| program(1.3) | |
| [buildInfo = dict<string, string>({{"coremlc-component-MIL", "3405.2.1"}, {"coremlc-version", "3405.2.1"}})] | |
| { | |
| func main<ios18>(tensor<fp16, [?, 64, 1, 4]> x) [FlexibleShapeInformation = tuple<tuple<string, dict<string, tensor<int32, [?]>>>, tuple<string, dict<string, dict<string, tensor<int32, [?]>>>>>((("DefaultShapes", {{"x", [12, 64, 1, 4]}}), ("EnumeratedShapes", {{"2e922f8e", {{"x", [24, 64, 1, 4]}}}, {"45c08af7", {{"x", [30, 64, 1, 4]}}}, {"66582a1a", {{"x", [10, 64, 1, 4]}}}, {"77b3f94e", {{"x", [20, 64, 1, 4]}}}, {"7f4b84c7", {{"x", [26, 64, 1, 4]}}}, {"7fe2a293", {{"x", [16, 64, 1, 4]}}}, {"8523ce80", {{"x", [22, 64, 1, 4]}}}, {"92b5cb49", {{"x", [8, 64, 1, 4]}}}, {"a76a4363", {{"x", [32, 64, 1, 4]}}}, {"b23494e7", {{"x", [14, 64, 1, 4]}}}, {"ba7abd9b", {{"x", [18, 64, 1, 4]}}}, {"cb00930d", {{"x", [28, 64, 1, 4]}}}, {"df778eb5", {{"x", [12, 64, 1, 4]}}}, {"e3e45208", {{"x", [1, 64, 1, 4]}}}})))] { | |
| string x_3_pad_type_0 = const()[name = string("x_3_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> x_3_strides_0 = const()[name = string("x_3_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> x_3_pad_0 = const()[name = string("x_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> x_3_dilations_0 = const()[name = string("x_3_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 x_3_groups_0 = const()[name = string("x_3_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 64, 1, 1]> var_75_to_fp16 = const()[name = string("op_75_to_fp16"), val = tensor<fp16, [1024, 64, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; | |
| tensor<fp16, [1024]> layer_in_proj_bias_to_fp16 = const()[name = string("layer_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131200)))]; | |
| tensor<fp16, [?, 1024, 1, 4]> x_3_cast_fp16 = conv(bias = layer_in_proj_bias_to_fp16, dilations = x_3_dilations_0, groups = x_3_groups_0, pad = x_3_pad_0, pad_type = x_3_pad_type_0, strides = x_3_strides_0, weight = var_75_to_fp16, x = x)[name = string("x_3_cast_fp16")]; | |
| fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)]; | |
| tensor<fp16, [?, 1024, 1, 4]> fill_like_0_cast_fp16 = fill_like(ref_tensor = x_3_cast_fp16, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")]; | |
| tensor<int32, [4]> var_92_begin_0 = const()[name = string("op_92_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [4]> var_92_end_0 = const()[name = string("op_92_end_0"), val = tensor<int32, [4]>([0, 1024, 1, 1])]; | |
| tensor<bool, [4]> var_92_end_mask_0 = const()[name = string("op_92_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])]; | |
| tensor<fp16, [?, 1024, 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 = fill_like_0_cast_fp16)[name = string("op_92_cast_fp16")]; | |
| tensor<fp16, [1, 1024, 1, 1]> var_93_to_fp16 = const()[name = string("op_93_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133312)))]; | |
| tensor<fp16, [?, 1024, 1, 1]> special_tokens_cast_fp16 = mul(x = var_92_cast_fp16, y = var_93_to_fp16)[name = string("special_tokens_cast_fp16")]; | |
| int32 var_96 = const()[name = string("op_96"), val = int32(3)]; | |
| bool x_5_interleave_0 = const()[name = string("x_5_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_5_cast_fp16 = concat(axis = var_96, interleave = x_5_interleave_0, values = (special_tokens_cast_fp16, x_3_cast_fp16))[name = string("x_5_cast_fp16")]; | |
| int32 var_108 = const()[name = string("op_108"), val = int32(-2)]; | |
| int32 var_112 = const()[name = string("op_112"), val = int32(1)]; | |
| int32 var_117 = const()[name = string("op_117"), val = int32(2)]; | |
| fp16 const_1_promoted_to_fp16 = const()[name = string("const_1_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_122_cast_fp16 = mul(x = x_5_cast_fp16, y = const_1_promoted_to_fp16)[name = string("op_122_cast_fp16")]; | |
| bool x_7_interleave_0 = const()[name = string("x_7_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2048, 1, 5]> x_7_cast_fp16 = concat(axis = var_112, interleave = x_7_interleave_0, values = (x_5_cast_fp16, var_122_cast_fp16))[name = string("x_7_cast_fp16")]; | |
| tensor<int32, [1]> out_1_axes_0 = const()[name = string("out_1_axes_0"), val = tensor<int32, [1]>([1])]; | |
| fp16 var_132_to_fp16 = const()[name = string("op_132_to_fp16"), val = fp16(0x1.5p-17)]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_132_to_fp16, x = x_7_cast_fp16)[name = string("out_1_cast_fp16")]; | |
| tensor<fp16, [1, 2048, 1, 1]> layer_encoder_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135424)))]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_3_cast_fp16 = mul(x = out_1_cast_fp16, y = layer_encoder_layers_0_input_layernorm_weight_to_fp16)[name = string("out_3_cast_fp16")]; | |
| tensor<int32, [2]> var_138_split_sizes_0 = const()[name = string("op_138_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; | |
| int32 var_138_axis_0 = const()[name = string("op_138_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_138_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_138_cast_fp16_1 = split(axis = var_138_axis_0, split_sizes = var_138_split_sizes_0, x = out_3_cast_fp16)[name = string("op_138_cast_fp16")]; | |
| string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 1024, 1, 1]> var_103_to_fp16 = const()[name = string("op_103_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139584)))]; | |
| tensor<fp16, [?, 1024, 1, 5]> query_states_1_cast_fp16 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = var_103_to_fp16, x = var_138_cast_fp16_0)[name = string("query_states_1_cast_fp16")]; | |
| string key_states_1_pad_type_0 = const()[name = string("key_states_1_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> key_states_1_strides_0 = const()[name = string("key_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> key_states_1_pad_0 = const()[name = string("key_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> key_states_1_dilations_0 = const()[name = string("key_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 key_states_1_groups_0 = const()[name = string("key_states_1_groups_0"), val = int32(1)]; | |
| tensor<fp16, [128, 1024, 1, 1]> var_104_to_fp16 = const()[name = string("op_104_to_fp16"), val = tensor<fp16, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2236800)))]; | |
| tensor<fp16, [?, 128, 1, 5]> key_states_1_cast_fp16 = conv(dilations = key_states_1_dilations_0, groups = key_states_1_groups_0, pad = key_states_1_pad_0, pad_type = key_states_1_pad_type_0, strides = key_states_1_strides_0, weight = var_104_to_fp16, x = var_138_cast_fp16_0)[name = string("key_states_1_cast_fp16")]; | |
| string value_states_1_pad_type_0 = const()[name = string("value_states_1_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> value_states_1_strides_0 = const()[name = string("value_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> value_states_1_pad_0 = const()[name = string("value_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> value_states_1_dilations_0 = const()[name = string("value_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 value_states_1_groups_0 = const()[name = string("value_states_1_groups_0"), val = int32(1)]; | |
| tensor<fp16, [128, 1024, 1, 1]> var_105_to_fp16 = const()[name = string("op_105_to_fp16"), val = tensor<fp16, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2499008)))]; | |
| tensor<fp16, [?, 128, 1, 5]> value_states_1_cast_fp16 = conv(dilations = value_states_1_dilations_0, groups = value_states_1_groups_0, pad = value_states_1_pad_0, pad_type = value_states_1_pad_type_0, strides = value_states_1_strides_0, weight = var_105_to_fp16, x = var_138_cast_fp16_0)[name = string("value_states_1_cast_fp16")]; | |
| tensor<int32, [4]> concat_0x = const()[name = string("concat_0x"), val = tensor<int32, [4]>([-1, 16, 64, 5])]; | |
| tensor<fp16, [?, 16, 64, 5]> embed_1_cast_fp16 = reshape(shape = concat_0x, x = query_states_1_cast_fp16)[name = string("embed_1_cast_fp16")]; | |
| tensor<int32, [4]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [4]>([-1, 2, 64, 5])]; | |
| tensor<fp16, [?, 2, 64, 5]> embed_3_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("embed_3_cast_fp16")]; | |
| tensor<int32, [4]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [4]>([-1, 2, 64, 5])]; | |
| tensor<fp16, [?, 2, 64, 5]> value_states_3_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("value_states_3_cast_fp16")]; | |
| tensor<fp16, [64, 5]> cos_to_fp16 = const()[name = string("cos_to_fp16"), val = tensor<fp16, [64, 5]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2761216)))]; | |
| tensor<fp16, [?, 16, 64, 5]> var_164_cast_fp16 = mul(x = embed_1_cast_fp16, y = cos_to_fp16)[name = string("op_164_cast_fp16")]; | |
| tensor<int32, [2]> var_165_split_sizes_0 = const()[name = string("op_165_split_sizes_0"), val = tensor<int32, [2]>([32, 32])]; | |
| int32 var_165_axis_0 = const()[name = string("op_165_axis_0"), val = int32(-2)]; | |
| tensor<fp16, [?, 16, 32, 5]> var_165_cast_fp16_0, tensor<fp16, [?, 16, 32, 5]> var_165_cast_fp16_1 = split(axis = var_165_axis_0, split_sizes = var_165_split_sizes_0, x = embed_1_cast_fp16)[name = string("op_165_cast_fp16")]; | |
| fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 16, 32, 5]> var_167_cast_fp16 = mul(x = var_165_cast_fp16_1, y = const_2_promoted_to_fp16)[name = string("op_167_cast_fp16")]; | |
| bool var_169_interleave_0 = const()[name = string("op_169_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 16, 64, 5]> var_169_cast_fp16 = concat(axis = var_108, interleave = var_169_interleave_0, values = (var_167_cast_fp16, var_165_cast_fp16_0))[name = string("op_169_cast_fp16")]; | |
| tensor<fp16, [64, 5]> sin_to_fp16 = const()[name = string("sin_to_fp16"), val = tensor<fp16, [64, 5]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2761920)))]; | |
| tensor<fp16, [?, 16, 64, 5]> var_170_cast_fp16 = mul(x = var_169_cast_fp16, y = sin_to_fp16)[name = string("op_170_cast_fp16")]; | |
| tensor<fp16, [?, 16, 64, 5]> query_states_3_cast_fp16 = add(x = var_164_cast_fp16, y = var_170_cast_fp16)[name = string("query_states_3_cast_fp16")]; | |
| tensor<fp16, [?, 2, 64, 5]> var_172_cast_fp16 = mul(x = embed_3_cast_fp16, y = cos_to_fp16)[name = string("op_172_cast_fp16")]; | |
| tensor<int32, [2]> var_173_split_sizes_0 = const()[name = string("op_173_split_sizes_0"), val = tensor<int32, [2]>([32, 32])]; | |
| int32 var_173_axis_0 = const()[name = string("op_173_axis_0"), val = int32(-2)]; | |
| tensor<fp16, [?, 2, 32, 5]> var_173_cast_fp16_0, tensor<fp16, [?, 2, 32, 5]> var_173_cast_fp16_1 = split(axis = var_173_axis_0, split_sizes = var_173_split_sizes_0, x = embed_3_cast_fp16)[name = string("op_173_cast_fp16")]; | |
| fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 2, 32, 5]> var_175_cast_fp16 = mul(x = var_173_cast_fp16_1, y = const_3_promoted_to_fp16)[name = string("op_175_cast_fp16")]; | |
| bool var_177_interleave_0 = const()[name = string("op_177_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2, 64, 5]> var_177_cast_fp16 = concat(axis = var_108, interleave = var_177_interleave_0, values = (var_175_cast_fp16, var_173_cast_fp16_0))[name = string("op_177_cast_fp16")]; | |
| tensor<fp16, [?, 2, 64, 5]> var_178_cast_fp16 = mul(x = var_177_cast_fp16, y = sin_to_fp16)[name = string("op_178_cast_fp16")]; | |
| tensor<fp16, [?, 2, 64, 5]> key_states_3_cast_fp16 = add(x = var_172_cast_fp16, y = var_178_cast_fp16)[name = string("key_states_3_cast_fp16")]; | |
| tensor<int32, [2]> var_183_split_sizes_0 = const()[name = string("op_183_split_sizes_0"), val = tensor<int32, [2]>([8, 8])]; | |
