];148            tensor<int32, [2]> var_40_dilations_0 = const()[name = string("op_40_dilations_0"), val = tensor<int32, [2]>([1, 1])];149            int32 var_40_groups_0 = const()[name = string("op_40_groups_0"), val = int32(1)];150            tensor<fp16, [2560, 640, 1, 1]> decoder_cells_0_hh_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2560, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1648896))), scale = tensor<fp16, [2560, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3287360))))[name = string("decoder_cells_0_hh_weight_to_fp16_quantized")];151            tensor<fp16, [2, 2560, 1, 1]> var_40_cast_fp16 = conv(dilations = var_40_dilations_0, groups = var_40_groups_0, pad = var_40_pad_0, pad_type = var_40_pad_type_0, strides = var_40_strides_0, weight = decoder_cells_0_hh_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("op_40_cast_fp16")];152            tensor<fp16, [2, 2560, 1, 1]> gates_1_cast_fp16 = add(x = var_34_cast_fp16, y = var_40_cast_fp16)[name = string("gates_1_cast_fp16")];153            tensor<int32, [4]> var_43_begin_0 = const()[name = string("op_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];154            tensor<int32, [4]> var_43_end_0 = const()[name = string("op_43_end_0"), val = tensor<int32, [4]>([2, 640, 1, 1])];155            tensor<bool, [4]> var_43_end_mask_0 = const()[name = string("op_43_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];156            tensor<fp16, [2, 640, 1, 1]> var_43_cast_fp16 = slice_by_index(begin = var_43_begin_0, end = var_43_end_0, end_mask = var_43_end_mask_0, x = gates_1_cast_fp16)[name = string("op_43_cast_fp16")];157            tensor<fp16, [2, 640, 1, 1]> i_1_cast_fp16 = sigmoid(x = var_43_cast_fp16)[name = string("i_1_cast_fp16")];158            tensor<int32, [4]> var_46_begin_0 = const()[name = string("op_46_begin_0"), val = tensor<int32, [4]>([0, 640, 0, 0])];159            tensor<int32, [4]> var_46_end_0 = const()[name = string("op_46_end_0"), val = tensor<int32, [4]>([2, 1280, 1, 1])];160            tensor<bool, [4]> var_46_end_mask_0 = const()[name = string("op_46_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];161            tensor<fp16, [2, 640, 1, 1]> var_46_cast_fp16 = slice_by_index(begin = var_46_begin_0, end = var_46_end_0, end_mask = var_46_end_mask_0, x = gates_1_cast_fp16)[name = string("op_46_cast_fp16")];162            tensor<fp16, [2, 640, 1, 1]> f_1_cast_fp16 = sigmoid(x = var_46_cast_fp16)[name = string("f_1_cast_fp16")];163            tensor<int32, [4]> var_49_begin_0 = const()[name = string("op_49_begin_0"), val = tensor<int32, [4]>([0, 1280, 0, 0])];164            tensor<int32, [4]> var_49_end_0 = const()[name = string("op_49_end_0"), val = tensor<int32, [4]>([2, 1920, 1, 1])];165            tensor<bool, [4]> var_49_end_mask_0 = const()[name = string("op_49_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];166            tensor<fp16, [2, 640, 1, 1]> var_49_cast_fp16 = slice_by_index(begin = var_49_begin_0, end = var_49_end_0, end_mask = var_49_end_mask_0, x = gates_1_cast_fp16)[name = string("op_49_cast_fp16")];167            tensor<fp16, [2, 640, 1, 1]> g_1_cast_fp16 = tanh(x = var_49_cast_fp16)[name = string("g_1_cast_fp16")];168            tensor<int32, [4]> var_52_begin_0 = const()[name = string("op_52_begin_0"), val = tensor<int32, [4]>([0, 1920, 0, 0])];169            tensor<int32, [4]> var_52_end_0 = const()[name = string("op_52_end_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];170            tensor<bool, [4]> var_52_end_mask_0 = const()[name = string("op_52_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];171            tensor<fp16, [2, 640, 1, 1]> var_52_cast_fp16 = slice_by_index(begin = var_52_begin_0, end = var_52_end_0, end_mask = var_52_end_mask_0, x = gates_1_cast_fp16)[name = string("op_52_cast_fp16")];172            tensor<fp16, [2, 640, 1, 1]> o_1_cast_fp16 = sigmoid(x = var_52_cast_fp16)[name = string("o_1_cast_fp16")];173            tensor<fp16, [2, 640, 1, 1]> var_54_cast_fp16 = mul(x = f_1_cast_fp16, y = c_1_cast_fp16)[name = string("op_54_cast_fp16")];174            tensor<fp16, [2, 640, 1, 1]> var_55_cast_fp16 = mul(x = i_1_cast_fp16, y = g_1_cast_fp16)[name = string("op_55_cast_fp16")];175            tensor<fp16, [2, 640, 1, 1]> c_new_1_cast_fp16 = add(x = var_54_cast_fp16, y = var_55_cast_fp16)[name = string("c_new_1_cast_fp16")];176            tensor<fp16, [2, 640, 1, 1]> var_57_cast_fp16 = tanh(x = c_new_1_cast_fp16)[name = string("op_57_cast_fp16")];177            tensor<fp16, [2, 640, 1, 1]> input_3_cast_fp16 = mul(x = o_1_cast_fp16, y = var_57_cast_fp16)[name = string("input_3_cast_fp16")];178            tensor<int32, [4]> input_5_begin_0 = const()[name = string("input_5_begin_0"), val = tensor<int32, [4]>([0, 640, 0, 0])];179            tensor<int32, [4]> input_5_end_0 = const()[name = string("input_5_end_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];180            tensor<bool, [4]> input_5_end_mask_0 = const()[name = string("input_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];181            tensor<fp16, [2, 640, 1, 1]> input_5_cast_fp16 = slice_by_index(begin = input_5_begin_0, end = input_5_end_0, end_mask = input_5_end_mask_0, x = h_in)[name = string("input_5_cast_fp16")];182            tensor<int32, [4]> c_begin_0 = const()[name = string("c_begin_0"), val = tensor<int32, [4]>([0, 640, 0, 0])];183            tensor<int32, [4]> c_end_0 = const()[name = string("c_end_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];184            tensor<bool, [4]> c_end_mask_0 = const()[name = string("c_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];185            tensor<fp16, [2, 640, 1, 1]> c_cast_fp16 = slice_by_index(begin = c_begin_0, end = c_end_0, end_mask = c_end_mask_0, x = c_in)[name = string("c_cast_fp16")];186            string var_74_pad_type_0 = const()[name = string("op_74_pad_type_0"), val = string("valid")];187            tensor<int32, [2]> var_74_strides_0 = const()[name = string("op_74_strides_0"), val = tensor<int32, [2]>([1, 1])];188            tensor<int32, [4]> var_74_pad_0 = const()[name = string("op_74_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];189            tensor<int32, [2]> var_74_dilations_0 = const()[name = string("op_74_dilations_0"), val = tensor<int32, [2]>([1, 1])];190            int32 var_74_groups_0 = const()[name = string("op_74_groups_0"), val = int32(1)];191            tensor<fp16, [2560, 640, 1, 1]> decoder_cells_1_ih_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2560, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3292544))), scale = tensor<fp16, [2560, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4931008))))[name = string("decoder_cells_1_ih_weight_to_fp16_quantized")];192            tensor<fp16, [2560]> decoder_cells_1_ih_bias_to_fp16 = const()[name = string("decoder_cells_1_ih_bias_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4936192)))];193            tensor<fp16, [2, 2560, 1, 1]> var_74_cast_fp16 = conv(bias = decoder_cells_1_ih_bias_to_fp16, dilations = var_74_dilations_0, groups = var_74_groups_0, pad = var_74_pad_0, pad_type = var_74_pad_type_0, strides = var_74_strides_0, weight = decoder_cells_1_ih_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("op_74_cast_fp16")];194            string var_80_pad_type_0 = const()[name = string("op_80_pad_type_0"), val = string("valid")];195            tensor<int32, [2]> var_80_strides_0 = const()[name = string("op_80_strides_0"), val = tensor<int32, [2]>([1, 1])];196            tensor<int32, [4]> var_80_pad_0 = const()[name = string("op_80_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];197            tensor<int32, [2]> var_80_dilations_0 = const()[name = string("op_80_dilations_0"), val = tensor<int32, [2]>([1, 1])];198            int32 var_80_groups_0 = const()[name = string("op_80_groups_0"), val = int32(1)];199            tensor<fp16, [2560, 640, 1, 1]> decoder_cells_1_hh_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2560, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4941376))), scale = tensor<fp16, [2560, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6579840))))[name = string("decoder_cells_1_hh_weight_to_fp16_quantized")];200            tensor<fp16, [2, 2560, 1, 1]> var_80_cast_fp16 = conv(dilations = var_80_dilations_0, groups = var_80_groups_0, pad = var_80_pad_0, pad_type = var_80_pad_type_0, strides = var_80_strides_0, weight = decoder_cells_1_hh_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("op_80_cast_fp16")];201            tensor<fp16, [2, 2560, 1, 1]> gates_cast_fp16 = add(x = var_74_cast_fp16, y = var_80_cast_fp16)[name = string("gates_cast_fp16")];202            tensor<int32, [4]> var_83_begin_0 = const()[name = string("op_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];203            tensor<int32, [4]> var_83_end_0 = const()[name = string("op_83_end_0"), val = tensor<int32, [4]>([2, 640, 1, 1])];204            tensor<bool, [4]> var_83_end_mask_0 = const()[name = string("op_83_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];205            tensor<fp16, [2, 640, 1, 1]> var_83_cast_fp16 = slice_by_index(begin = var_83_begin_0, end = var_83_end_0, end_mask = var_83_end_mask_0, x = gates_cast_fp16)[name = string("op_83_cast_fp16")];206            tensor<fp16, [2, 640, 1, 1]> i_cast_fp16 = sigmoid(x = var_83_cast_fp16)[name = string("i_cast_fp16")];207            tensor<int32, [4]> var_86_begin_0 = const()[name = string("op_86_begin_0"), val = tensor<int32, [4]>([0, 640, 0, 0])];208            tensor<int32, [4]> var_86_end_0 = const()[name = string("op_86_end_0"), val = tensor<int32, [4]>([2, 1280, 1, 1])];209            tensor<bool, [4]> var_86_end_mask_0 = const()[name = string("op_86_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];210            tensor<fp16, [2, 640, 1, 1]> var_86_cast_fp16 = slice_by_index(begin = var_86_begin_0, end = var_86_end_0, end_mask = var_86_end_mask_0, x = gates_cast_fp16)[name = string("op_86_cast_fp16")];211            tensor<fp16, [2, 640, 1, 1]> f_cast_fp16 = sigmoid(x = var_86_cast_fp16)[name = string("f_cast_fp16")];212            tensor<int32, [4]> var_89_begin_0 = const()[name = string("op_89_begin_0"), val = tensor<int32, [4]>([0, 1280, 0, 0])];213            tensor<int32, [4]> var_89_end_0 = const()[name = string("op_89_end_0"), val = tensor<int32, [4]>([2, 1920, 1, 1])];214            tensor<bool, [4]> var_89_end_mask_0 = const()[name = string("op_89_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];215            tensor<fp16, [2, 640, 1, 1]> var_89_cast_fp16 = slice_by_index(begin = var_89_begin_0, end = var_89_end_0, end_mask = var_89_end_mask_0, x = gates_cast_fp16)[name = string("op_89_cast_fp16")];216            tensor<fp16, [2, 640, 1, 1]> g_cast_fp16 = tanh(x = var_89_cast_fp16)[name = string("g_cast_fp16")];217            tensor<int32, [4]> var_92_begin_0 = const()[name = string("op_92_begin_0"), val = tensor<int32, [4]>([0, 1920, 0, 0])];218            tensor<int32, [4]> var_92_end_0 = const()[name = string("op_92_end_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];219            tensor<bool, [4]> var_92_end_mask_0 = const()[name = string("op_92_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];220            tensor<fp16, [2, 640, 1, 1]> var_92_cast_fp16 = slice_by_index(begin = var_92_begin_0, end = var_92_end_0, end_mask = var_92_end_mask_0, x = gates_cast_fp16)[name = string("op_92_cast_fp16")];221            tensor<fp16, [2, 640, 1, 1]> o_cast_fp16 = sigmoid(x = var_92_cast_fp16)[name = string("o_cast_fp16")];222            tensor<fp16, [2, 640, 1, 1]> var_94_cast_fp16 = mul(x = f_cast_fp16, y = c_cast_fp16)[name = string("op_94_cast_fp16")];223            tensor<fp16, [2, 640, 1, 1]> var_95_cast_fp16 = mul(x = i_cast_fp16, y = g_cast_fp16)[name = string("op_95_cast_fp16")];224            tensor<fp16, [2, 640, 1, 1]> c_new_cast_fp16 = add(x = var_94_cast_fp16, y = var_95_cast_fp16)[name = string("c_new_cast_fp16")];225            tensor<fp16, [2, 640, 1, 1]> var_97_cast_fp16 = tanh(x = c_new_cast_fp16)[name = string("op_97_cast_fp16")];226            tensor<fp16, [2, 640, 1, 1]> input_7_cast_fp16 = mul(x = o_cast_fp16, y = var_97_cast_fp16)[name = string("input_7_cast_fp16")];227            string pred_pad_type_0 = const()[name = string("pred_pad_type_0"), val = string("valid")];228            tensor<int32, [2]> pred_strides_0 = const()[name = string("pred_strides_0"), val = tensor<int32, [2]>([1, 1])];229            tensor<int32, [4]> pred_pad_0 = const()[name = string("pred_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];230            tensor<int32, [2]> pred_dilations_0 = const()[name = string("pred_dilations_0"), val = tensor<int32, [2]>([1, 1])];231            int32 pred_groups_0 = const()[name = string("pred_groups_0"), val = int32(1)];232            tensor<fp16, [640, 640, 1, 1]> decoder_joint_pred_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [640, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6585024))), scale = tensor<fp16, [640, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6994688))))[name = string("decoder_joint_pred_weight_to_fp16_quantized")];233            tensor<fp16, [640]> decoder_joint_pred_bias_to_fp16 = const()[name = string("decoder_joint_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6996032)))];234            tensor<fp16, [2, 640, 1, 1]> pred_cast_fp16 = conv(bias = decoder_joint_pred_bias_to_fp16, dilations = pred_dilations_0, groups = pred_groups_0, pad = pred_pad_0, pad_type = pred_pad_type_0, strides = pred_strides_0, weight = decoder_joint_pred_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("pred_cast_fp16")];235            bool var_110_interleave_0 = const()[name = string("op_110_interleave_0"), val = bool(false)];236            tensor<fp16, [2, 1280, 1, 1]> h_out = concat(axis = var_14, interleave = var_110_interleave_0, values = (input_3_cast_fp16, input_7_cast_fp16))[name = string("op_110_cast_fp16")];237            bool var_112_interleave_0 = const()[name = string("op_112_interleave_0"), val = bool(false)];238            tensor<fp16, [2, 1280, 1, 1]> c_out = concat(axis = var_14, interleave = var_112_interleave_0, values = (c_new_1_cast_fp16, c_new_cast_fp16))[name = string("op_112_cast_fp16")];239            tensor<fp16, [2, 640, 1, 8]> var_122_cast_fp16 = add(x = enc_step, y = pred_cast_fp16)[name = string("op_122_cast_fp16")];240            tensor<fp16, [2, 640, 1, 8]> input_cast_fp16 = relu(x = var_122_cast_fp16)[name = string("input_cast_fp16")];241            string x_pad_type_0 = const()[name = string("x_pad_type_0"), val = string("valid")];242            tensor<int32, [2]> x_strides_0 = const()[name = string("x_strides_0"), val = tensor<int32, [2]>([1, 1])];243            tensor<int32, [4]> x_pad_0 = const()[name = string("x_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];244            tensor<int32, [2]> x_dilations_0 = const()[name = string("x_dilations_0"), val = tensor<int32, [2]>([1, 1])];245            int32 x_groups_0 = const()[name = string("x_groups_0"), val = int32(1)];246            tensor<fp16, [8198, 640, 1, 1]> joint_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [8198, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6997376))), scale = tensor<fp16, [8198, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12244160))))[name = string("joint_out_weight_to_fp16_quantized")];247            tensor<fp16, [8198]> joint_out_bias_to_fp16 = const()[name = string("joint_out_bias_to_fp16"), val = tensor<fp16, [8198]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12260672)))];248            tensor<fp16, [2, 8198, 1, 8]> logits = conv(bias = joint_out_bias_to_fp16, dilations = x_dilations_0, groups = x_groups_0, pad = x_pad_0, pad_type = x_pad_type_0, strides = x_strides_0, weight = joint_out_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("x_cast_fp16")];249        } -> (logits, h_out, c_out);250    func decoder_4<ios18>(tensor<fp16, [4, 1280, 1, 1]> c_in, tensor<fp16, [4, 640, 1, 1]> embed, tensor<fp16, [4, 640, 1, 8]> enc_step, tensor<fp16, [4, 1280, 1, 1]> h_in) {251            int32 var_14 = const()[name = string("op_14"), val = int32(1)];252            tensor<int32, [4]> input_1_begin_0 = const()[name = string("input_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];253            tensor<int32, [4]> input_1_end_0 = const()[name = string("input_1_end_0"), val = tensor<int32, [4]>([4, 640, 1, 1])];254            tensor<bool, [4]> input_1_end_mask_0 = const()[name = string("input_1_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];255            tensor<fp16, [4, 640, 1, 1]> input_1_cast_fp16 = slice_by_index(begin = input_1_begin_0, end = input_1_end_0, end_mask = input_1_end_mask_0, x = h_in)[name = string("input_1_cast_fp16")];256            tensor<int32, [4]> c_1_begin_0 = const()[name = string("c_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];257            tensor<int32, [4]> c_1_end_0 = const()[name = string("c_1_end_0"), val = tensor<int32, [4]>([4, 640, 1, 1])];258            tensor<bool, [4]> c_1_end_mask_0 = const()[name = string("c_1_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];259            tensor<fp16, [4, 640, 1, 1]> c_1_cast_fp16 = slice_by_index(begin = c_1_begin_0, end = c_1_end_0, end_mask = c_1_end_mask_0, x = c_in)[name = string("c_1_cast_fp16")];260            string var_34_pad_type_0 = const()[name = string("op_34_pad_type_0"), val = string("valid")];261            tensor<int32, [2]> var_34_strides_0 = const()[name = string("op_34_strides_0"), val = tensor<int32, [2]>([1, 1])];262            tensor<int32, [4]> var_34_pad_0 = const()[name = string("op_34_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];263            tensor<int32, [2]> var_34_dilations_0 = const()[name = string("op_34_dilations_0"), val = tensor<int32, [2]>([1, 1])];264            int32 var_34_groups_0 = const()[name = string("op_34_groups_0"), val = int32(1)];265            tensor<fp16, [2560, 640, 1, 1]> decoder_cells_0_ih_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2560, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [2560, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1638528))))[name = string("decoder_cells_0_ih_weight_to_fp16_quantized")];266            tensor<fp16, [2560]> decoder_cells_0_ih_bias_to_fp16 = const()[name = string("decoder_cells_0_ih_bias_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1643712)))];267            tensor<fp16, [4, 2560, 1, 1]> var_34_cast_fp16 = conv(bias = decoder_cells_0_ih_bias_to_fp16, dilations = var_34_dilations_0, groups = var_34_groups_0, pad = var_34_pad_0, pad_type = var_34_pad_type_0, strides = var_34_strides_0, weight = decoder_cells_0_ih_weight_to_fp16_quantized, x = embed)[name = string("op_34_cast_fp16")];268            string var_40_pad_type_0 = const()[name = string("op_40_pad_type_0"), val = string("valid")];269            tensor<int32, [2]> var_40_strides_0 = const()[name = string("op_40_strides_0"), val = tensor<int32, [2]>([1, 1])];270            tensor<int32, [4]> var_40_pad_0 = const()[name = string("op_40_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];271            tensor<int32, [2]> var_40_dilations_0 = const()[name = string("op_40_dilations_0"), val = tensor<int32, [2]>([1, 1])];272            int32 var_40_groups_0 = const()[name = string("op_40_groups_0"), val = int32(1)];273            tensor<fp16, [2560, 640, 1, 1]> decoder_cells_0_hh_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2560, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1648896))), scale = tensor<fp16, [2560, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3287360))))[name = string("decoder_cells_0_hh_weight_to_fp16_quantized")];274            tensor<fp16, [4, 2560, 1, 1]> var_40_cast_fp16 = conv(dilations = var_40_dilations_0, groups = var_40_groups_0, pad = var_40_pad_0, pad_type = var_40_pad_type_0, strides = var_40_strides_0, weight = decoder_cells_0_hh_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("op_40_cast_fp16")];275            tensor<fp16, [4, 2560, 1, 1]> gates_1_cast_fp16 = add(x = var_34_cast_fp16, y = var_40_cast_fp16)[name = string("gates_1_cast_fp16")];276            tensor<int32, [4]> var_43_begin_0 = const()[name = string("op_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];277            tensor<int32, [4]> var_43_end_0 = const()[name = string("op_43_end_0"), val = tensor<int32, [4]>([4, 640, 1, 1])];278            tensor<bool, [4]> var_43_end_mask_0 = const()[name = string("op_43_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];279            tensor<fp16, [4, 640, 1, 1]> var_43_cast_fp16 = slice_by_index(begin = var_43_begin_0, end = var_43_end_0, end_mask = var_43_end_mask_0, x = gates_1_cast_fp16)[name = string("op_43_cast_fp16")];280            tensor<fp16, [4, 640, 1, 1]> i_1_cast_fp16 = sigmoid(x = var_43_cast_fp16)[name = string("i_1_cast_fp16")];281            tensor<int32, [4]> var_46_begin_0 = const()[name = string("op_46_begin_0"), val = tensor<int32, [4]>([0, 640, 0, 0])];282            tensor<int32, [4]> var_46_end_0 = const()[name = string("op_46_end_0"), val = tensor<int32, [4]>([4, 1280, 1, 1])];283            tensor<bool, [4]> var_46_end_mask_0 = const()[name = string("op_46_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];284            tensor<fp16, [4, 640, 1, 1]> var_46_cast_fp16 = slice_by_index(begin = var_46_begin_0, end = var_46_end_0, end_mask = var_46_end_mask_0, x = gates_1_cast_fp16)[name = string("op_46_cast_fp16")];285            tensor<fp16, [4, 640, 1, 1]> f_1_cast_fp16 = sigmoid(x = var_46_cast_fp16)[name = string("f_1_cast_fp16")];286            tensor<int32, [4]> var_49_begin_0 = const()[name = string("op_49_begin_0"), val = tensor<int32, [4]>([0, 1280, 0, 0])];287            tensor<int32, [4]> var_49_end_0 = const()[name = string("op_49_end_0"), val = tensor<int32, [4]>([4, 1920, 1, 1])];288            tensor<bool, [4]> var_49_end_mask_0 = const()[name = string("op_49_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];289            tensor<fp16, [4, 640, 1, 1]> var_49_cast_fp16 = slice_by_index(begin = var_49_begin_0, end = var_49_end_0, end_mask = var_49_end_mask_0, x = gates_1_cast_fp16)[name = string("op_49_cast_fp16")];290            tensor<fp16, [4, 640, 1, 1]> g_1_cast_fp16 = tanh(x = var_49_cast_fp16)[name = string("g_1_cast_fp16")];291            tensor<int32, [4]> var_52_begin_0 = const()[name = string("op_52_begin_0"), val = tensor<int32, [4]>([0, 1920, 0, 0])];292            tensor<int32, [4]> var_52_end_0 = const()[name = string("op_52_end_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];293            tensor<bool, [4]> var_52_end_mask_0 = const()[name = string("op_52_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];294            tensor<fp16, [4, 640, 1, 1]> var_52_cast_fp16 = slice_by_index(begin = var_52_begin_0, end = var_52_end_0, end_mask = var_52_end_mask_0, x = gates_1_cast_fp16)[name = string("op_52_cast_fp16")];295            tensor<fp16, [4, 640, 1, 1]> o_1_cast_fp16 = sigmoid(x = var_52_cast_fp16)[name = string("o_1_cast_fp16")];296            tensor<fp16, [4, 640, 1, 1]> var_54_cast_fp16 = mul(x = f_1_cast_fp16, y = c_1_cast_fp16)[name = string("op_54_cast_fp16")];297            tensor<fp16, [4, 640, 1, 1]> var_55_cast_fp16 = mul(x = i_1_cast_fp16, y = g_1_cast_fp16)[name = string("op_55_cast_fp16")];298            tensor<fp16, [4, 640, 1, 1]> c_new_1_cast_fp16 = add(x = var_54_cast_fp16, y = var_55_cast_fp16)[name = string("c_new_1_cast_fp16")];299            tensor<fp16, [4, 640, 1, 1]> var_57_cast_fp16 = tanh(x = c_new_1_cast_fp16)[name = string("op_57_cast_fp16")];300            tensor<fp16, [4, 640, 1, 1]> input_3_cast_fp16 = mul(x = o_1_cast_fp16, y = var_57_cast_fp16)[name = string("input_3_cast_fp16")];301            tensor<int32, [4]> input_5_begin_0 = const()[name = string("input_5_begin_0"), val = tensor<int32, [4]>([0, 640, 0, 0])];302            tensor<int32, [4]> input_5_end_0 = const()[name = string("input_5_end_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];303            tensor<bool, [4]> input_5_end_mask_0 = const()[name = string("input_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];304            tensor<fp16, [4, 640, 1, 1]> input_5_cast_fp16 = slice_by_index(begin = input_5_begin_0, end = input_5_end_0, end_mask = input_5_end_mask_0, x = h_in)[name = string("input_5_cast_fp16")];305            tensor<int32, [4]> c_begin_0 = const()[name = string("c_begin_0"), val = tensor<int32, [4]>([0, 640, 0, 0])];306            tensor<int32, [4]> c_end_0 = const()[name = string("c_end_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];307            tensor<bool, [4]> c_end_mask_0 = const()[name = string("c_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];308            tensor<fp16, [4, 640, 1, 1]> c_cast_fp16 = slice_by_index(begin = c_begin_0, end = c_end_0, end_mask = c_end_mask_0, x = c_in)[name = string("c_cast_fp16")];309            string var_74_pad_type_0 = const()[name = string("op_74_pad_type_0"), val = string("valid")];310            tensor<int32, [2]> var_74_strides_0 = const()[name = string("op_74_strides_0"), val = tensor<int32, [2]>([1, 1])];311            tensor<int32, [4]> var_74_pad_0 = const()[name = string("op_74_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];312            tensor<int32, [2]> var_74_dilations_0 = const()[name = string("op_74_dilations_0"), val = tensor<int32, [2]>([1, 1])];313            int32 var_74_groups_0 = const()[name = string("op_74_groups_0"), val = int32(1)];314            tensor<fp16, [2560, 640, 1, 1]> decoder_cells_1_ih_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2560, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3292544))), scale = tensor<fp16, [2560, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4931008))))[name = string("decoder_cells_1_ih_weight_to_fp16_quantized")];315            tensor<fp16, [2560]> decoder_cells_1_ih_bias_to_fp16 = const()[name = string("decoder_cells_1_ih_bias_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4936192)))];316            tensor<fp16, [4, 2560, 1, 1]> var_74_cast_fp16 = conv(bias = decoder_cells_1_ih_bias_to_fp16, dilations = var_74_dilations_0, groups = var_74_groups_0, pad = var_74_pad_0, pad_type = var_74_pad_type_0, strides = var_74_strides_0, weight = decoder_cells_1_ih_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("op_74_cast_fp16")];317            string var_80_pad_type_0 = const()[name = string("op_80_pad_type_0"), val = string("valid")];318            tensor<int32, [2]> var_80_strides_0 = const()[name = string("op_80_strides_0"), val = tensor<int32, [2]>([1, 1])];319            tensor<int32, [4]> var_80_pad_0 = const()[name = string("op_80_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];320            tensor<int32, [2]> var_80_dilations_0 = const()[name = string("op_80_dilations_0"), val = tensor<int32, [2]>([1, 1])];321            int32 var_80_groups_0 = const()[name = string("op_80_groups_0"), val = int32(1)];322            tensor<fp16, [2560, 640, 1, 1]> decoder_cells_1_hh_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2560, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4941376))), scale = tensor<fp16, [2560, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6579840))))[name = string("decoder_cells_1_hh_weight_to_fp16_quantized")];323            tensor<fp16, [4, 2560, 1, 1]> var_80_cast_fp16 = conv(dilations = var_80_dilations_0, groups = var_80_groups_0, pad = var_80_pad_0, pad_type = var_80_pad_type_0, strides = var_80_strides_0, weight = decoder_cells_1_hh_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("op_80_cast_fp16")];324            tensor<fp16, [4, 2560, 1, 1]> gates_cast_fp16 = add(x = var_74_cast_fp16, y = var_80_cast_fp16)[name = string("gates_cast_fp16")];325            tensor<int32, [4]> var_83_begin_0 = const()[name = string("op_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];326            tensor<int32, [4]> var_83_end_0 = const()[name = string("op_83_end_0"), val = tensor<int32, [4]>([4, 640, 1, 1])];327            tensor<bool, [4]> var_83_end_mask_0 = const()[name = string("op_83_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];328            tensor<fp16, [4, 640, 1, 1]> var_83_cast_fp16 = slice_by_index(begin = var_83_begin_0, end = var_83_end_0, end_mask = var_83_end_mask_0, x = gates_cast_fp16)[name = string("op_83_cast_fp16")];329            tensor<fp16, [4, 640, 1, 1]> i_cast_fp16 = sigmoid(x = var_83_cast_fp16)[name = string("i_cast_fp16")];330            tensor<int32, [4]> var_86_begin_0 = const()[name = string("op_86_begin_0"), val = tensor<int32, [4]>([0, 640, 0, 0])];331            tensor<int32, [4]> var_86_end_0 = const()[name = string("op_86_end_0"), val = tensor<int32, [4]>([4, 1280, 1, 1])];332            tensor<bool, [4]> var_86_end_mask_0 = const()[name = string("op_86_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];333            tensor<fp16, [4, 640, 1, 1]> var_86_cast_fp16 = slice_by_index(begin = var_86_begin_0, end = var_86_end_0, end_mask = var_86_end_mask_0, x = gates_cast_fp16)[name = string("op_86_cast_fp16")];334            tensor<fp16, [4, 640, 1, 1]> f_cast_fp16 = sigmoid(x = var_86_cast_fp16)[name = string("f_cast_fp16")];335            tensor<int32, [4]> var_89_begin_0 = const()[name = string("op_89_begin_0"), val = tensor<int32, [4]>([0, 1280, 0, 0])];336            tensor<int32, [4]> var_89_end_0 = const()[name = string("op_89_end_0"), val = tensor<int32, [4]>([4, 1920, 1, 1])];337            tensor<bool, [4]> var_89_end_mask_0 = const()[name = string("op_89_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];338            tensor<fp16, [4, 640, 1, 1]> var_89_cast_fp16 = slice_by_index(begin = var_89_begin_0, end = var_89_end_0, end_mask = var_89_end_mask_0, x = gates_cast_fp16)[name = string("op_89_cast_fp16")];339            tensor<fp16, [4, 640, 1, 1]> g_cast_fp16 = tanh(x = var_89_cast_fp16)[name = string("g_cast_fp16")];340            tensor<int32, [4]> var_92_begin_0 = const()[name = string("op_92_begin_0"), val = tensor<int32, [4]>([0, 1920, 0, 0])];341            tensor<int32, [4]> var_92_end_0 = const()[name = string("op_92_end_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];342            tensor<bool, [4]> var_92_end_mask_0 = const()[name = string("op_92_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];343            tensor<fp16, [4, 640, 1, 1]> var_92_cast_fp16 = slice_by_index(begin = var_92_begin_0, end = var_92_end_0, end_mask = var_92_end_mask_0, x = gates_cast_fp16)[name = string("op_92_cast_fp16")];344            tensor<fp16, [4, 640, 1, 1]> o_cast_fp16 = sigmoid(x = var_92_cast_fp16)[name = string("o_cast_fp16")];345            tensor<fp16, [4, 640, 1, 1]> var_94_cast_fp16 = mul(x = f_cast_fp16, y = c_cast_fp16)[name = string("op_94_cast_fp16")];346            tensor<fp16, [4, 640, 1, 1]> var_95_cast_fp16 = mul(x = i_cast_fp16, y = g_cast_fp16)[name = string("op_95_cast_fp16")];347            tensor<fp16, [4, 640, 1, 1]> c_new_cast_fp16 = add(x = var_94_cast_fp16, y = var_95_cast_fp16)[name = string("c_new_cast_fp16")];348            tensor<fp16, [4, 640, 1, 1]> var_97_cast_fp16 = tanh(x = c_new_cast_fp16)[name = string("op_97_cast_fp16")];349            tensor<fp16, [4, 640, 1, 1]> input_7_cast_fp16 = mul(x = o_cast_fp16, y = var_97_cast_fp16)[name = string("input_7_cast_fp16")];350            string pred_pad_type_0 = const()[name = string("pred_pad_type_0"), val = string("valid")];351            tensor<int32, [2]> pred_strides_0 = const()[name = string("pred_strides_0"), val = tensor<int32, [2]>([1, 1])];352            tensor<int32, [4]> pred_pad_0 = const()[name = string("pred_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];353            tensor<int32, [2]> pred_dilations_0 = const()[name = string("pred_dilations_0"), val = tensor<int32, [2]>([1, 1])];354            int32 pred_groups_0 = const()[name = string("pred_groups_0"), val = int32(1)];355            tensor<fp16, [640, 640, 1, 1]> decoder_joint_pred_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [640, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6585024))), scale = tensor<fp16, [640, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6994688))))[name = string("decoder_joint_pred_weight_to_fp16_quantized")];356            tensor<fp16, [640]> decoder_joint_pred_bias_to_fp16 = const()[name = string("decoder_joint_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6996032)))];357            tensor<fp16, [4, 640, 1, 1]> pred_cast_fp16 = conv(bias = decoder_joint_pred_bias_to_fp16, dilations = pred_dilations_0, groups = pred_groups_0, pad = pred_pad_0, pad_type = pred_pad_type_0, strides = pred_strides_0, weight = decoder_joint_pred_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("pred_cast_fp16")];358            bool var_110_interleave_0 = const()[name = string("op_110_interleave_0"), val = bool(false)];359            tensor<fp16, [4, 1280, 1, 1]> h_out = concat(axis = var_14, interleave = var_110_interleave_0, values = (input_3_cast_fp16, input_7_cast_fp16))[name = string("op_110_cast_fp16")];360            bool var_112_interleave_0 = const()[name = string("op_112_interleave_0"), val = bool(false)];361            tensor<fp16, [4, 1280, 1, 1]> c_out = concat(axis = var_14, interleave = var_112_interleave_0, values = (c_new_1_cast_fp16, c_new_cast_fp16))[name = string("op_112_cast_fp16")];362            tensor<fp16, [4, 640, 1, 8]> var_122_cast_fp16 = add(x = enc_step, y = pred_cast_fp16)[name = string("op_122_cast_fp16")];363            tensor<fp16, [4, 640, 1, 8]> input_cast_fp16 = relu(x = var_122_cast_fp16)[name = string("input_cast_fp16")];364            string x_pad_type_0 = const()[name = string("x_pad_type_0"), val = string("valid")];365            tensor<int32, [2]> x_strides_0 = const()[name = string("x_strides_0"), val = tensor<int32, [2]>([1, 1])];366            tensor<int32, [4]> x_pad_0 = const()[name = string("x_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];367            tensor<int32, [2]> x_dilations_0 = const()[name = string("x_dilations_0"), val = tensor<int32, [2]>([1, 1])];368            int32 x_groups_0 = const()[name = string("x_groups_0"), val = int32(1)];369            tensor<fp16, [8198, 640, 1, 1]> joint_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [8198, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6997376))), scale = tensor<fp16, [8198, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12244160))))[name = string("joint_out_weight_to_fp16_quantized")];370            tensor<fp16, [8198]> joint_out_bias_to_fp16 = const()[name = string("joint_out_bias_to_fp16"), val = tensor<fp16, [8198]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12260672)))];371            tensor<fp16, [4, 8198, 1, 8]> logits = conv(bias = joint_out_bias_to_fp16, dilations = x_dilations_0, groups = x_groups_0, pad = x_pad_0, pad_type = x_pad_type_0, strides = x_strides_0, weight = joint_out_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("x_cast_fp16")];372        } -> (logits, h_out, c_out);373    func decoder_8<ios18>(tensor<fp16, [8, 1280, 1, 1]> c_in, tensor<fp16, [8, 640, 1, 1]> embed, tensor<fp16, [8, 640, 1, 8]> enc_step, tensor<fp16, [8, 1280, 1, 1]> h_in) {374            int32 var_14 = const()[name = string("op_14"), val = int32(1)];375            tensor<int32, [4]> input_1_begin_0 = const()[name = string("input_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];376            tensor<int32, [4]> input_1_end_0 = const()[name = string("input_1_end_0"), val = tensor<int32, [4]>([8, 640, 1, 1])];377            tensor<bool, [4]> input_1_end_mask_0 = const()[name = string("input_1_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];378            tensor<fp16, [8, 640, 1, 1]> input_1_cast_fp16 = slice_by_index(begin = input_1_begin_0, end = input_1_end_0, end_mask = input_1_end_mask_0, x = h_in)[name = string("input_1_cast_fp16")];379            tensor<int32, [4]> c_1_begin_0 = const()[name = string("c_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];380            tensor<int32, [4]> c_1_end_0 = const()[name = string("c_1_end_0"), val = tensor<int32, [4]>([8, 640, 1, 1])];381            tensor<bool, [4]> c_1_end_mask_0 = const()[name = string("c_1_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];382            tensor<fp16, [8, 640, 1, 1]> c_1_cast_fp16 = slice_by_index(begin = c_1_begin_0, end = c_1_end_0, end_mask = c_1_end_mask_0, x = c_in)[name = string("c_1_cast_fp16")];383            string var_34_pad_type_0 = const()[name = string("op_34_pad_type_0"), val = string("valid")];384            tensor<int32, [2]> var_34_strides_0 = const()[name = string("op_34_strides_0"), val = tensor<int32, [2]>([1, 1])];385            tensor<int32, [4]> var_34_pad_0 = const()[name = string("op_34_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];386            tensor<int32, [2]> var_34_dilations_0 = const()[name = string("op_34_dilations_0"), val = tensor<int32, [2]>([1, 1])];387            int32 var_34_groups_0 = const()[name = string("op_34_groups_0"), val = int32(1)];388            tensor<fp16, [2560, 640, 1, 1]> decoder_cells_0_ih_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2560, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [2560, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1638528))))[name = string("decoder_cells_0_ih_weight_to_fp16_quantized")];389            tensor<fp16, [2560]> decoder_cells_0_ih_bias_to_fp16 = const()[name = string("decoder_cells_0_ih_bias_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1643712)))];390            tensor<fp16, [8, 2560, 1, 1]> var_34_cast_fp16 = conv(bias = decoder_cells_0_ih_bias_to_fp16, dilations = var_34_dilations_0, groups = var_34_groups_0, pad = var_34_pad_0, pad_type = var_34_pad_type_0, strides = var_34_strides_0, weight = decoder_cells_0_ih_weight_to_fp16_quantized, x = embed)[name = string("op_34_cast_fp16")];391            string var_40_pad_type_0 = const()[name = string("op_40_pad_type_0"), val = string("valid")];392            tensor<int32, [2]> var_40_strides_0 = const()[name = string("op_40_strides_0"), val = tensor<int32, [2]>([1, 1])];393            tensor<int32, [4]> var_40_pad_0 = const()[name = string("op_40_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];394            tensor<int32, [2]> var_40_dilations_0 = const()[name = string("op_40_dilations_0"), val = tensor<int32, [2]>([1, 1])];395            int32 var_40_groups_0 = const()[name = string("op_40_groups_0"), val = int32(1)];396            tensor<fp16, [2560, 640, 1, 1]> decoder_cells_0_hh_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2560, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1648896))), scale = tensor<fp16, [2560, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3287360))))[name = string("decoder_cells_0_hh_weight_to_fp16_quantized")];397            tensor<fp16, [8, 2560, 1, 1]> var_40_cast_fp16 = conv(dilations = var_40_dilations_0, groups = var_40_groups_0, pad = var_40_pad_0, pad_type = var_40_pad_type_0, strides = var_40_strides_0, weight = decoder_cells_0_hh_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("op_40_cast_fp16")];398            tensor<fp16, [8, 2560, 1, 1]> gates_1_cast_fp16 = add(x = var_34_cast_fp16, y = var_40_cast_fp16)[name = string("gates_1_cast_fp16")];399            tensor<int32, [4]> var_43_begin_0 = const()[name = string("op_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];400            tensor<int32, [4]> var_43_end_0 = const()[name = string("op_43_end_0"), val = tensor<int32, [4]>([8, 640, 1, 1])];401            tensor<bool, [4]> var_43_end_mask_0 = const()[name = string("op_43_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];402            tensor<fp16, [8, 640, 1, 1]> var_43_cast_fp16 = slice_by_index(begin = var_43_begin_0, end = var_43_end_0, end_mask = var_43_end_mask_0, x = gates_1_cast_fp16)[name = string("op_43_cast_fp16")];403            tensor<fp16, [8, 640, 1, 1]> i_1_cast_fp16 = sigmoid(x = var_43_cast_fp16)[name = string("i_1_cast_fp16")];404            tensor<int32, [4]> var_46_begin_0 = const()[name = string("op_46_begin_0"), val = tensor<int32, [4]>([0, 640, 0, 0])];405            tensor<int32, [4]> var_46_end_0 = const()[name = string("op_46_end_0"), val = tensor<int32, [4]>([8, 1280, 1, 1])];406            tensor<bool, [4]> var_46_end_mask_0 = const()[name = string("op_46_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];407            tensor<fp16, [8, 640, 1, 1]> var_46_cast_fp16 = slice_by_index(begin = var_46_begin_0, end = var_46_end_0, end_mask = var_46_end_mask_0, x = gates_1_cast_fp16)[name = string("op_46_cast_fp16")];408            tensor<fp16, [8, 640, 1, 1]> f_1_cast_fp16 = sigmoid(x = var_46_cast_fp16)[name = string("f_1_cast_fp16")];409            tensor<int32, [4]> var_49_begin_0 = const()[name = string("op_49_begin_0"), val = tensor<int32, [4]>([0, 1280, 0, 0])];410            tensor<int32, [4]> var_49_end_0 = const()[name = string("op_49_end_0"), val = tensor<int32, [4]>([8, 1920, 1, 1])];411            tensor<bool, [4]> var_49_end_mask_0 = const()[name = string("op_49_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];412            tensor<fp16, [8, 640, 1, 1]> var_49_cast_fp16 = slice_by_index(begin = var_49_begin_0, end = var_49_end_0, end_mask = var_49_end_mask_0, x = gates_1_cast_fp16)[name = string("op_49_cast_fp16")];413            tensor<fp16, [8, 640, 1, 1]> g_1_cast_fp16 = tanh(x = var_49_cast_fp16)[name = string("g_1_cast_fp16")];414            tensor<int32, [4]> var_52_begin_0 = const()[name = string("op_52_begin_0"), val = tensor<int32, [4]>([0, 1920, 0, 0])];415            tensor<int32, [4]> var_52_end_0 = const()[name = string("op_52_end_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];416            tensor<bool, [4]> var_52_end_mask_0 = const()[name = string("op_52_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];417            tensor<fp16, [8, 640, 1, 1]> var_52_cast_fp16 = slice_by_index(begin = var_52_begin_0, end = var_52_end_0, end_mask = var_52_end_mask_0, x = gates_1_cast_fp16)[name = string("op_52_cast_fp16")];418            tensor<fp16, [8, 640, 1, 1]> o_1_cast_fp16 = sigmoid(x = var_52_cast_fp16)[name = string("o_1_cast_fp16")];419            tensor<fp16, [8, 640, 1, 1]> var_54_cast_fp16 = mul(x = f_1_cast_fp16, y = c_1_cast_fp16)[name = string("op_54_cast_fp16")];420            tensor<fp16, [8, 640, 1, 1]> var_55_cast_fp16 = mul(x = i_1_cast_fp16, y = g_1_cast_fp16)[name = string("op_55_cast_fp16")];421            tensor<fp16, [8, 640, 1, 1]> c_new_1_cast_fp16 = add(x = var_54_cast_fp16, y = var_55_cast_fp16)[name = string("c_new_1_cast_fp16")];422            tensor<fp16, [8, 640, 1, 1]> var_57_cast_fp16 = tanh(x = c_new_1_cast_fp16)[name = string("op_57_cast_fp16")];423            tensor<fp16, [8, 640, 1, 1]> input_3_cast_fp16 = mul(x = o_1_cast_fp16, y = var_57_cast_fp16)[name = string("input_3_cast_fp16")];424            tensor<int32, [4]> input_5_begin_0 = const()[name = string("input_5_begin_0"), val = tensor<int32, [4]>([0, 640, 0, 0])];425            tensor<int32, [4]> input_5_end_0 = const()[name = string("input_5_end_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];426            tensor<bool, [4]> input_5_end_mask_0 = const()[name = string("input_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];427            tensor<fp16, [8, 640, 1, 1]> input_5_cast_fp16 = slice_by_index(begin = input_5_begin_0, end = input_5_end_0, end_mask = input_5_end_mask_0, x = h_in)[name = string("input_5_cast_fp16")];428            tensor<int32, [4]> c_begin_0 = const()[name = string("c_begin_0"), val = tensor<int32, [4]>([0, 640, 0, 0])];429            tensor<int32, [4]> c_end_0 = const()[name = string("c_end_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];430            tensor<bool, [4]> c_end_mask_0 = const()[name = string("c_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];431            tensor<fp16, [8, 640, 1, 1]> c_cast_fp16 = slice_by_index(begin = c_begin_0, end = c_end_0, end_mask = c_end_mask_0, x = c_in)[name = string("c_cast_fp16")];432            string var_74_pad_type_0 = const()[name = string("op_74_pad_type_0"), val = string("valid")];433            tensor<int32, [2]> var_74_strides_0 = const()[name = string("op_74_strides_0"), val = tensor<int32, [2]>([1, 1])];434            tensor<int32, [4]> var_74_pad_0 = const()[name = string("op_74_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];435            tensor<int32, [2]> var_74_dilations_0 = const()[name = string("op_74_dilations_0"), val = tensor<int32, [2]>([1, 1])];436            int32 var_74_groups_0 = const()[name = string("op_74_groups_0"), val = int32(1)];437            tensor<fp16, [2560, 640, 1, 1]> decoder_cells_1_ih_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2560, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3292544))), scale = tensor<fp16, [2560, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4931008))))[name = string("decoder_cells_1_ih_weight_to_fp16_quantized")];438            tensor<fp16, [2560]> decoder_cells_1_ih_bias_to_fp16 = const()[name = string("decoder_cells_1_ih_bias_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4936192)))];439            tensor<fp16, [8, 2560, 1, 1]> var_74_cast_fp16 = conv(bias = decoder_cells_1_ih_bias_to_fp16, dilations = var_74_dilations_0, groups = var_74_groups_0, pad = var_74_pad_0, pad_type = var_74_pad_type_0, strides = var_74_strides_0, weight = decoder_cells_1_ih_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("op_74_cast_fp16")];440            string var_80_pad_type_0 = const()[name = string("op_80_pad_type_0"), val = string("valid")];441            tensor<int32, [2]> var_80_strides_0 = const()[name = string("op_80_strides_0"), val = tensor<int32, [2]>([1, 1])];442            tensor<int32, [4]> var_80_pad_0 = const()[name = string("op_80_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];443            tensor<int32, [2]> var_80_dilations_0 = const()[name = string("op_80_dilations_0"), val = tensor<int32, [2]>([1, 1])];444            int32 var_80_groups_0 = const()[name = string("op_80_groups_0"), val = int32(1)];445            tensor<fp16, [2560, 640, 1, 1]> decoder_cells_1_hh_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2560, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4941376))), scale = tensor<fp16, [2560, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6579840))))[name = string("decoder_cells_1_hh_weight_to_fp16_quantized")];446            tensor<fp16, [8, 2560, 1, 1]> var_80_cast_fp16 = conv(dilations = var_80_dilations_0, groups = var_80_groups_0, pad = var_80_pad_0, pad_type = var_80_pad_type_0, strides = var_80_strides_0, weight = decoder_cells_1_hh_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("op_80_cast_fp16")];447            tensor<fp16, [8, 2560, 1, 1]> gates_cast_fp16 = add(x = var_74_cast_fp16, y = var_80_cast_fp16)[name = string("gates_cast_fp16")];448            tensor<int32, [4]> var_83_begin_0 = const()[name = string("op_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];449            tensor<int32, [4]> var_83_end_0 = const()[name = string("op_83_end_0"), val = tensor<int32, [4]>([8, 640, 1, 1])];450            tensor<bool, [4]> var_83_end_mask_0 = const()[name = string("op_83_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];451            tensor<fp16, [8, 640, 1, 1]> var_83_cast_fp16 = slice_by_index(begin = var_83_begin_0, end = var_83_end_0, end_mask = var_83_end_mask_0, x = gates_cast_fp16)[name = string("op_83_cast_fp16")];452            tensor<fp16, [8, 640, 1, 1]> i_cast_fp16 = sigmoid(x = var_83_cast_fp16)[name = string("i_cast_fp16")];453            tensor<int32, [4]> var_86_begin_0 = const()[name = string("op_86_begin_0"), val = tensor<int32, [4]>([0, 640, 0, 0])];454            tensor<int32, [4]> var_86_end_0 = const()[name = string("op_86_end_0"), val = tensor<int32, [4]>([8, 1280, 1, 1])];455            tensor<bool, [4]> var_86_end_mask_0 = const()[name = string("op_86_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];456            tensor<fp16, [8, 640, 1, 1]> var_86_cast_fp16 = slice_by_index(begin = var_86_begin_0, end = var_86_end_0, end_mask = var_86_end_mask_0, x = gates_cast_fp16)[name = string("op_86_cast_fp16")];457            tensor<fp16, [8, 640, 1, 1]> f_cast_fp16 = sigmoid(x = var_86_cast_fp16)[name = string("f_cast_fp16")];458            tensor<int32, [4]> var_89_begin_0 = const()[name = string("op_89_begin_0"), val = tensor<int32, [4]>([0, 1280, 0, 0])];459            tensor<int32, [4]> var_89_end_0 = const()[name = string("op_89_end_0"), val = tensor<int32, [4]>([8, 1920, 1, 1])];460            tensor<bool, [4]> var_89_end_mask_0 = const()[name = string("op_89_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];461            tensor<fp16, [8, 640, 1, 1]> var_89_cast_fp16 = slice_by_index(begin = var_89_begin_0, end = var_89_end_0, end_mask = var_89_end_mask_0, x = gates_cast_fp16)[name = string("op_89_cast_fp16")];462            tensor<fp16, [8, 640, 1, 1]> g_cast_fp16 = tanh(x = var_89_cast_fp16)[name = string("g_cast_fp16")];463            tensor<int32, [4]> var_92_begin_0 = const()[name = string("op_92_begin_0"), val = tensor<int32, [4]>([0, 1920, 0, 0])];464            tensor<int32, [4]> var_92_end_0 = const()[name = string("op_92_end_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];465            tensor<bool, [4]> var_92_end_mask_0 = const()[name = string("op_92_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];466            tensor<fp16, [8, 640, 1, 1]> var_92_cast_fp16 = slice_by_index(begin = var_92_begin_0, end = var_92_end_0, end_mask = var_92_end_mask_0, x = gates_cast_fp16)[name = string("op_92_cast_fp16")];467            tensor<fp16, [8, 640, 1, 1]> o_cast_fp16 = sigmoid(x = var_92_cast_fp16)[name = string("o_cast_fp16")];468            tensor<fp16, [8, 640, 1, 1]> var_94_cast_fp16 = mul(x = f_cast_fp16, y = c_cast_fp16)[name = string("op_94_cast_fp16")];469            tensor<fp16, [8, 640, 1, 1]> var_95_cast_fp16 = mul(x = i_cast_fp16, y = g_cast_fp16)[name = string("op_95_cast_fp16")];470            tensor<fp16, [8, 640, 1, 1]> c_new_cast_fp16 = add(x = var_94_cast_fp16, y = var_95_cast_fp16)[name = string("c_new_cast_fp16")];471            tensor<fp16, [8, 640, 1, 1]> var_97_cast_fp16 = tanh(x = c_new_cast_fp16)[name = string("op_97_cast_fp16")];472            tensor<fp16, [8, 640, 1, 1]> input_7_cast_fp16 = mul(x = o_cast_fp16, y = var_97_cast_fp16)[name = string("input_7_cast_fp16")];473            string pred_pad_type_0 = const()[name = string("pred_pad_type_0"), val = string("valid")];474            tensor<int32, [2]> pred_strides_0 = const()[name = string("pred_strides_0"), val = tensor<int32, [2]>([1, 1])];475            tensor<int32, [4]> pred_pad_0 = const()[name = string("pred_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];476            tensor<int32, [2]> pred_dilations_0 = const()[name = string("pred_dilations_0"), val = tensor<int32, [2]>([1, 1])];477            int32 pred_groups_0 = const()[name = string("pred_groups_0"), val = int32(1)];478            tensor<fp16, [640, 640, 1, 1]> decoder_joint_pred_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [640, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6585024))), scale = tensor<fp16, [640, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6994688))))[name = string("decoder_joint_pred_weight_to_fp16_quantized")];479            tensor<fp16, [640]> decoder_joint_pred_bias_to_fp16 = const()[name = string("decoder_joint_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6996032)))];480            tensor<fp16, [8, 640, 1, 1]> pred_cast_fp16 = conv(bias = decoder_joint_pred_bias_to_fp16, dilations = pred_dilations_0, groups = pred_groups_0, pad = pred_pad_0, pad_type = pred_pad_type_0, strides = pred_strides_0, weight = decoder_joint_pred_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("pred_cast_fp16")];481            bool var_110_interleave_0 = const()[name = string("op_110_interleave_0"), val = bool(false)];482            tensor<fp16, [8, 1280, 1, 1]> h_out = concat(axis = var_14, interleave = var_110_interleave_0, values = (input_3_cast_fp16, input_7_cast_fp16))[name = string("op_110_cast_fp16")];483            bool var_112_interleave_0 = const()[name = string("op_112_interleave_0"), val = bool(false)];484            tensor<fp16, [8, 1280, 1, 1]> c_out = concat(axis = var_14, interleave = var_112_interleave_0, values = (c_new_1_cast_fp16, c_new_cast_fp16))[name = string("op_112_cast_fp16")];485            tensor<fp16, [8, 640, 1, 8]> var_122_cast_fp16 = add(x = enc_step, y = pred_cast_fp16)[name = string("op_122_cast_fp16")];486            tensor<fp16, [8, 640, 1, 8]> input_cast_fp16 = relu(x = var_122_cast_fp16)[name = string("input_cast_fp16")];487            string x_pad_type_0 = const()[name = string("x_pad_type_0"), val = string("valid")];488            tensor<int32, [2]> x_strides_0 = const()[name = string("x_strides_0"), val = tensor<int32, [2]>([1, 1])];489            tensor<int32, [4]> x_pad_0 = const()[name = string("x_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];490            tensor<int32, [2]> x_dilations_0 = const()[name = string("x_dilations_0"), val = tensor<int32, [2]>([1, 1])];491            int32 x_groups_0 = const()[name = string("x_groups_0"), val = int32(1)];492            tensor<fp16, [8198, 640, 1, 1]> joint_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [8198, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6997376))), scale = tensor<fp16, [8198, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12244160))))[name = string("joint_out_weight_to_fp16_quantized")];493            tensor<fp16, [8198]> joint_out_bias_to_fp16 = const()[name = string("joint_out_bias_to_fp16"), val = tensor<fp16, [8198]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12260672)))];494            tensor<fp16, [8, 8198, 1, 8]> logits = conv(bias = joint_out_bias_to_fp16, dilations = x_dilations_0, groups = x_groups_0, pad = x_pad_0, pad_type = x_pad_type_0, strides = x_strides_0, weight = joint_out_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("x_cast_fp16")];495        } -> (logits, h_out, c_out);496    func main<ios18>(tensor<fp16, [16, 1280, 1, 1]> c_in, tensor<fp16, [16, 640, 1, 1]> embed, tensor<fp16, [16, 640, 1, 8]> enc_step, tensor<fp16, [16, 1280, 1, 1]> h_in) {497            int32 var_14 = const()[name = string("op_14"), val = int32(1)];498            tensor<int32, [4]> input_1_begin_0 = const()[name = string("input_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];499            tensor<int32, [4]> input_1_end_0 = const()[name = string("input_1_end_0"), val = tensor<int32, [4]>([16, 640, 1, 1])];500            tensor<bool, [4]> input_1_end_mask_0 = const()[name = string("input_1_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];501            tensor<fp16, [16, 640, 1, 1]> input_1_cast_fp16 = slice_by_index(begin = input_1_begin_0, end = input_1_end_0, end_mask = input_1_end_mask_0, x = h_in)[name = string("input_1_cast_fp16")];502            tensor<int32, [4]> c_1_begin_0 = const()[name = string("c_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];503            tensor<int32, [4]> c_1_end_0 = const()[name = string("c_1_end_0"), val = tensor<int32, [4]>([16, 640, 1, 1])];504            tensor<bool, [4]> c_1_end_mask_0 = const()[name = string("c_1_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];505            tensor<fp16, [16, 640, 1, 1]> c_1_cast_fp16 = slice_by_index(begin = c_1_begin_0, end = c_1_end_0, end_mask = c_1_end_mask_0, x = c_in)[name = string("c_1_cast_fp16")];506            string var_34_pad_type_0 = const()[name = string("op_34_pad_type_0"), val = string("valid")];507            tensor<int32, [2]> var_34_strides_0 = const()[name = string("op_34_strides_0"), val = tensor<int32, [2]>([1, 1])];508            tensor<int32, [4]> var_34_pad_0 = const()[name = string("op_34_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];509            tensor<int32, [2]> var_34_dilations_0 = const()[name = string("op_34_dilations_0"), val = tensor<int32, [2]>([1, 1])];510            int32 var_34_groups_0 = const()[name = string("op_34_groups_0"), val = int32(1)];511            tensor<fp16, [2560, 640, 1, 1]> decoder_cells_0_ih_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2560, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [2560, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1638528))))[name = string("decoder_cells_0_ih_weight_to_fp16_quantized")];512            tensor<fp16, [2560]> decoder_cells_0_ih_bias_to_fp16 = const()[name = string("decoder_cells_0_ih_bias_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1643712)))];513            tensor<fp16, [16, 2560, 1, 1]> var_34_cast_fp16 = conv(bias = decoder_cells_0_ih_bias_to_fp16, dilations = var_34_dilations_0, groups = var_34_groups_0, pad = var_34_pad_0, pad_type = var_34_pad_type_0, strides = var_34_strides_0, weight = decoder_cells_0_ih_weight_to_fp16_quantized, x = embed)[name = string("op_34_cast_fp16")];514            string var_40_pad_type_0 = const()[name = string("op_40_pad_type_0"), val = string("valid")];515            tensor<int32, [2]> var_40_strides_0 = const()[name = string("op_40_strides_0"), val = tensor<int32, [2]>([1, 1])];516            tensor<int32, [4]> var_40_pad_0 = const()[name = string("op_40_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];517            tensor<int32, [2]> var_40_dilations_0 = const()[name = string("op_40_dilations_0"), val = tensor<int32, [2]>([1, 1])];518            int32 var_40_groups_0 = const()[name = string("op_40_groups_0"), val = int32(1)];519            tensor<fp16, [2560, 640, 1, 1]> decoder_cells_0_hh_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2560, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1648896))), scale = tensor<fp16, [2560, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3287360))))[name = string("decoder_cells_0_hh_weight_to_fp16_quantized")];520            tensor<fp16, [16, 2560, 1, 1]> var_40_cast_fp16 = conv(dilations = var_40_dilations_0, groups = var_40_groups_0, pad = var_40_pad_0, pad_type = var_40_pad_type_0, strides = var_40_strides_0, weight = decoder_cells_0_hh_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("op_40_cast_fp16")];521            tensor<fp16, [16, 2560, 1, 1]> gates_1_cast_fp16 = add(x = var_34_cast_fp16, y = var_40_cast_fp16)[name = string("gates_1_cast_fp16")];522            tensor<int32, [4]> var_43_begin_0 = const()[name = string("op_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];523            tensor<int32, [4]> var_43_end_0 = const()[name = string("op_43_end_0"), val = tensor<int32, [4]>([16, 640, 1, 1])];524            tensor<bool, [4]> var_43_end_mask_0 = const()[name = string("op_43_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];525            tensor<fp16, [16, 640, 1, 1]> var_43_cast_fp16 = slice_by_index(begin = var_43_begin_0, end = var_43_end_0, end_mask = var_43_end_mask_0, x = gates_1_cast_fp16)[name = string("op_43_cast_fp16")];526            tensor<fp16, [16, 640, 1, 1]> i_1_cast_fp16 = sigmoid(x = var_43_cast_fp16)[name = string("i_1_cast_fp16")];527            tensor<int32, [4]> var_46_begin_0 = const()[name = string("op_46_begin_0"), val = tensor<int32, [4]>([0, 640, 0, 0])];528            tensor<int32, [4]> var_46_end_0 = const()[name = string("op_46_end_0"), val = tensor<int32, [4]>([16, 1280, 1, 1])];529            tensor<bool, [4]> var_46_end_mask_0 = const()[name = string("op_46_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];530            tensor<fp16, [16, 640, 1, 1]> var_46_cast_fp16 = slice_by_index(begin = var_46_begin_0, end = var_46_end_0, end_mask = var_46_end_mask_0, x = gates_1_cast_fp16)[name = string("op_46_cast_fp16")];531            tensor<fp16, [16, 640, 1, 1]> f_1_cast_fp16 = sigmoid(x = var_46_cast_fp16)[name = string("f_1_cast_fp16")];532            tensor<int32, [4]> var_49_begin_0 = const()[name = string("op_49_begin_0"), val = tensor<int32, [4]>([0, 1280, 0, 0])];533            tensor<int32, [4]> var_49_end_0 = const()[name = string("op_49_end_0"), val = tensor<int32, [4]>([16, 1920, 1, 1])];534            tensor<bool, [4]> var_49_end_mask_0 = const()[name = string("op_49_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];535            tensor<fp16, [16, 640, 1, 1]> var_49_cast_fp16 = slice_by_index(begin = var_49_begin_0, end = var_49_end_0, end_mask = var_49_end_mask_0, x = gates_1_cast_fp16)[name = string("op_49_cast_fp16")];536            tensor<fp16, [16, 640, 1, 1]> g_1_cast_fp16 = tanh(x = var_49_cast_fp16)[name = string("g_1_cast_fp16")];537            tensor<int32, [4]> var_52_begin_0 = const()[name = string("op_52_begin_0"), val = tensor<int32, [4]>([0, 1920, 0, 0])];538            tensor<int32, [4]> var_52_end_0 = const()[name = string("op_52_end_0"), val = tensor<int32, [4]>([16, 1, 1, 1])];539            tensor<bool, [4]> var_52_end_mask_0 = const()[name = string("op_52_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];540            tensor<fp16, [16, 640, 1, 1]> var_52_cast_fp16 = slice_by_index(begin = var_52_begin_0, end = var_52_end_0, end_mask = var_52_end_mask_0, x = gates_1_cast_fp16)[name = string("op_52_cast_fp16")];541            tensor<fp16, [16, 640, 1, 1]> o_1_cast_fp16 = sigmoid(x = var_52_cast_fp16)[name = string("o_1_cast_fp16")];542            tensor<fp16, [16, 640, 1, 1]> var_54_cast_fp16 = mul(x = f_1_cast_fp16, y = c_1_cast_fp16)[name = string("op_54_cast_fp16")];543            tensor<fp16, [16, 640, 1, 1]> var_55_cast_fp16 = mul(x = i_1_cast_fp16, y = g_1_cast_fp16)[name = string("op_55_cast_fp16")];544            tensor<fp16, [16, 640, 1, 1]> c_new_1_cast_fp16 = add(x = var_54_cast_fp16, y = var_55_cast_fp16)[name = string("c_new_1_cast_fp16")];545            tensor<fp16, [16, 640, 1, 1]> var_57_cast_fp16 = tanh(x = c_new_1_cast_fp16)[name = string("op_57_cast_fp16")];546            tensor<fp16, [16, 640, 1, 1]> input_3_cast_fp16 = mul(x = o_1_cast_fp16, y = var_57_cast_fp16)[name = string("input_3_cast_fp16")];547            tensor<int32, [4]> input_5_begin_0 = const()[name = string("input_5_begin_0"), val = tensor<int32, [4]>([0, 640, 0, 0])];548            tensor<int32, [4]> input_5_end_0 = const()[name = string("input_5_end_0"), val = tensor<int32, [4]>([16, 1, 1, 1])];549            tensor<bool, [4]> input_5_end_mask_0 = const()[name = string("input_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];550            tensor<fp16, [16, 640, 1, 1]> input_5_cast_fp16 = slice_by_index(begin = input_5_begin_0, end = input_5_end_0, end_mask = input_5_end_mask_0, x = h_in)[name = string("input_5_cast_fp16")];551            tensor<int32, [4]> c_begin_0 = const()[name = string("c_begin_0"), val = tensor<int32, [4]>([0, 640, 0, 0])];552            tensor<int32, [4]> c_end_0 = const()[name = string("c_end_0"), val = tensor<int32, [4]>([16, 1, 1, 1])];553            tensor<bool, [4]> c_end_mask_0 = const()[name = string("c_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];554            tensor<fp16, [16, 640, 1, 1]> c_cast_fp16 = slice_by_index(begin = c_begin_0, end = c_end_0, end_mask = c_end_mask_0, x = c_in)[name = string("c_cast_fp16")];555            string var_74_pad_type_0 = const()[name = string("op_74_pad_type_0"), val = string("valid")];556            tensor<int32, [2]> var_74_strides_0 = const()[name = string("op_74_strides_0"), val = tensor<int32, [2]>([1, 1])];557            tensor<int32, [4]> var_74_pad_0 = const()[name = string("op_74_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];558            tensor<int32, [2]> var_74_dilations_0 = const()[name = string("op_74_dilations_0"), val = tensor<int32, [2]>([1, 1])];559            int32 var_74_groups_0 = const()[name = string("op_74_groups_0"), val = int32(1)];560            tensor<fp16, [2560, 640, 1, 1]> decoder_cells_1_ih_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2560, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3292544))), scale = tensor<fp16, [2560, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4931008))))[name = string("decoder_cells_1_ih_weight_to_fp16_quantized")];561            