| int32 var_183_axis_0 = const()[name = string("op_183_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 8, 64, 5]> var_183_cast_fp16_0, tensor<fp16, [?, 8, 64, 5]> var_183_cast_fp16_1 = split(axis = var_183_axis_0, split_sizes = var_183_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_183_cast_fp16")]; | |
| tensor<int32, [2]> var_185_split_sizes_0 = const()[name = string("op_185_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_185_axis_0 = const()[name = string("op_185_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1, 64, 5]> var_185_cast_fp16_0, tensor<fp16, [?, 1, 64, 5]> var_185_cast_fp16_1 = split(axis = var_185_axis_0, split_sizes = var_185_split_sizes_0, x = key_states_3_cast_fp16)[name = string("op_185_cast_fp16")]; | |
| tensor<int32, [2]> var_187_split_sizes_0 = const()[name = string("op_187_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_187_axis_0 = const()[name = string("op_187_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1, 64, 5]> var_187_cast_fp16_0, tensor<fp16, [?, 1, 64, 5]> var_187_cast_fp16_1 = split(axis = var_187_axis_0, split_sizes = var_187_split_sizes_0, x = value_states_3_cast_fp16)[name = string("op_187_cast_fp16")]; | |
| bool attn_weights_1_transpose_x_1 = const()[name = string("attn_weights_1_transpose_x_1"), val = bool(true)]; | |
| bool attn_weights_1_transpose_y_1 = const()[name = string("attn_weights_1_transpose_y_1"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_1_cast_fp16 = matmul(transpose_x = attn_weights_1_transpose_x_1, transpose_y = attn_weights_1_transpose_y_1, x = var_185_cast_fp16_0, y = var_183_cast_fp16_0)[name = string("attn_weights_1_cast_fp16")]; | |
| fp16 _inversed_attn_weights_3_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_3_y_0_to_fp16"), val = fp16(0x1p-3)]; | |
| tensor<fp16, [?, 8, 5, 5]> _inversed_attn_weights_3_cast_fp16 = mul(x = attn_weights_1_cast_fp16, y = _inversed_attn_weights_3_y_0_to_fp16)[name = string("_inversed_attn_weights_3_cast_fp16")]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_5_cast_fp16 = softmax(axis = var_117, x = _inversed_attn_weights_3_cast_fp16)[name = string("attn_weights_5_cast_fp16")]; | |
| bool var_194_transpose_x_0 = const()[name = string("op_194_transpose_x_0"), val = bool(false)]; | |
| bool var_194_transpose_y_0 = const()[name = string("op_194_transpose_y_0"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 64, 5]> var_194_cast_fp16 = matmul(transpose_x = var_194_transpose_x_0, transpose_y = var_194_transpose_y_0, x = var_187_cast_fp16_0, y = attn_weights_5_cast_fp16)[name = string("op_194_cast_fp16")]; | |
| bool attn_weights_7_transpose_x_1 = const()[name = string("attn_weights_7_transpose_x_1"), val = bool(true)]; | |
| bool attn_weights_7_transpose_y_1 = const()[name = string("attn_weights_7_transpose_y_1"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_7_cast_fp16 = matmul(transpose_x = attn_weights_7_transpose_x_1, transpose_y = attn_weights_7_transpose_y_1, x = var_185_cast_fp16_1, y = var_183_cast_fp16_1)[name = string("attn_weights_7_cast_fp16")]; | |
| fp16 _inversed_attn_weights_9_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_9_y_0_to_fp16"), val = fp16(0x1p-3)]; | |
| tensor<fp16, [?, 8, 5, 5]> _inversed_attn_weights_9_cast_fp16 = mul(x = attn_weights_7_cast_fp16, y = _inversed_attn_weights_9_y_0_to_fp16)[name = string("_inversed_attn_weights_9_cast_fp16")]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_11_cast_fp16 = softmax(axis = var_117, x = _inversed_attn_weights_9_cast_fp16)[name = string("attn_weights_11_cast_fp16")]; | |
| bool attn_output_1_transpose_x_0 = const()[name = string("attn_output_1_transpose_x_0"), val = bool(false)]; | |
| bool attn_output_1_transpose_y_0 = const()[name = string("attn_output_1_transpose_y_0"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 64, 5]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = var_187_cast_fp16_1, y = attn_weights_11_cast_fp16)[name = string("attn_output_1_cast_fp16")]; | |
| bool attn_output_3_interleave_0 = const()[name = string("attn_output_3_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 16, 64, 5]> attn_output_3_cast_fp16 = concat(axis = var_112, interleave = attn_output_3_interleave_0, values = (var_194_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")]; | |
| tensor<int32, [4]> concat_3x = const()[name = string("concat_3x"), val = tensor<int32, [4]>([-1, 1024, 1, 5])]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_11_cast_fp16 = reshape(shape = concat_3x, x = attn_output_3_cast_fp16)[name = string("x_11_cast_fp16")]; | |
| string hidden_states_3_pad_type_0 = const()[name = string("hidden_states_3_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> hidden_states_3_strides_0 = const()[name = string("hidden_states_3_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = string("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> hidden_states_3_dilations_0 = const()[name = string("hidden_states_3_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 hidden_states_3_groups_0 = const()[name = string("hidden_states_3_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 1024, 1, 1]> var_111_to_fp16 = const()[name = string("op_111_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2762624)))]; | |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_3_cast_fp16 = conv(dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = var_111_to_fp16, x = x_11_cast_fp16)[name = string("hidden_states_3_cast_fp16")]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_13_cast_fp16 = add(x = x_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = string("x_13_cast_fp16")]; | |
| fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_213_cast_fp16 = mul(x = x_13_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_213_cast_fp16")]; | |
| bool x_15_interleave_0 = const()[name = string("x_15_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2048, 1, 5]> x_15_cast_fp16 = concat(axis = var_112, interleave = x_15_interleave_0, values = (x_13_cast_fp16, var_213_cast_fp16))[name = string("x_15_cast_fp16")]; | |
| tensor<int32, [1]> out_7_axes_0 = const()[name = string("out_7_axes_0"), val = tensor<int32, [1]>([1])]; | |
| fp16 var_223_to_fp16 = const()[name = string("op_223_to_fp16"), val = fp16(0x1.5p-17)]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_223_to_fp16, x = x_15_cast_fp16)[name = string("out_7_cast_fp16")]; | |
| tensor<fp16, [1, 2048, 1, 1]> layer_encoder_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4859840)))]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_9_cast_fp16 = mul(x = out_7_cast_fp16, y = layer_encoder_layers_0_post_attention_layernorm_weight_to_fp16)[name = string("out_9_cast_fp16")]; | |
| tensor<int32, [2]> var_229_split_sizes_0 = const()[name = string("op_229_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; | |
| int32 var_229_axis_0 = const()[name = string("op_229_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_229_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_229_cast_fp16_1 = split(axis = var_229_axis_0, split_sizes = var_229_split_sizes_0, x = out_9_cast_fp16)[name = string("op_229_cast_fp16")]; | |
| string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)]; | |
| tensor<fp16, [4096, 1024, 1, 1]> var_98_to_fp16 = const()[name = string("op_98_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4864000)))]; | |
| tensor<fp16, [?, 4096, 1, 5]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = var_98_to_fp16, x = var_229_cast_fp16_0)[name = string("input_1_cast_fp16")]; | |
| tensor<fp16, [?, 4096, 1, 5]> var_237_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_237_cast_fp16")]; | |
| string var_242_pad_type_0 = const()[name = string("op_242_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> var_242_strides_0 = const()[name = string("op_242_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> var_242_pad_0 = const()[name = string("op_242_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> var_242_dilations_0 = const()[name = string("op_242_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_242_groups_0 = const()[name = string("op_242_groups_0"), val = int32(1)]; | |
| tensor<fp16, [4096, 1024, 1, 1]> var_99_to_fp16 = const()[name = string("op_99_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13252672)))]; | |
| tensor<fp16, [?, 4096, 1, 5]> var_242_cast_fp16 = conv(dilations = var_242_dilations_0, groups = var_242_groups_0, pad = var_242_pad_0, pad_type = var_242_pad_type_0, strides = var_242_strides_0, weight = var_99_to_fp16, x = var_229_cast_fp16_0)[name = string("op_242_cast_fp16")]; | |
| tensor<fp16, [?, 4096, 1, 5]> x_21_cast_fp16 = mul(x = var_237_cast_fp16, y = var_242_cast_fp16)[name = string("x_21_cast_fp16")]; | |
| string hidden_states_5_pad_type_0 = const()[name = string("hidden_states_5_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> hidden_states_5_strides_0 = const()[name = string("hidden_states_5_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> hidden_states_5_pad_0 = const()[name = string("hidden_states_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> hidden_states_5_dilations_0 = const()[name = string("hidden_states_5_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 hidden_states_5_groups_0 = const()[name = string("hidden_states_5_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 4096, 1, 1]> var_100_to_fp16 = const()[name = string("op_100_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21641344)))]; | |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_5_cast_fp16 = conv(dilations = hidden_states_5_dilations_0, groups = hidden_states_5_groups_0, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = hidden_states_5_strides_0, weight = var_100_to_fp16, x = x_21_cast_fp16)[name = string("hidden_states_5_cast_fp16")]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_23_cast_fp16 = add(x = x_13_cast_fp16, y = hidden_states_5_cast_fp16)[name = string("x_23_cast_fp16")]; | |
| int32 var_260 = const()[name = string("op_260"), val = int32(-2)]; | |
| int32 var_264 = const()[name = string("op_264"), val = int32(1)]; | |
| int32 var_269 = const()[name = string("op_269"), val = int32(2)]; | |
| fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_274_cast_fp16 = mul(x = x_23_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_274_cast_fp16")]; | |
| bool x_25_interleave_0 = const()[name = string("x_25_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2048, 1, 5]> x_25_cast_fp16 = concat(axis = var_264, interleave = x_25_interleave_0, values = (x_23_cast_fp16, var_274_cast_fp16))[name = string("x_25_cast_fp16")]; | |
| tensor<int32, [1]> out_13_axes_0 = const()[name = string("out_13_axes_0"), val = tensor<int32, [1]>([1])]; | |
| fp16 var_284_to_fp16 = const()[name = string("op_284_to_fp16"), val = fp16(0x1.5p-17)]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_284_to_fp16, x = x_25_cast_fp16)[name = string("out_13_cast_fp16")]; | |
| tensor<fp16, [1, 2048, 1, 1]> layer_encoder_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30030016)))]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_15_cast_fp16 = mul(x = out_13_cast_fp16, y = layer_encoder_layers_1_input_layernorm_weight_to_fp16)[name = string("out_15_cast_fp16")]; | |
| tensor<int32, [2]> var_290_split_sizes_0 = const()[name = string("op_290_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; | |
| int32 var_290_axis_0 = const()[name = string("op_290_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_290_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_290_cast_fp16_1 = split(axis = var_290_axis_0, split_sizes = var_290_split_sizes_0, x = out_15_cast_fp16)[name = string("op_290_cast_fp16")]; | |
| string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 1024, 1, 1]> var_255_to_fp16 = const()[name = string("op_255_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30034176)))]; | |
| tensor<fp16, [?, 1024, 1, 5]> query_states_7_cast_fp16 = conv(dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = var_255_to_fp16, x = var_290_cast_fp16_0)[name = string("query_states_7_cast_fp16")]; | |
| string key_states_7_pad_type_0 = const()[name = string("key_states_7_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> key_states_7_strides_0 = const()[name = string("key_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> key_states_7_pad_0 = const()[name = string("key_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> key_states_7_dilations_0 = const()[name = string("key_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 key_states_7_groups_0 = const()[name = string("key_states_7_groups_0"), val = int32(1)]; | |
| tensor<fp16, [128, 1024, 1, 1]> var_256_to_fp16 = const()[name = string("op_256_to_fp16"), val = tensor<fp16, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32131392)))]; | |
| tensor<fp16, [?, 128, 1, 5]> key_states_7_cast_fp16 = conv(dilations = key_states_7_dilations_0, groups = key_states_7_groups_0, pad = key_states_7_pad_0, pad_type = key_states_7_pad_type_0, strides = key_states_7_strides_0, weight = var_256_to_fp16, x = var_290_cast_fp16_0)[name = string("key_states_7_cast_fp16")]; | |
| string value_states_7_pad_type_0 = const()[name = string("value_states_7_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> value_states_7_strides_0 = const()[name = string("value_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> value_states_7_pad_0 = const()[name = string("value_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> value_states_7_dilations_0 = const()[name = string("value_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 value_states_7_groups_0 = const()[name = string("value_states_7_groups_0"), val = int32(1)]; | |
| tensor<fp16, [128, 1024, 1, 1]> var_257_to_fp16 = const()[name = string("op_257_to_fp16"), val = tensor<fp16, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32393600)))]; | |
| tensor<fp16, [?, 128, 1, 5]> value_states_7_cast_fp16 = conv(dilations = value_states_7_dilations_0, groups = value_states_7_groups_0, pad = value_states_7_pad_0, pad_type = value_states_7_pad_type_0, strides = value_states_7_strides_0, weight = var_257_to_fp16, x = var_290_cast_fp16_0)[name = string("value_states_7_cast_fp16")]; | |
| tensor<int32, [4]> concat_4x = const()[name = string("concat_4x"), val = tensor<int32, [4]>([-1, 16, 64, 5])]; | |
| tensor<fp16, [?, 16, 64, 5]> embed_5_cast_fp16 = reshape(shape = concat_4x, x = query_states_7_cast_fp16)[name = string("embed_5_cast_fp16")]; | |
| tensor<int32, [4]> concat_5x = const()[name = string("concat_5x"), val = tensor<int32, [4]>([-1, 2, 64, 5])]; | |
| tensor<fp16, [?, 2, 64, 5]> embed_7_cast_fp16 = reshape(shape = concat_5x, x = key_states_7_cast_fp16)[name = string("embed_7_cast_fp16")]; | |
| tensor<int32, [4]> concat_6x = const()[name = string("concat_6x"), val = tensor<int32, [4]>([-1, 2, 64, 5])]; | |
| tensor<fp16, [?, 2, 64, 5]> value_states_9_cast_fp16 = reshape(shape = concat_6x, x = value_states_7_cast_fp16)[name = string("value_states_9_cast_fp16")]; | |
| tensor<fp16, [?, 16, 64, 5]> var_316_cast_fp16 = mul(x = embed_5_cast_fp16, y = cos_to_fp16)[name = string("op_316_cast_fp16")]; | |
| tensor<int32, [2]> var_317_split_sizes_0 = const()[name = string("op_317_split_sizes_0"), val = tensor<int32, [2]>([32, 32])]; | |
| int32 var_317_axis_0 = const()[name = string("op_317_axis_0"), val = int32(-2)]; | |
| tensor<fp16, [?, 16, 32, 5]> var_317_cast_fp16_0, tensor<fp16, [?, 16, 32, 5]> var_317_cast_fp16_1 = split(axis = var_317_axis_0, split_sizes = var_317_split_sizes_0, x = embed_5_cast_fp16)[name = string("op_317_cast_fp16")]; | |
| fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 16, 32, 5]> var_319_cast_fp16 = mul(x = var_317_cast_fp16_1, y = const_6_promoted_to_fp16)[name = string("op_319_cast_fp16")]; | |
| bool var_321_interleave_0 = const()[name = string("op_321_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 16, 64, 5]> var_321_cast_fp16 = concat(axis = var_260, interleave = var_321_interleave_0, values = (var_319_cast_fp16, var_317_cast_fp16_0))[name = string("op_321_cast_fp16")]; | |
| tensor<fp16, [?, 16, 64, 5]> var_322_cast_fp16 = mul(x = var_321_cast_fp16, y = sin_to_fp16)[name = string("op_322_cast_fp16")]; | |
| tensor<fp16, [?, 16, 64, 5]> query_states_9_cast_fp16 = add(x = var_316_cast_fp16, y = var_322_cast_fp16)[name = string("query_states_9_cast_fp16")]; | |
| tensor<fp16, [?, 2, 64, 5]> var_324_cast_fp16 = mul(x = embed_7_cast_fp16, y = cos_to_fp16)[name = string("op_324_cast_fp16")]; | |
| tensor<int32, [2]> var_325_split_sizes_0 = const()[name = string("op_325_split_sizes_0"), val = tensor<int32, [2]>([32, 32])]; | |
| int32 var_325_axis_0 = const()[name = string("op_325_axis_0"), val = int32(-2)]; | |
| tensor<fp16, [?, 2, 32, 5]> var_325_cast_fp16_0, tensor<fp16, [?, 2, 32, 5]> var_325_cast_fp16_1 = split(axis = var_325_axis_0, split_sizes = var_325_split_sizes_0, x = embed_7_cast_fp16)[name = string("op_325_cast_fp16")]; | |
| fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 2, 32, 5]> var_327_cast_fp16 = mul(x = var_325_cast_fp16_1, y = const_7_promoted_to_fp16)[name = string("op_327_cast_fp16")]; | |
| bool var_329_interleave_0 = const()[name = string("op_329_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2, 64, 5]> var_329_cast_fp16 = concat(axis = var_260, interleave = var_329_interleave_0, values = (var_327_cast_fp16, var_325_cast_fp16_0))[name = string("op_329_cast_fp16")]; | |
| tensor<fp16, [?, 2, 64, 5]> var_330_cast_fp16 = mul(x = var_329_cast_fp16, y = sin_to_fp16)[name = string("op_330_cast_fp16")]; | |
| tensor<fp16, [?, 2, 64, 5]> key_states_9_cast_fp16 = add(x = var_324_cast_fp16, y = var_330_cast_fp16)[name = string("key_states_9_cast_fp16")]; | |
| tensor<int32, [2]> var_335_split_sizes_0 = const()[name = string("op_335_split_sizes_0"), val = tensor<int32, [2]>([8, 8])]; | |
| int32 var_335_axis_0 = const()[name = string("op_335_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 8, 64, 5]> var_335_cast_fp16_0, tensor<fp16, [?, 8, 64, 5]> var_335_cast_fp16_1 = split(axis = var_335_axis_0, split_sizes = var_335_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_335_cast_fp16")]; | |
| tensor<int32, [2]> var_337_split_sizes_0 = const()[name = string("op_337_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_337_axis_0 = const()[name = string("op_337_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1, 64, 5]> var_337_cast_fp16_0, tensor<fp16, [?, 1, 64, 5]> var_337_cast_fp16_1 = split(axis = var_337_axis_0, split_sizes = var_337_split_sizes_0, x = key_states_9_cast_fp16)[name = string("op_337_cast_fp16")]; | |
| tensor<int32, [2]> var_339_split_sizes_0 = const()[name = string("op_339_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_339_axis_0 = const()[name = string("op_339_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1, 64, 5]> var_339_cast_fp16_0, tensor<fp16, [?, 1, 64, 5]> var_339_cast_fp16_1 = split(axis = var_339_axis_0, split_sizes = var_339_split_sizes_0, x = value_states_9_cast_fp16)[name = string("op_339_cast_fp16")]; | |
| bool attn_weights_13_transpose_x_1 = const()[name = string("attn_weights_13_transpose_x_1"), val = bool(true)]; | |
| bool attn_weights_13_transpose_y_1 = const()[name = string("attn_weights_13_transpose_y_1"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_13_cast_fp16 = matmul(transpose_x = attn_weights_13_transpose_x_1, transpose_y = attn_weights_13_transpose_y_1, x = var_337_cast_fp16_0, y = var_335_cast_fp16_0)[name = string("attn_weights_13_cast_fp16")]; | |
| fp16 _inversed_attn_weights_15_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_15_y_0_to_fp16"), val = fp16(0x1p-3)]; | |
| tensor<fp16, [?, 8, 5, 5]> _inversed_attn_weights_15_cast_fp16 = mul(x = attn_weights_13_cast_fp16, y = _inversed_attn_weights_15_y_0_to_fp16)[name = string("_inversed_attn_weights_15_cast_fp16")]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_17_cast_fp16 = softmax(axis = var_269, x = _inversed_attn_weights_15_cast_fp16)[name = string("attn_weights_17_cast_fp16")]; | |
| bool var_346_transpose_x_0 = const()[name = string("op_346_transpose_x_0"), val = bool(false)]; | |
| bool var_346_transpose_y_0 = const()[name = string("op_346_transpose_y_0"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 64, 5]> var_346_cast_fp16 = matmul(transpose_x = var_346_transpose_x_0, transpose_y = var_346_transpose_y_0, x = var_339_cast_fp16_0, y = attn_weights_17_cast_fp16)[name = string("op_346_cast_fp16")]; | |
| bool attn_weights_19_transpose_x_1 = const()[name = string("attn_weights_19_transpose_x_1"), val = bool(true)]; | |
| bool attn_weights_19_transpose_y_1 = const()[name = string("attn_weights_19_transpose_y_1"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_19_cast_fp16 = matmul(transpose_x = attn_weights_19_transpose_x_1, transpose_y = attn_weights_19_transpose_y_1, x = var_337_cast_fp16_1, y = var_335_cast_fp16_1)[name = string("attn_weights_19_cast_fp16")]; | |
| fp16 _inversed_attn_weights_21_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_21_y_0_to_fp16"), val = fp16(0x1p-3)]; | |
| tensor<fp16, [?, 8, 5, 5]> _inversed_attn_weights_21_cast_fp16 = mul(x = attn_weights_19_cast_fp16, y = _inversed_attn_weights_21_y_0_to_fp16)[name = string("_inversed_attn_weights_21_cast_fp16")]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_23_cast_fp16 = softmax(axis = var_269, x = _inversed_attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")]; | |
| bool attn_output_5_transpose_x_0 = const()[name = string("attn_output_5_transpose_x_0"), val = bool(false)]; | |
| bool attn_output_5_transpose_y_0 = const()[name = string("attn_output_5_transpose_y_0"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 64, 5]> attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = var_339_cast_fp16_1, y = attn_weights_23_cast_fp16)[name = string("attn_output_5_cast_fp16")]; | |
| bool attn_output_7_interleave_0 = const()[name = string("attn_output_7_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 16, 64, 5]> attn_output_7_cast_fp16 = concat(axis = var_264, interleave = attn_output_7_interleave_0, values = (var_346_cast_fp16, attn_output_5_cast_fp16))[name = string("attn_output_7_cast_fp16")]; | |
| tensor<int32, [4]> concat_7x = const()[name = string("concat_7x"), val = tensor<int32, [4]>([-1, 1024, 1, 5])]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_29_cast_fp16 = reshape(shape = concat_7x, x = attn_output_7_cast_fp16)[name = string("x_29_cast_fp16")]; | |
| string hidden_states_9_pad_type_0 = const()[name = string("hidden_states_9_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> hidden_states_9_strides_0 = const()[name = string("hidden_states_9_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> hidden_states_9_pad_0 = const()[name = string("hidden_states_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> hidden_states_9_dilations_0 = const()[name = string("hidden_states_9_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 hidden_states_9_groups_0 = const()[name = string("hidden_states_9_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 1024, 1, 1]> var_263_to_fp16 = const()[name = string("op_263_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32655808)))]; | |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_9_cast_fp16 = conv(dilations = hidden_states_9_dilations_0, groups = hidden_states_9_groups_0, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = hidden_states_9_strides_0, weight = var_263_to_fp16, x = x_29_cast_fp16)[name = string("hidden_states_9_cast_fp16")]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_31_cast_fp16 = add(x = x_23_cast_fp16, y = hidden_states_9_cast_fp16)[name = string("x_31_cast_fp16")]; | |
| fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_365_cast_fp16 = mul(x = x_31_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_365_cast_fp16")]; | |
| bool x_33_interleave_0 = const()[name = string("x_33_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2048, 1, 5]> x_33_cast_fp16 = concat(axis = var_264, interleave = x_33_interleave_0, values = (x_31_cast_fp16, var_365_cast_fp16))[name = string("x_33_cast_fp16")]; | |
| tensor<int32, [1]> out_19_axes_0 = const()[name = string("out_19_axes_0"), val = tensor<int32, [1]>([1])]; | |
| fp16 var_375_to_fp16 = const()[name = string("op_375_to_fp16"), val = fp16(0x1.5p-17)]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_375_to_fp16, x = x_33_cast_fp16)[name = string("out_19_cast_fp16")]; | |
| tensor<fp16, [1, 2048, 1, 1]> layer_encoder_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34753024)))]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_21_cast_fp16 = mul(x = out_19_cast_fp16, y = layer_encoder_layers_1_post_attention_layernorm_weight_to_fp16)[name = string("out_21_cast_fp16")]; | |
| tensor<int32, [2]> var_381_split_sizes_0 = const()[name = string("op_381_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; | |
| int32 var_381_axis_0 = const()[name = string("op_381_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_381_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_381_cast_fp16_1 = split(axis = var_381_axis_0, split_sizes = var_381_split_sizes_0, x = out_21_cast_fp16)[name = string("op_381_cast_fp16")]; | |
| string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)]; | |
| tensor<fp16, [4096, 1024, 1, 1]> var_250_to_fp16 = const()[name = string("op_250_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34757184)))]; | |
| tensor<fp16, [?, 4096, 1, 5]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = var_250_to_fp16, x = var_381_cast_fp16_0)[name = string("input_3_cast_fp16")]; | |
| tensor<fp16, [?, 4096, 1, 5]> var_389_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_389_cast_fp16")]; | |
| string var_394_pad_type_0 = const()[name = string("op_394_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> var_394_strides_0 = const()[name = string("op_394_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> var_394_pad_0 = const()[name = string("op_394_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> var_394_dilations_0 = const()[name = string("op_394_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_394_groups_0 = const()[name = string("op_394_groups_0"), val = int32(1)]; | |
| tensor<fp16, [4096, 1024, 1, 1]> var_251_to_fp16 = const()[name = string("op_251_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43145856)))]; | |
| tensor<fp16, [?, 4096, 1, 5]> var_394_cast_fp16 = conv(dilations = var_394_dilations_0, groups = var_394_groups_0, pad = var_394_pad_0, pad_type = var_394_pad_type_0, strides = var_394_strides_0, weight = var_251_to_fp16, x = var_381_cast_fp16_0)[name = string("op_394_cast_fp16")]; | |
| tensor<fp16, [?, 4096, 1, 5]> x_39_cast_fp16 = mul(x = var_389_cast_fp16, y = var_394_cast_fp16)[name = string("x_39_cast_fp16")]; | |
| string hidden_states_11_pad_type_0 = const()[name = string("hidden_states_11_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> hidden_states_11_strides_0 = const()[name = string("hidden_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> hidden_states_11_pad_0 = const()[name = string("hidden_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> hidden_states_11_dilations_0 = const()[name = string("hidden_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 hidden_states_11_groups_0 = const()[name = string("hidden_states_11_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 4096, 1, 1]> var_252_to_fp16 = const()[name = string("op_252_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51534528)))]; | |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_11_cast_fp16 = conv(dilations = hidden_states_11_dilations_0, groups = hidden_states_11_groups_0, pad = hidden_states_11_pad_0, pad_type = hidden_states_11_pad_type_0, strides = hidden_states_11_strides_0, weight = var_252_to_fp16, x = x_39_cast_fp16)[name = string("hidden_states_11_cast_fp16")]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_41_cast_fp16 = add(x = x_31_cast_fp16, y = hidden_states_11_cast_fp16)[name = string("x_41_cast_fp16")]; | |
| int32 var_412 = const()[name = string("op_412"), val = int32(-2)]; | |
| int32 var_416 = const()[name = string("op_416"), val = int32(1)]; | |
| int32 var_421 = const()[name = string("op_421"), val = int32(2)]; | |
| fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_426_cast_fp16 = mul(x = x_41_cast_fp16, y = const_9_promoted_to_fp16)[name = string("op_426_cast_fp16")]; | |
| bool x_43_interleave_0 = const()[name = string("x_43_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2048, 1, 5]> x_43_cast_fp16 = concat(axis = var_416, interleave = x_43_interleave_0, values = (x_41_cast_fp16, var_426_cast_fp16))[name = string("x_43_cast_fp16")]; | |
| tensor<int32, [1]> out_25_axes_0 = const()[name = string("out_25_axes_0"), val = tensor<int32, [1]>([1])]; | |
| fp16 var_436_to_fp16 = const()[name = string("op_436_to_fp16"), val = fp16(0x1.5p-17)]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_436_to_fp16, x = x_43_cast_fp16)[name = string("out_25_cast_fp16")]; | |
| tensor<fp16, [1, 2048, 1, 1]> layer_encoder_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59923200)))]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_27_cast_fp16 = mul(x = out_25_cast_fp16, y = layer_encoder_layers_2_input_layernorm_weight_to_fp16)[name = string("out_27_cast_fp16")]; | |
| tensor<int32, [2]> var_442_split_sizes_0 = const()[name = string("op_442_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; | |
| int32 var_442_axis_0 = const()[name = string("op_442_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_442_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_442_cast_fp16_1 = split(axis = var_442_axis_0, split_sizes = var_442_split_sizes_0, x = out_27_cast_fp16)[name = string("op_442_cast_fp16")]; | |
| string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 1024, 1, 1]> var_407_to_fp16 = const()[name = string("op_407_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59927360)))]; | |
| tensor<fp16, [?, 1024, 1, 5]> query_states_13_cast_fp16 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = var_407_to_fp16, x = var_442_cast_fp16_0)[name = string("query_states_13_cast_fp16")]; | |
| string key_states_13_pad_type_0 = const()[name = string("key_states_13_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> key_states_13_strides_0 = const()[name = string("key_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> key_states_13_pad_0 = const()[name = string("key_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> key_states_13_dilations_0 = const()[name = string("key_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 key_states_13_groups_0 = const()[name = string("key_states_13_groups_0"), val = int32(1)]; | |
| tensor<fp16, [128, 1024, 1, 1]> var_408_to_fp16 = const()[name = string("op_408_to_fp16"), val = tensor<fp16, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62024576)))]; | |
| tensor<fp16, [?, 128, 1, 5]> key_states_13_cast_fp16 = conv(dilations = key_states_13_dilations_0, groups = key_states_13_groups_0, pad = key_states_13_pad_0, pad_type = key_states_13_pad_type_0, strides = key_states_13_strides_0, weight = var_408_to_fp16, x = var_442_cast_fp16_0)[name = string("key_states_13_cast_fp16")]; | |
| string value_states_13_pad_type_0 = const()[name = string("value_states_13_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> value_states_13_strides_0 = const()[name = string("value_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> value_states_13_pad_0 = const()[name = string("value_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> value_states_13_dilations_0 = const()[name = string("value_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 value_states_13_groups_0 = const()[name = string("value_states_13_groups_0"), val = int32(1)]; | |
| tensor<fp16, [128, 1024, 1, 1]> var_409_to_fp16 = const()[name = string("op_409_to_fp16"), val = tensor<fp16, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62286784)))]; | |
| tensor<fp16, [?, 128, 1, 5]> value_states_13_cast_fp16 = conv(dilations = value_states_13_dilations_0, groups = value_states_13_groups_0, pad = value_states_13_pad_0, pad_type = value_states_13_pad_type_0, strides = value_states_13_strides_0, weight = var_409_to_fp16, x = var_442_cast_fp16_0)[name = string("value_states_13_cast_fp16")]; | |
| tensor<int32, [4]> concat_8x = const()[name = string("concat_8x"), val = tensor<int32, [4]>([-1, 16, 64, 5])]; | |
| tensor<fp16, [?, 16, 64, 5]> embed_9_cast_fp16 = reshape(shape = concat_8x, x = query_states_13_cast_fp16)[name = string("embed_9_cast_fp16")]; | |
| tensor<int32, [4]> concat_9x = const()[name = string("concat_9x"), val = tensor<int32, [4]>([-1, 2, 64, 5])]; | |
| tensor<fp16, [?, 2, 64, 5]> embed_11_cast_fp16 = reshape(shape = concat_9x, x = key_states_13_cast_fp16)[name = string("embed_11_cast_fp16")]; | |
| tensor<int32, [4]> concat_10x = const()[name = string("concat_10x"), val = tensor<int32, [4]>([-1, 2, 64, 5])]; | |
| tensor<fp16, [?, 2, 64, 5]> value_states_15_cast_fp16 = reshape(shape = concat_10x, x = value_states_13_cast_fp16)[name = string("value_states_15_cast_fp16")]; | |
| tensor<fp16, [?, 16, 64, 5]> var_468_cast_fp16 = mul(x = embed_9_cast_fp16, y = cos_to_fp16)[name = string("op_468_cast_fp16")]; | |
| tensor<int32, [2]> var_469_split_sizes_0 = const()[name = string("op_469_split_sizes_0"), val = tensor<int32, [2]>([32, 32])]; | |
| int32 var_469_axis_0 = const()[name = string("op_469_axis_0"), val = int32(-2)]; | |
| tensor<fp16, [?, 16, 32, 5]> var_469_cast_fp16_0, tensor<fp16, [?, 16, 32, 5]> var_469_cast_fp16_1 = split(axis = var_469_axis_0, split_sizes = var_469_split_sizes_0, x = embed_9_cast_fp16)[name = string("op_469_cast_fp16")]; | |
| fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 16, 32, 5]> var_471_cast_fp16 = mul(x = var_469_cast_fp16_1, y = const_10_promoted_to_fp16)[name = string("op_471_cast_fp16")]; | |
| bool var_473_interleave_0 = const()[name = string("op_473_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 16, 64, 5]> var_473_cast_fp16 = concat(axis = var_412, interleave = var_473_interleave_0, values = (var_471_cast_fp16, var_469_cast_fp16_0))[name = string("op_473_cast_fp16")]; | |
| tensor<fp16, [?, 16, 64, 5]> var_474_cast_fp16 = mul(x = var_473_cast_fp16, y = sin_to_fp16)[name = string("op_474_cast_fp16")]; | |
| tensor<fp16, [?, 16, 64, 5]> query_states_15_cast_fp16 = add(x = var_468_cast_fp16, y = var_474_cast_fp16)[name = string("query_states_15_cast_fp16")]; | |
| tensor<fp16, [?, 2, 64, 5]> var_476_cast_fp16 = mul(x = embed_11_cast_fp16, y = cos_to_fp16)[name = string("op_476_cast_fp16")]; | |
| tensor<int32, [2]> var_477_split_sizes_0 = const()[name = string("op_477_split_sizes_0"), val = tensor<int32, [2]>([32, 32])]; | |
| int32 var_477_axis_0 = const()[name = string("op_477_axis_0"), val = int32(-2)]; | |
| tensor<fp16, [?, 2, 32, 5]> var_477_cast_fp16_0, tensor<fp16, [?, 2, 32, 5]> var_477_cast_fp16_1 = split(axis = var_477_axis_0, split_sizes = var_477_split_sizes_0, x = embed_11_cast_fp16)[name = string("op_477_cast_fp16")]; | |
| fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 2, 32, 5]> var_479_cast_fp16 = mul(x = var_477_cast_fp16_1, y = const_11_promoted_to_fp16)[name = string("op_479_cast_fp16")]; | |
| bool var_481_interleave_0 = const()[name = string("op_481_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2, 64, 5]> var_481_cast_fp16 = concat(axis = var_412, interleave = var_481_interleave_0, values = (var_479_cast_fp16, var_477_cast_fp16_0))[name = string("op_481_cast_fp16")]; | |
| tensor<fp16, [?, 2, 64, 5]> var_482_cast_fp16 = mul(x = var_481_cast_fp16, y = sin_to_fp16)[name = string("op_482_cast_fp16")]; | |
| tensor<fp16, [?, 2, 64, 5]> key_states_15_cast_fp16 = add(x = var_476_cast_fp16, y = var_482_cast_fp16)[name = string("key_states_15_cast_fp16")]; | |
| tensor<int32, [2]> var_487_split_sizes_0 = const()[name = string("op_487_split_sizes_0"), val = tensor<int32, [2]>([8, 8])]; | |
| int32 var_487_axis_0 = const()[name = string("op_487_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 8, 64, 5]> var_487_cast_fp16_0, tensor<fp16, [?, 8, 64, 5]> var_487_cast_fp16_1 = split(axis = var_487_axis_0, split_sizes = var_487_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_487_cast_fp16")]; | |
| tensor<int32, [2]> var_489_split_sizes_0 = const()[name = string("op_489_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_489_axis_0 = const()[name = string("op_489_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1, 64, 5]> var_489_cast_fp16_0, tensor<fp16, [?, 1, 64, 5]> var_489_cast_fp16_1 = split(axis = var_489_axis_0, split_sizes = var_489_split_sizes_0, x = key_states_15_cast_fp16)[name = string("op_489_cast_fp16")]; | |
| tensor<int32, [2]> var_491_split_sizes_0 = const()[name = string("op_491_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_491_axis_0 = const()[name = string("op_491_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1, 64, 5]> var_491_cast_fp16_0, tensor<fp16, [?, 1, 64, 5]> var_491_cast_fp16_1 = split(axis = var_491_axis_0, split_sizes = var_491_split_sizes_0, x = value_states_15_cast_fp16)[name = string("op_491_cast_fp16")]; | |
| bool attn_weights_25_transpose_x_1 = const()[name = string("attn_weights_25_transpose_x_1"), val = bool(true)]; | |
| bool attn_weights_25_transpose_y_1 = const()[name = string("attn_weights_25_transpose_y_1"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_25_cast_fp16 = matmul(transpose_x = attn_weights_25_transpose_x_1, transpose_y = attn_weights_25_transpose_y_1, x = var_489_cast_fp16_0, y = var_487_cast_fp16_0)[name = string("attn_weights_25_cast_fp16")]; | |
| fp16 _inversed_attn_weights_27_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_27_y_0_to_fp16"), val = fp16(0x1p-3)]; | |
| tensor<fp16, [?, 8, 5, 5]> _inversed_attn_weights_27_cast_fp16 = mul(x = attn_weights_25_cast_fp16, y = _inversed_attn_weights_27_y_0_to_fp16)[name = string("_inversed_attn_weights_27_cast_fp16")]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_29_cast_fp16 = softmax(axis = var_421, x = _inversed_attn_weights_27_cast_fp16)[name = string("attn_weights_29_cast_fp16")]; | |
| bool var_498_transpose_x_0 = const()[name = string("op_498_transpose_x_0"), val = bool(false)]; | |
| bool var_498_transpose_y_0 = const()[name = string("op_498_transpose_y_0"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 64, 5]> var_498_cast_fp16 = matmul(transpose_x = var_498_transpose_x_0, transpose_y = var_498_transpose_y_0, x = var_491_cast_fp16_0, y = attn_weights_29_cast_fp16)[name = string("op_498_cast_fp16")]; | |
| bool attn_weights_31_transpose_x_1 = const()[name = string("attn_weights_31_transpose_x_1"), val = bool(true)]; | |
| bool attn_weights_31_transpose_y_1 = const()[name = string("attn_weights_31_transpose_y_1"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_31_cast_fp16 = matmul(transpose_x = attn_weights_31_transpose_x_1, transpose_y = attn_weights_31_transpose_y_1, x = var_489_cast_fp16_1, y = var_487_cast_fp16_1)[name = string("attn_weights_31_cast_fp16")]; | |
| fp16 _inversed_attn_weights_33_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_33_y_0_to_fp16"), val = fp16(0x1p-3)]; | |
| tensor<fp16, [?, 8, 5, 5]> _inversed_attn_weights_33_cast_fp16 = mul(x = attn_weights_31_cast_fp16, y = _inversed_attn_weights_33_y_0_to_fp16)[name = string("_inversed_attn_weights_33_cast_fp16")]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_35_cast_fp16 = softmax(axis = var_421, x = _inversed_attn_weights_33_cast_fp16)[name = string("attn_weights_35_cast_fp16")]; | |
| bool attn_output_9_transpose_x_0 = const()[name = string("attn_output_9_transpose_x_0"), val = bool(false)]; | |
| bool attn_output_9_transpose_y_0 = const()[name = string("attn_output_9_transpose_y_0"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 64, 5]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = var_491_cast_fp16_1, y = attn_weights_35_cast_fp16)[name = string("attn_output_9_cast_fp16")]; | |
| bool attn_output_11_interleave_0 = const()[name = string("attn_output_11_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 16, 64, 5]> attn_output_11_cast_fp16 = concat(axis = var_416, interleave = attn_output_11_interleave_0, values = (var_498_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")]; | |
| tensor<int32, [4]> concat_11x = const()[name = string("concat_11x"), val = tensor<int32, [4]>([-1, 1024, 1, 5])]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_47_cast_fp16 = reshape(shape = concat_11x, x = attn_output_11_cast_fp16)[name = string("x_47_cast_fp16")]; | |
| string hidden_states_15_pad_type_0 = const()[name = string("hidden_states_15_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> hidden_states_15_strides_0 = const()[name = string("hidden_states_15_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> hidden_states_15_pad_0 = const()[name = string("hidden_states_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> hidden_states_15_dilations_0 = const()[name = string("hidden_states_15_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 hidden_states_15_groups_0 = const()[name = string("hidden_states_15_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 1024, 1, 1]> var_415_to_fp16 = const()[name = string("op_415_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62548992)))]; | |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_15_cast_fp16 = conv(dilations = hidden_states_15_dilations_0, groups = hidden_states_15_groups_0, pad = hidden_states_15_pad_0, pad_type = hidden_states_15_pad_type_0, strides = hidden_states_15_strides_0, weight = var_415_to_fp16, x = x_47_cast_fp16)[name = string("hidden_states_15_cast_fp16")]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_49_cast_fp16 = add(x = x_41_cast_fp16, y = hidden_states_15_cast_fp16)[name = string("x_49_cast_fp16")]; | |
| fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_517_cast_fp16 = mul(x = x_49_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_517_cast_fp16")]; | |
| bool x_51_interleave_0 = const()[name = string("x_51_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2048, 1, 5]> x_51_cast_fp16 = concat(axis = var_416, interleave = x_51_interleave_0, values = (x_49_cast_fp16, var_517_cast_fp16))[name = string("x_51_cast_fp16")]; | |
| tensor<int32, [1]> out_31_axes_0 = const()[name = string("out_31_axes_0"), val = tensor<int32, [1]>([1])]; | |
| fp16 var_527_to_fp16 = const()[name = string("op_527_to_fp16"), val = fp16(0x1.5p-17)]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_527_to_fp16, x = x_51_cast_fp16)[name = string("out_31_cast_fp16")]; | |
| tensor<fp16, [1, 2048, 1, 1]> layer_encoder_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64646208)))]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_33_cast_fp16 = mul(x = out_31_cast_fp16, y = layer_encoder_layers_2_post_attention_layernorm_weight_to_fp16)[name = string("out_33_cast_fp16")]; | |
| tensor<int32, [2]> var_533_split_sizes_0 = const()[name = string("op_533_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; | |
| int32 var_533_axis_0 = const()[name = string("op_533_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_533_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_533_cast_fp16_1 = split(axis = var_533_axis_0, split_sizes = var_533_split_sizes_0, x = out_33_cast_fp16)[name = string("op_533_cast_fp16")]; | |
| string input_5_pad_type_0 = const()[name = string("input_5_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> input_5_strides_0 = const()[name = string("input_5_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> input_5_dilations_0 = const()[name = string("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 input_5_groups_0 = const()[name = string("input_5_groups_0"), val = int32(1)]; | |
| tensor<fp16, [4096, 1024, 1, 1]> var_402_to_fp16 = const()[name = string("op_402_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64650368)))]; | |
| tensor<fp16, [?, 4096, 1, 5]> input_5_cast_fp16 = conv(dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = var_402_to_fp16, x = var_533_cast_fp16_0)[name = string("input_5_cast_fp16")]; | |
| tensor<fp16, [?, 4096, 1, 5]> var_541_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_541_cast_fp16")]; | |
| string var_546_pad_type_0 = const()[name = string("op_546_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> var_546_strides_0 = const()[name = string("op_546_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> var_546_pad_0 = const()[name = string("op_546_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> var_546_dilations_0 = const()[name = string("op_546_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_546_groups_0 = const()[name = string("op_546_groups_0"), val = int32(1)]; | |
| tensor<fp16, [4096, 1024, 1, 1]> var_403_to_fp16 = const()[name = string("op_403_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73039040)))]; | |
| tensor<fp16, [?, 4096, 1, 5]> var_546_cast_fp16 = conv(dilations = var_546_dilations_0, groups = var_546_groups_0, pad = var_546_pad_0, pad_type = var_546_pad_type_0, strides = var_546_strides_0, weight = var_403_to_fp16, x = var_533_cast_fp16_0)[name = string("op_546_cast_fp16")]; | |
| tensor<fp16, [?, 4096, 1, 5]> x_57_cast_fp16 = mul(x = var_541_cast_fp16, y = var_546_cast_fp16)[name = string("x_57_cast_fp16")]; | |
| string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 4096, 1, 1]> var_404_to_fp16 = const()[name = string("op_404_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81427712)))]; | |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_17_cast_fp16 = conv(dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = var_404_to_fp16, x = x_57_cast_fp16)[name = string("hidden_states_17_cast_fp16")]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_59_cast_fp16 = add(x = x_49_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("x_59_cast_fp16")]; | |
| int32 var_564 = const()[name = string("op_564"), val = int32(-2)]; | |
| int32 var_568 = const()[name = string("op_568"), val = int32(1)]; | |
| int32 var_573 = const()[name = string("op_573"), val = int32(2)]; | |
| fp16 const_13_promoted_to_fp16 = const()[name = string("const_13_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_578_cast_fp16 = mul(x = x_59_cast_fp16, y = const_13_promoted_to_fp16)[name = string("op_578_cast_fp16")]; | |
| bool x_61_interleave_0 = const()[name = string("x_61_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2048, 1, 5]> x_61_cast_fp16 = concat(axis = var_568, interleave = x_61_interleave_0, values = (x_59_cast_fp16, var_578_cast_fp16))[name = string("x_61_cast_fp16")]; | |
| tensor<int32, [1]> out_37_axes_0 = const()[name = string("out_37_axes_0"), val = tensor<int32, [1]>([1])]; | |