tensor<fp16, [2560]> decoder_cells_1_ih_bias_to_fp16 = const()[name = string("decoder_cells_1_ih_bias_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4936192)))];562            tensor<fp16, [16, 2560, 1, 1]> var_74_cast_fp16 = conv(bias = decoder_cells_1_ih_bias_to_fp16, dilations = var_74_dilations_0, groups = var_74_groups_0, pad = var_74_pad_0, pad_type = var_74_pad_type_0, strides = var_74_strides_0, weight = decoder_cells_1_ih_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("op_74_cast_fp16")];563            string var_80_pad_type_0 = const()[name = string("op_80_pad_type_0"), val = string("valid")];564            tensor<int32, [2]> var_80_strides_0 = const()[name = string("op_80_strides_0"), val = tensor<int32, [2]>([1, 1])];565            tensor<int32, [4]> var_80_pad_0 = const()[name = string("op_80_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];566            tensor<int32, [2]> var_80_dilations_0 = const()[name = string("op_80_dilations_0"), val = tensor<int32, [2]>([1, 1])];567            int32 var_80_groups_0 = const()[name = string("op_80_groups_0"), val = int32(1)];568            tensor<fp16, [2560, 640, 1, 1]> decoder_cells_1_hh_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2560, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4941376))), scale = tensor<fp16, [2560, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6579840))))[name = string("decoder_cells_1_hh_weight_to_fp16_quantized")];569            tensor<fp16, [16, 2560, 1, 1]> var_80_cast_fp16 = conv(dilations = var_80_dilations_0, groups = var_80_groups_0, pad = var_80_pad_0, pad_type = var_80_pad_type_0, strides = var_80_strides_0, weight = decoder_cells_1_hh_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("op_80_cast_fp16")];570            tensor<fp16, [16, 2560, 1, 1]> gates_cast_fp16 = add(x = var_74_cast_fp16, y = var_80_cast_fp16)[name = string("gates_cast_fp16")];571            tensor<int32, [4]> var_83_begin_0 = const()[name = string("op_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];572            tensor<int32, [4]> var_83_end_0 = const()[name = string("op_83_end_0"), val = tensor<int32, [4]>([16, 640, 1, 1])];573            tensor<bool, [4]> var_83_end_mask_0 = const()[name = string("op_83_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];574            tensor<fp16, [16, 640, 1, 1]> var_83_cast_fp16 = slice_by_index(begin = var_83_begin_0, end = var_83_end_0, end_mask = var_83_end_mask_0, x = gates_cast_fp16)[name = string("op_83_cast_fp16")];575            tensor<fp16, [16, 640, 1, 1]> i_cast_fp16 = sigmoid(x = var_83_cast_fp16)[name = string("i_cast_fp16")];576            tensor<int32, [4]> var_86_begin_0 = const()[name = string("op_86_begin_0"), val = tensor<int32, [4]>([0, 640, 0, 0])];577            tensor<int32, [4]> var_86_end_0 = const()[name = string("op_86_end_0"), val = tensor<int32, [4]>([16, 1280, 1, 1])];578            tensor<bool, [4]> var_86_end_mask_0 = const()[name = string("op_86_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];579            tensor<fp16, [16, 640, 1, 1]> var_86_cast_fp16 = slice_by_index(begin = var_86_begin_0, end = var_86_end_0, end_mask = var_86_end_mask_0, x = gates_cast_fp16)[name = string("op_86_cast_fp16")];580            tensor<fp16, [16, 640, 1, 1]> f_cast_fp16 = sigmoid(x = var_86_cast_fp16)[name = string("f_cast_fp16")];581            tensor<int32, [4]> var_89_begin_0 = const()[name = string("op_89_begin_0"), val = tensor<int32, [4]>([0, 1280, 0, 0])];582            tensor<int32, [4]> var_89_end_0 = const()[name = string("op_89_end_0"), val = tensor<int32, [4]>([16, 1920, 1, 1])];583            tensor<bool, [4]> var_89_end_mask_0 = const()[name = string("op_89_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];584            tensor<fp16, [16, 640, 1, 1]> var_89_cast_fp16 = slice_by_index(begin = var_89_begin_0, end = var_89_end_0, end_mask = var_89_end_mask_0, x = gates_cast_fp16)[name = string("op_89_cast_fp16")];585            tensor<fp16, [16, 640, 1, 1]> g_cast_fp16 = tanh(x = var_89_cast_fp16)[name = string("g_cast_fp16")];586            tensor<int32, [4]> var_92_begin_0 = const()[name = string("op_92_begin_0"), val = tensor<int32, [4]>([0, 1920, 0, 0])];587            tensor<int32, [4]> var_92_end_0 = const()[name = string("op_92_end_0"), val = tensor<int32, [4]>([16, 1, 1, 1])];588            tensor<bool, [4]> var_92_end_mask_0 = const()[name = string("op_92_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];589            tensor<fp16, [16, 640, 1, 1]> var_92_cast_fp16 = slice_by_index(begin = var_92_begin_0, end = var_92_end_0, end_mask = var_92_end_mask_0, x = gates_cast_fp16)[name = string("op_92_cast_fp16")];590            tensor<fp16, [16, 640, 1, 1]> o_cast_fp16 = sigmoid(x = var_92_cast_fp16)[name = string("o_cast_fp16")];591            tensor<fp16, [16, 640, 1, 1]> var_94_cast_fp16 = mul(x = f_cast_fp16, y = c_cast_fp16)[name = string("op_94_cast_fp16")];592            tensor<fp16, [16, 640, 1, 1]> var_95_cast_fp16 = mul(x = i_cast_fp16, y = g_cast_fp16)[name = string("op_95_cast_fp16")];593            tensor<fp16, [16, 640, 1, 1]> c_new_cast_fp16 = add(x = var_94_cast_fp16, y = var_95_cast_fp16)[name = string("c_new_cast_fp16")];594            tensor<fp16, [16, 640, 1, 1]> var_97_cast_fp16 = tanh(x = c_new_cast_fp16)[name = string("op_97_cast_fp16")];595            tensor<fp16, [16, 640, 1, 1]> input_7_cast_fp16 = mul(x = o_cast_fp16, y = var_97_cast_fp16)[name = string("input_7_cast_fp16")];596            string pred_pad_type_0 = const()[name = string("pred_pad_type_0"), val = string("valid")];597            tensor<int32, [2]> pred_strides_0 = const()[name = string("pred_strides_0"), val = tensor<int32, [2]>([1, 1])];598            tensor<int32, [4]> pred_pad_0 = const()[name = string("pred_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];599            tensor<int32, [2]> pred_dilations_0 = const()[name = string("pred_dilations_0"), val = tensor<int32, [2]>([1, 1])];600            int32 pred_groups_0 = const()[name = string("pred_groups_0"), val = int32(1)];601            tensor<fp16, [640, 640, 1, 1]> decoder_joint_pred_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [640, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6585024))), scale = tensor<fp16, [640, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6994688))))[name = string("decoder_joint_pred_weight_to_fp16_quantized")];602            tensor<fp16, [640]> decoder_joint_pred_bias_to_fp16 = const()[name = string("decoder_joint_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6996032)))];603            tensor<fp16, [16, 640, 1, 1]> pred_cast_fp16 = conv(bias = decoder_joint_pred_bias_to_fp16, dilations = pred_dilations_0, groups = pred_groups_0, pad = pred_pad_0, pad_type = pred_pad_type_0, strides = pred_strides_0, weight = decoder_joint_pred_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("pred_cast_fp16")];604            bool var_110_interleave_0 = const()[name = string("op_110_interleave_0"), val = bool(false)];605            tensor<fp16, [16, 1280, 1, 1]> h_out = concat(axis = var_14, interleave = var_110_interleave_0, values = (input_3_cast_fp16, input_7_cast_fp16))[name = string("op_110_cast_fp16")];606            bool var_112_interleave_0 = const()[name = string("op_112_interleave_0"), val = bool(false)];607            tensor<fp16, [16, 1280, 1, 1]> c_out = concat(axis = var_14, interleave = var_112_interleave_0, values = (c_new_1_cast_fp16, c_new_cast_fp16))[name = string("op_112_cast_fp16")];608            tensor<fp16, [16, 640, 1, 8]> var_122_cast_fp16 = add(x = enc_step, y = pred_cast_fp16)[name = string("op_122_cast_fp16")];609            tensor<fp16, [16, 640, 1, 8]> input_cast_fp16 = relu(x = var_122_cast_fp16)[name = string("input_cast_fp16")];610            string x_pad_type_0 = const()[name = string("x_pad_type_0"), val = string("valid")];611            tensor<int32, [2]> x_strides_0 = const()[name = string("x_strides_0"), val = tensor<int32, [2]>([1, 1])];612            tensor<int32, [4]> x_pad_0 = const()[name = string("x_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];613            tensor<int32, [2]> x_dilations_0 = const()[name = string("x_dilations_0"), val = tensor<int32, [2]>([1, 1])];614            int32 x_groups_0 = const()[name = string("x_groups_0"), val = int32(1)];615            tensor<fp16, [8198, 640, 1, 1]> joint_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [8198, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6997376))), scale = tensor<fp16, [8198, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12244160))))[name = string("joint_out_weight_to_fp16_quantized")];616            tensor<fp16, [8198]> joint_out_bias_to_fp16 = const()[name = string("joint_out_bias_to_fp16"), val = tensor<fp16, [8198]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12260672)))];617            tensor<fp16, [16, 8198, 1, 8]> logits = conv(bias = joint_out_bias_to_fp16, dilations = x_dilations_0, groups = x_groups_0, pad = x_pad_0, pad_type = x_pad_type_0, strides = x_strides_0, weight = joint_out_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("x_cast_fp16")];618        } -> (logits, h_out, c_out);619}