| fp16 var_588_to_fp16 = const()[name = string("op_588_to_fp16"), val = fp16(0x1.5p-17)]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_588_to_fp16, x = x_61_cast_fp16)[name = string("out_37_cast_fp16")]; | |
| tensor<fp16, [1, 2048, 1, 1]> layer_encoder_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89816384)))]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_39_cast_fp16 = mul(x = out_37_cast_fp16, y = layer_encoder_layers_3_input_layernorm_weight_to_fp16)[name = string("out_39_cast_fp16")]; | |
| tensor<int32, [2]> var_594_split_sizes_0 = const()[name = string("op_594_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; | |
| int32 var_594_axis_0 = const()[name = string("op_594_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_594_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_594_cast_fp16_1 = split(axis = var_594_axis_0, split_sizes = var_594_split_sizes_0, x = out_39_cast_fp16)[name = string("op_594_cast_fp16")]; | |
| string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> query_states_19_pad_0 = const()[name = string("query_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> query_states_19_dilations_0 = const()[name = string("query_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 query_states_19_groups_0 = const()[name = string("query_states_19_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 1024, 1, 1]> var_559_to_fp16 = const()[name = string("op_559_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89820544)))]; | |
| tensor<fp16, [?, 1024, 1, 5]> query_states_19_cast_fp16 = conv(dilations = query_states_19_dilations_0, groups = query_states_19_groups_0, pad = query_states_19_pad_0, pad_type = query_states_19_pad_type_0, strides = query_states_19_strides_0, weight = var_559_to_fp16, x = var_594_cast_fp16_0)[name = string("query_states_19_cast_fp16")]; | |
| string key_states_19_pad_type_0 = const()[name = string("key_states_19_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> key_states_19_strides_0 = const()[name = string("key_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> key_states_19_pad_0 = const()[name = string("key_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> key_states_19_dilations_0 = const()[name = string("key_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 key_states_19_groups_0 = const()[name = string("key_states_19_groups_0"), val = int32(1)]; | |
| tensor<fp16, [128, 1024, 1, 1]> var_560_to_fp16 = const()[name = string("op_560_to_fp16"), val = tensor<fp16, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91917760)))]; | |
| tensor<fp16, [?, 128, 1, 5]> key_states_19_cast_fp16 = conv(dilations = key_states_19_dilations_0, groups = key_states_19_groups_0, pad = key_states_19_pad_0, pad_type = key_states_19_pad_type_0, strides = key_states_19_strides_0, weight = var_560_to_fp16, x = var_594_cast_fp16_0)[name = string("key_states_19_cast_fp16")]; | |
| string value_states_19_pad_type_0 = const()[name = string("value_states_19_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> value_states_19_strides_0 = const()[name = string("value_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> value_states_19_pad_0 = const()[name = string("value_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> value_states_19_dilations_0 = const()[name = string("value_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 value_states_19_groups_0 = const()[name = string("value_states_19_groups_0"), val = int32(1)]; | |
| tensor<fp16, [128, 1024, 1, 1]> var_561_to_fp16 = const()[name = string("op_561_to_fp16"), val = tensor<fp16, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92179968)))]; | |
| tensor<fp16, [?, 128, 1, 5]> value_states_19_cast_fp16 = conv(dilations = value_states_19_dilations_0, groups = value_states_19_groups_0, pad = value_states_19_pad_0, pad_type = value_states_19_pad_type_0, strides = value_states_19_strides_0, weight = var_561_to_fp16, x = var_594_cast_fp16_0)[name = string("value_states_19_cast_fp16")]; | |
| tensor<int32, [4]> concat_12x = const()[name = string("concat_12x"), val = tensor<int32, [4]>([-1, 16, 64, 5])]; | |
| tensor<fp16, [?, 16, 64, 5]> embed_13_cast_fp16 = reshape(shape = concat_12x, x = query_states_19_cast_fp16)[name = string("embed_13_cast_fp16")]; | |
| tensor<int32, [4]> concat_13x = const()[name = string("concat_13x"), val = tensor<int32, [4]>([-1, 2, 64, 5])]; | |
| tensor<fp16, [?, 2, 64, 5]> embed_15_cast_fp16 = reshape(shape = concat_13x, x = key_states_19_cast_fp16)[name = string("embed_15_cast_fp16")]; | |
| tensor<int32, [4]> concat_14x = const()[name = string("concat_14x"), val = tensor<int32, [4]>([-1, 2, 64, 5])]; | |
| tensor<fp16, [?, 2, 64, 5]> value_states_21_cast_fp16 = reshape(shape = concat_14x, x = value_states_19_cast_fp16)[name = string("value_states_21_cast_fp16")]; | |
| tensor<fp16, [?, 16, 64, 5]> var_620_cast_fp16 = mul(x = embed_13_cast_fp16, y = cos_to_fp16)[name = string("op_620_cast_fp16")]; | |
| tensor<int32, [2]> var_621_split_sizes_0 = const()[name = string("op_621_split_sizes_0"), val = tensor<int32, [2]>([32, 32])]; | |
| int32 var_621_axis_0 = const()[name = string("op_621_axis_0"), val = int32(-2)]; | |
| tensor<fp16, [?, 16, 32, 5]> var_621_cast_fp16_0, tensor<fp16, [?, 16, 32, 5]> var_621_cast_fp16_1 = split(axis = var_621_axis_0, split_sizes = var_621_split_sizes_0, x = embed_13_cast_fp16)[name = string("op_621_cast_fp16")]; | |
| fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 16, 32, 5]> var_623_cast_fp16 = mul(x = var_621_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_623_cast_fp16")]; | |
| bool var_625_interleave_0 = const()[name = string("op_625_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 16, 64, 5]> var_625_cast_fp16 = concat(axis = var_564, interleave = var_625_interleave_0, values = (var_623_cast_fp16, var_621_cast_fp16_0))[name = string("op_625_cast_fp16")]; | |
| tensor<fp16, [?, 16, 64, 5]> var_626_cast_fp16 = mul(x = var_625_cast_fp16, y = sin_to_fp16)[name = string("op_626_cast_fp16")]; | |
| tensor<fp16, [?, 16, 64, 5]> query_states_21_cast_fp16 = add(x = var_620_cast_fp16, y = var_626_cast_fp16)[name = string("query_states_21_cast_fp16")]; | |
| tensor<fp16, [?, 2, 64, 5]> var_628_cast_fp16 = mul(x = embed_15_cast_fp16, y = cos_to_fp16)[name = string("op_628_cast_fp16")]; | |
| tensor<int32, [2]> var_629_split_sizes_0 = const()[name = string("op_629_split_sizes_0"), val = tensor<int32, [2]>([32, 32])]; | |
| int32 var_629_axis_0 = const()[name = string("op_629_axis_0"), val = int32(-2)]; | |
| tensor<fp16, [?, 2, 32, 5]> var_629_cast_fp16_0, tensor<fp16, [?, 2, 32, 5]> var_629_cast_fp16_1 = split(axis = var_629_axis_0, split_sizes = var_629_split_sizes_0, x = embed_15_cast_fp16)[name = string("op_629_cast_fp16")]; | |
| fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 2, 32, 5]> var_631_cast_fp16 = mul(x = var_629_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_631_cast_fp16")]; | |
| bool var_633_interleave_0 = const()[name = string("op_633_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2, 64, 5]> var_633_cast_fp16 = concat(axis = var_564, interleave = var_633_interleave_0, values = (var_631_cast_fp16, var_629_cast_fp16_0))[name = string("op_633_cast_fp16")]; | |
| tensor<fp16, [?, 2, 64, 5]> var_634_cast_fp16 = mul(x = var_633_cast_fp16, y = sin_to_fp16)[name = string("op_634_cast_fp16")]; | |
| tensor<fp16, [?, 2, 64, 5]> key_states_21_cast_fp16 = add(x = var_628_cast_fp16, y = var_634_cast_fp16)[name = string("key_states_21_cast_fp16")]; | |
| tensor<int32, [2]> var_639_split_sizes_0 = const()[name = string("op_639_split_sizes_0"), val = tensor<int32, [2]>([8, 8])]; | |
| int32 var_639_axis_0 = const()[name = string("op_639_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 8, 64, 5]> var_639_cast_fp16_0, tensor<fp16, [?, 8, 64, 5]> var_639_cast_fp16_1 = split(axis = var_639_axis_0, split_sizes = var_639_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_639_cast_fp16")]; | |
| tensor<int32, [2]> var_641_split_sizes_0 = const()[name = string("op_641_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_641_axis_0 = const()[name = string("op_641_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1, 64, 5]> var_641_cast_fp16_0, tensor<fp16, [?, 1, 64, 5]> var_641_cast_fp16_1 = split(axis = var_641_axis_0, split_sizes = var_641_split_sizes_0, x = key_states_21_cast_fp16)[name = string("op_641_cast_fp16")]; | |
| tensor<int32, [2]> var_643_split_sizes_0 = const()[name = string("op_643_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_643_axis_0 = const()[name = string("op_643_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1, 64, 5]> var_643_cast_fp16_0, tensor<fp16, [?, 1, 64, 5]> var_643_cast_fp16_1 = split(axis = var_643_axis_0, split_sizes = var_643_split_sizes_0, x = value_states_21_cast_fp16)[name = string("op_643_cast_fp16")]; | |
| bool attn_weights_37_transpose_x_1 = const()[name = string("attn_weights_37_transpose_x_1"), val = bool(true)]; | |
| bool attn_weights_37_transpose_y_1 = const()[name = string("attn_weights_37_transpose_y_1"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_37_cast_fp16 = matmul(transpose_x = attn_weights_37_transpose_x_1, transpose_y = attn_weights_37_transpose_y_1, x = var_641_cast_fp16_0, y = var_639_cast_fp16_0)[name = string("attn_weights_37_cast_fp16")]; | |
| fp16 _inversed_attn_weights_39_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_39_y_0_to_fp16"), val = fp16(0x1p-3)]; | |
| tensor<fp16, [?, 8, 5, 5]> _inversed_attn_weights_39_cast_fp16 = mul(x = attn_weights_37_cast_fp16, y = _inversed_attn_weights_39_y_0_to_fp16)[name = string("_inversed_attn_weights_39_cast_fp16")]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_41_cast_fp16 = softmax(axis = var_573, x = _inversed_attn_weights_39_cast_fp16)[name = string("attn_weights_41_cast_fp16")]; | |
| bool var_650_transpose_x_0 = const()[name = string("op_650_transpose_x_0"), val = bool(false)]; | |
| bool var_650_transpose_y_0 = const()[name = string("op_650_transpose_y_0"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 64, 5]> var_650_cast_fp16 = matmul(transpose_x = var_650_transpose_x_0, transpose_y = var_650_transpose_y_0, x = var_643_cast_fp16_0, y = attn_weights_41_cast_fp16)[name = string("op_650_cast_fp16")]; | |
| bool attn_weights_43_transpose_x_1 = const()[name = string("attn_weights_43_transpose_x_1"), val = bool(true)]; | |
| bool attn_weights_43_transpose_y_1 = const()[name = string("attn_weights_43_transpose_y_1"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_43_cast_fp16 = matmul(transpose_x = attn_weights_43_transpose_x_1, transpose_y = attn_weights_43_transpose_y_1, x = var_641_cast_fp16_1, y = var_639_cast_fp16_1)[name = string("attn_weights_43_cast_fp16")]; | |
| fp16 _inversed_attn_weights_45_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_45_y_0_to_fp16"), val = fp16(0x1p-3)]; | |
| tensor<fp16, [?, 8, 5, 5]> _inversed_attn_weights_45_cast_fp16 = mul(x = attn_weights_43_cast_fp16, y = _inversed_attn_weights_45_y_0_to_fp16)[name = string("_inversed_attn_weights_45_cast_fp16")]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_47_cast_fp16 = softmax(axis = var_573, x = _inversed_attn_weights_45_cast_fp16)[name = string("attn_weights_47_cast_fp16")]; | |
| bool attn_output_13_transpose_x_0 = const()[name = string("attn_output_13_transpose_x_0"), val = bool(false)]; | |
| bool attn_output_13_transpose_y_0 = const()[name = string("attn_output_13_transpose_y_0"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 64, 5]> attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = var_643_cast_fp16_1, y = attn_weights_47_cast_fp16)[name = string("attn_output_13_cast_fp16")]; | |
| bool attn_output_15_interleave_0 = const()[name = string("attn_output_15_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 16, 64, 5]> attn_output_15_cast_fp16 = concat(axis = var_568, interleave = attn_output_15_interleave_0, values = (var_650_cast_fp16, attn_output_13_cast_fp16))[name = string("attn_output_15_cast_fp16")]; | |
| tensor<int32, [4]> concat_15x = const()[name = string("concat_15x"), val = tensor<int32, [4]>([-1, 1024, 1, 5])]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_65_cast_fp16 = reshape(shape = concat_15x, x = attn_output_15_cast_fp16)[name = string("x_65_cast_fp16")]; | |
| string hidden_states_21_pad_type_0 = const()[name = string("hidden_states_21_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> hidden_states_21_strides_0 = const()[name = string("hidden_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> hidden_states_21_pad_0 = const()[name = string("hidden_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> hidden_states_21_dilations_0 = const()[name = string("hidden_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 hidden_states_21_groups_0 = const()[name = string("hidden_states_21_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 1024, 1, 1]> var_567_to_fp16 = const()[name = string("op_567_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92442176)))]; | |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_21_cast_fp16 = conv(dilations = hidden_states_21_dilations_0, groups = hidden_states_21_groups_0, pad = hidden_states_21_pad_0, pad_type = hidden_states_21_pad_type_0, strides = hidden_states_21_strides_0, weight = var_567_to_fp16, x = x_65_cast_fp16)[name = string("hidden_states_21_cast_fp16")]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_67_cast_fp16 = add(x = x_59_cast_fp16, y = hidden_states_21_cast_fp16)[name = string("x_67_cast_fp16")]; | |
| fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_669_cast_fp16 = mul(x = x_67_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_669_cast_fp16")]; | |
| bool x_69_interleave_0 = const()[name = string("x_69_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2048, 1, 5]> x_69_cast_fp16 = concat(axis = var_568, interleave = x_69_interleave_0, values = (x_67_cast_fp16, var_669_cast_fp16))[name = string("x_69_cast_fp16")]; | |
| tensor<int32, [1]> out_43_axes_0 = const()[name = string("out_43_axes_0"), val = tensor<int32, [1]>([1])]; | |
| fp16 var_679_to_fp16 = const()[name = string("op_679_to_fp16"), val = fp16(0x1.5p-17)]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_679_to_fp16, x = x_69_cast_fp16)[name = string("out_43_cast_fp16")]; | |
| tensor<fp16, [1, 2048, 1, 1]> layer_encoder_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94539392)))]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_45_cast_fp16 = mul(x = out_43_cast_fp16, y = layer_encoder_layers_3_post_attention_layernorm_weight_to_fp16)[name = string("out_45_cast_fp16")]; | |
| tensor<int32, [2]> var_685_split_sizes_0 = const()[name = string("op_685_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; | |
| int32 var_685_axis_0 = const()[name = string("op_685_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_685_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_685_cast_fp16_1 = split(axis = var_685_axis_0, split_sizes = var_685_split_sizes_0, x = out_45_cast_fp16)[name = string("op_685_cast_fp16")]; | |
| string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)]; | |
| tensor<fp16, [4096, 1024, 1, 1]> var_554_to_fp16 = const()[name = string("op_554_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94543552)))]; | |
| tensor<fp16, [?, 4096, 1, 5]> input_7_cast_fp16 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = var_554_to_fp16, x = var_685_cast_fp16_0)[name = string("input_7_cast_fp16")]; | |
| tensor<fp16, [?, 4096, 1, 5]> var_693_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_693_cast_fp16")]; | |
| string var_698_pad_type_0 = const()[name = string("op_698_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> var_698_strides_0 = const()[name = string("op_698_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> var_698_pad_0 = const()[name = string("op_698_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> var_698_dilations_0 = const()[name = string("op_698_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_698_groups_0 = const()[name = string("op_698_groups_0"), val = int32(1)]; | |
| tensor<fp16, [4096, 1024, 1, 1]> var_555_to_fp16 = const()[name = string("op_555_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102932224)))]; | |
| tensor<fp16, [?, 4096, 1, 5]> var_698_cast_fp16 = conv(dilations = var_698_dilations_0, groups = var_698_groups_0, pad = var_698_pad_0, pad_type = var_698_pad_type_0, strides = var_698_strides_0, weight = var_555_to_fp16, x = var_685_cast_fp16_0)[name = string("op_698_cast_fp16")]; | |
| tensor<fp16, [?, 4096, 1, 5]> x_75_cast_fp16 = mul(x = var_693_cast_fp16, y = var_698_cast_fp16)[name = string("x_75_cast_fp16")]; | |
| string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 4096, 1, 1]> var_556_to_fp16 = const()[name = string("op_556_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111320896)))]; | |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_23_cast_fp16 = conv(dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = var_556_to_fp16, x = x_75_cast_fp16)[name = string("hidden_states_23_cast_fp16")]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_77_cast_fp16 = add(x = x_67_cast_fp16, y = hidden_states_23_cast_fp16)[name = string("x_77_cast_fp16")]; | |
| int32 var_716 = const()[name = string("op_716"), val = int32(-2)]; | |
| int32 var_720 = const()[name = string("op_720"), val = int32(1)]; | |
| int32 var_725 = const()[name = string("op_725"), val = int32(2)]; | |
| fp16 const_17_promoted_to_fp16 = const()[name = string("const_17_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_730_cast_fp16 = mul(x = x_77_cast_fp16, y = const_17_promoted_to_fp16)[name = string("op_730_cast_fp16")]; | |
| bool x_79_interleave_0 = const()[name = string("x_79_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2048, 1, 5]> x_79_cast_fp16 = concat(axis = var_720, interleave = x_79_interleave_0, values = (x_77_cast_fp16, var_730_cast_fp16))[name = string("x_79_cast_fp16")]; | |
| tensor<int32, [1]> out_49_axes_0 = const()[name = string("out_49_axes_0"), val = tensor<int32, [1]>([1])]; | |
| fp16 var_740_to_fp16 = const()[name = string("op_740_to_fp16"), val = fp16(0x1.5p-17)]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_740_to_fp16, x = x_79_cast_fp16)[name = string("out_49_cast_fp16")]; | |
| tensor<fp16, [1, 2048, 1, 1]> layer_encoder_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119709568)))]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_51_cast_fp16 = mul(x = out_49_cast_fp16, y = layer_encoder_layers_4_input_layernorm_weight_to_fp16)[name = string("out_51_cast_fp16")]; | |
| tensor<int32, [2]> var_746_split_sizes_0 = const()[name = string("op_746_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; | |
| int32 var_746_axis_0 = const()[name = string("op_746_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_746_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_746_cast_fp16_1 = split(axis = var_746_axis_0, split_sizes = var_746_split_sizes_0, x = out_51_cast_fp16)[name = string("op_746_cast_fp16")]; | |
| string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> query_states_25_pad_0 = const()[name = string("query_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> query_states_25_dilations_0 = const()[name = string("query_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 query_states_25_groups_0 = const()[name = string("query_states_25_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 1024, 1, 1]> var_711_to_fp16 = const()[name = string("op_711_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119713728)))]; | |
| tensor<fp16, [?, 1024, 1, 5]> query_states_25_cast_fp16 = conv(dilations = query_states_25_dilations_0, groups = query_states_25_groups_0, pad = query_states_25_pad_0, pad_type = query_states_25_pad_type_0, strides = query_states_25_strides_0, weight = var_711_to_fp16, x = var_746_cast_fp16_0)[name = string("query_states_25_cast_fp16")]; | |
| string key_states_25_pad_type_0 = const()[name = string("key_states_25_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> key_states_25_strides_0 = const()[name = string("key_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> key_states_25_pad_0 = const()[name = string("key_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> key_states_25_dilations_0 = const()[name = string("key_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 key_states_25_groups_0 = const()[name = string("key_states_25_groups_0"), val = int32(1)]; | |
| tensor<fp16, [128, 1024, 1, 1]> var_712_to_fp16 = const()[name = string("op_712_to_fp16"), val = tensor<fp16, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121810944)))]; | |
| tensor<fp16, [?, 128, 1, 5]> key_states_25_cast_fp16 = conv(dilations = key_states_25_dilations_0, groups = key_states_25_groups_0, pad = key_states_25_pad_0, pad_type = key_states_25_pad_type_0, strides = key_states_25_strides_0, weight = var_712_to_fp16, x = var_746_cast_fp16_0)[name = string("key_states_25_cast_fp16")]; | |
| string value_states_25_pad_type_0 = const()[name = string("value_states_25_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> value_states_25_strides_0 = const()[name = string("value_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> value_states_25_pad_0 = const()[name = string("value_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> value_states_25_dilations_0 = const()[name = string("value_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 value_states_25_groups_0 = const()[name = string("value_states_25_groups_0"), val = int32(1)]; | |
| tensor<fp16, [128, 1024, 1, 1]> var_713_to_fp16 = const()[name = string("op_713_to_fp16"), val = tensor<fp16, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122073152)))]; | |
| tensor<fp16, [?, 128, 1, 5]> value_states_25_cast_fp16 = conv(dilations = value_states_25_dilations_0, groups = value_states_25_groups_0, pad = value_states_25_pad_0, pad_type = value_states_25_pad_type_0, strides = value_states_25_strides_0, weight = var_713_to_fp16, x = var_746_cast_fp16_0)[name = string("value_states_25_cast_fp16")]; | |
| tensor<int32, [4]> concat_16x = const()[name = string("concat_16x"), val = tensor<int32, [4]>([-1, 16, 64, 5])]; | |
| tensor<fp16, [?, 16, 64, 5]> embed_17_cast_fp16 = reshape(shape = concat_16x, x = query_states_25_cast_fp16)[name = string("embed_17_cast_fp16")]; | |
| tensor<int32, [4]> concat_17x = const()[name = string("concat_17x"), val = tensor<int32, [4]>([-1, 2, 64, 5])]; | |
| tensor<fp16, [?, 2, 64, 5]> embed_19_cast_fp16 = reshape(shape = concat_17x, x = key_states_25_cast_fp16)[name = string("embed_19_cast_fp16")]; | |
| tensor<int32, [4]> concat_18x = const()[name = string("concat_18x"), val = tensor<int32, [4]>([-1, 2, 64, 5])]; | |
| tensor<fp16, [?, 2, 64, 5]> value_states_27_cast_fp16 = reshape(shape = concat_18x, x = value_states_25_cast_fp16)[name = string("value_states_27_cast_fp16")]; | |
| tensor<fp16, [?, 16, 64, 5]> var_772_cast_fp16 = mul(x = embed_17_cast_fp16, y = cos_to_fp16)[name = string("op_772_cast_fp16")]; | |
| tensor<int32, [2]> var_773_split_sizes_0 = const()[name = string("op_773_split_sizes_0"), val = tensor<int32, [2]>([32, 32])]; | |
| int32 var_773_axis_0 = const()[name = string("op_773_axis_0"), val = int32(-2)]; | |
| tensor<fp16, [?, 16, 32, 5]> var_773_cast_fp16_0, tensor<fp16, [?, 16, 32, 5]> var_773_cast_fp16_1 = split(axis = var_773_axis_0, split_sizes = var_773_split_sizes_0, x = embed_17_cast_fp16)[name = string("op_773_cast_fp16")]; | |
| fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 16, 32, 5]> var_775_cast_fp16 = mul(x = var_773_cast_fp16_1, y = const_18_promoted_to_fp16)[name = string("op_775_cast_fp16")]; | |
| bool var_777_interleave_0 = const()[name = string("op_777_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 16, 64, 5]> var_777_cast_fp16 = concat(axis = var_716, interleave = var_777_interleave_0, values = (var_775_cast_fp16, var_773_cast_fp16_0))[name = string("op_777_cast_fp16")]; | |
| tensor<fp16, [?, 16, 64, 5]> var_778_cast_fp16 = mul(x = var_777_cast_fp16, y = sin_to_fp16)[name = string("op_778_cast_fp16")]; | |
| tensor<fp16, [?, 16, 64, 5]> query_states_27_cast_fp16 = add(x = var_772_cast_fp16, y = var_778_cast_fp16)[name = string("query_states_27_cast_fp16")]; | |
| tensor<fp16, [?, 2, 64, 5]> var_780_cast_fp16 = mul(x = embed_19_cast_fp16, y = cos_to_fp16)[name = string("op_780_cast_fp16")]; | |
| tensor<int32, [2]> var_781_split_sizes_0 = const()[name = string("op_781_split_sizes_0"), val = tensor<int32, [2]>([32, 32])]; | |
| int32 var_781_axis_0 = const()[name = string("op_781_axis_0"), val = int32(-2)]; | |
| tensor<fp16, [?, 2, 32, 5]> var_781_cast_fp16_0, tensor<fp16, [?, 2, 32, 5]> var_781_cast_fp16_1 = split(axis = var_781_axis_0, split_sizes = var_781_split_sizes_0, x = embed_19_cast_fp16)[name = string("op_781_cast_fp16")]; | |
| fp16 const_19_promoted_to_fp16 = const()[name = string("const_19_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 2, 32, 5]> var_783_cast_fp16 = mul(x = var_781_cast_fp16_1, y = const_19_promoted_to_fp16)[name = string("op_783_cast_fp16")]; | |
| bool var_785_interleave_0 = const()[name = string("op_785_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2, 64, 5]> var_785_cast_fp16 = concat(axis = var_716, interleave = var_785_interleave_0, values = (var_783_cast_fp16, var_781_cast_fp16_0))[name = string("op_785_cast_fp16")]; | |
| tensor<fp16, [?, 2, 64, 5]> var_786_cast_fp16 = mul(x = var_785_cast_fp16, y = sin_to_fp16)[name = string("op_786_cast_fp16")]; | |
| tensor<fp16, [?, 2, 64, 5]> key_states_27_cast_fp16 = add(x = var_780_cast_fp16, y = var_786_cast_fp16)[name = string("key_states_27_cast_fp16")]; | |
| tensor<int32, [2]> var_791_split_sizes_0 = const()[name = string("op_791_split_sizes_0"), val = tensor<int32, [2]>([8, 8])]; | |
| int32 var_791_axis_0 = const()[name = string("op_791_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 8, 64, 5]> var_791_cast_fp16_0, tensor<fp16, [?, 8, 64, 5]> var_791_cast_fp16_1 = split(axis = var_791_axis_0, split_sizes = var_791_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_791_cast_fp16")]; | |
| tensor<int32, [2]> var_793_split_sizes_0 = const()[name = string("op_793_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_793_axis_0 = const()[name = string("op_793_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1, 64, 5]> var_793_cast_fp16_0, tensor<fp16, [?, 1, 64, 5]> var_793_cast_fp16_1 = split(axis = var_793_axis_0, split_sizes = var_793_split_sizes_0, x = key_states_27_cast_fp16)[name = string("op_793_cast_fp16")]; | |
| tensor<int32, [2]> var_795_split_sizes_0 = const()[name = string("op_795_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_795_axis_0 = const()[name = string("op_795_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1, 64, 5]> var_795_cast_fp16_0, tensor<fp16, [?, 1, 64, 5]> var_795_cast_fp16_1 = split(axis = var_795_axis_0, split_sizes = var_795_split_sizes_0, x = value_states_27_cast_fp16)[name = string("op_795_cast_fp16")]; | |
| bool attn_weights_49_transpose_x_1 = const()[name = string("attn_weights_49_transpose_x_1"), val = bool(true)]; | |
| bool attn_weights_49_transpose_y_1 = const()[name = string("attn_weights_49_transpose_y_1"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_49_cast_fp16 = matmul(transpose_x = attn_weights_49_transpose_x_1, transpose_y = attn_weights_49_transpose_y_1, x = var_793_cast_fp16_0, y = var_791_cast_fp16_0)[name = string("attn_weights_49_cast_fp16")]; | |
| fp16 _inversed_attn_weights_51_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_51_y_0_to_fp16"), val = fp16(0x1p-3)]; | |
| tensor<fp16, [?, 8, 5, 5]> _inversed_attn_weights_51_cast_fp16 = mul(x = attn_weights_49_cast_fp16, y = _inversed_attn_weights_51_y_0_to_fp16)[name = string("_inversed_attn_weights_51_cast_fp16")]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_53_cast_fp16 = softmax(axis = var_725, x = _inversed_attn_weights_51_cast_fp16)[name = string("attn_weights_53_cast_fp16")]; | |
| bool var_802_transpose_x_0 = const()[name = string("op_802_transpose_x_0"), val = bool(false)]; | |
| bool var_802_transpose_y_0 = const()[name = string("op_802_transpose_y_0"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 64, 5]> var_802_cast_fp16 = matmul(transpose_x = var_802_transpose_x_0, transpose_y = var_802_transpose_y_0, x = var_795_cast_fp16_0, y = attn_weights_53_cast_fp16)[name = string("op_802_cast_fp16")]; | |
| bool attn_weights_55_transpose_x_1 = const()[name = string("attn_weights_55_transpose_x_1"), val = bool(true)]; | |
| bool attn_weights_55_transpose_y_1 = const()[name = string("attn_weights_55_transpose_y_1"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_55_cast_fp16 = matmul(transpose_x = attn_weights_55_transpose_x_1, transpose_y = attn_weights_55_transpose_y_1, x = var_793_cast_fp16_1, y = var_791_cast_fp16_1)[name = string("attn_weights_55_cast_fp16")]; | |
| fp16 _inversed_attn_weights_57_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_57_y_0_to_fp16"), val = fp16(0x1p-3)]; | |
| tensor<fp16, [?, 8, 5, 5]> _inversed_attn_weights_57_cast_fp16 = mul(x = attn_weights_55_cast_fp16, y = _inversed_attn_weights_57_y_0_to_fp16)[name = string("_inversed_attn_weights_57_cast_fp16")]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_59_cast_fp16 = softmax(axis = var_725, x = _inversed_attn_weights_57_cast_fp16)[name = string("attn_weights_59_cast_fp16")]; | |
| bool attn_output_17_transpose_x_0 = const()[name = string("attn_output_17_transpose_x_0"), val = bool(false)]; | |
| bool attn_output_17_transpose_y_0 = const()[name = string("attn_output_17_transpose_y_0"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 64, 5]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = var_795_cast_fp16_1, y = attn_weights_59_cast_fp16)[name = string("attn_output_17_cast_fp16")]; | |
| bool attn_output_19_interleave_0 = const()[name = string("attn_output_19_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 16, 64, 5]> attn_output_19_cast_fp16 = concat(axis = var_720, interleave = attn_output_19_interleave_0, values = (var_802_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")]; | |
| tensor<int32, [4]> concat_19x = const()[name = string("concat_19x"), val = tensor<int32, [4]>([-1, 1024, 1, 5])]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_83_cast_fp16 = reshape(shape = concat_19x, x = attn_output_19_cast_fp16)[name = string("x_83_cast_fp16")]; | |
| string hidden_states_27_pad_type_0 = const()[name = string("hidden_states_27_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = string("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = string("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> hidden_states_27_dilations_0 = const()[name = string("hidden_states_27_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 hidden_states_27_groups_0 = const()[name = string("hidden_states_27_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 1024, 1, 1]> var_719_to_fp16 = const()[name = string("op_719_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122335360)))]; | |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_27_cast_fp16 = conv(dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = var_719_to_fp16, x = x_83_cast_fp16)[name = string("hidden_states_27_cast_fp16")]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_85_cast_fp16 = add(x = x_77_cast_fp16, y = hidden_states_27_cast_fp16)[name = string("x_85_cast_fp16")]; | |
| fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_821_cast_fp16 = mul(x = x_85_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_821_cast_fp16")]; | |
| bool x_87_interleave_0 = const()[name = string("x_87_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2048, 1, 5]> x_87_cast_fp16 = concat(axis = var_720, interleave = x_87_interleave_0, values = (x_85_cast_fp16, var_821_cast_fp16))[name = string("x_87_cast_fp16")]; | |
| tensor<int32, [1]> out_55_axes_0 = const()[name = string("out_55_axes_0"), val = tensor<int32, [1]>([1])]; | |
| fp16 var_831_to_fp16 = const()[name = string("op_831_to_fp16"), val = fp16(0x1.5p-17)]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_55_cast_fp16 = layer_norm(axes = out_55_axes_0, epsilon = var_831_to_fp16, x = x_87_cast_fp16)[name = string("out_55_cast_fp16")]; | |
| tensor<fp16, [1, 2048, 1, 1]> layer_encoder_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124432576)))]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_57_cast_fp16 = mul(x = out_55_cast_fp16, y = layer_encoder_layers_4_post_attention_layernorm_weight_to_fp16)[name = string("out_57_cast_fp16")]; | |
| tensor<int32, [2]> var_837_split_sizes_0 = const()[name = string("op_837_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; | |
| int32 var_837_axis_0 = const()[name = string("op_837_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_837_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_837_cast_fp16_1 = split(axis = var_837_axis_0, split_sizes = var_837_split_sizes_0, x = out_57_cast_fp16)[name = string("op_837_cast_fp16")]; | |
| string input_9_pad_type_0 = const()[name = string("input_9_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> input_9_strides_0 = const()[name = string("input_9_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> input_9_dilations_0 = const()[name = string("input_9_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 input_9_groups_0 = const()[name = string("input_9_groups_0"), val = int32(1)]; | |
| tensor<fp16, [4096, 1024, 1, 1]> var_706_to_fp16 = const()[name = string("op_706_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124436736)))]; | |
| tensor<fp16, [?, 4096, 1, 5]> input_9_cast_fp16 = conv(dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = var_706_to_fp16, x = var_837_cast_fp16_0)[name = string("input_9_cast_fp16")]; | |
| tensor<fp16, [?, 4096, 1, 5]> var_845_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_845_cast_fp16")]; | |
| string var_850_pad_type_0 = const()[name = string("op_850_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> var_850_strides_0 = const()[name = string("op_850_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> var_850_pad_0 = const()[name = string("op_850_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> var_850_dilations_0 = const()[name = string("op_850_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_850_groups_0 = const()[name = string("op_850_groups_0"), val = int32(1)]; | |
| tensor<fp16, [4096, 1024, 1, 1]> var_707_to_fp16 = const()[name = string("op_707_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132825408)))]; | |
| tensor<fp16, [?, 4096, 1, 5]> var_850_cast_fp16 = conv(dilations = var_850_dilations_0, groups = var_850_groups_0, pad = var_850_pad_0, pad_type = var_850_pad_type_0, strides = var_850_strides_0, weight = var_707_to_fp16, x = var_837_cast_fp16_0)[name = string("op_850_cast_fp16")]; | |
| tensor<fp16, [?, 4096, 1, 5]> x_93_cast_fp16 = mul(x = var_845_cast_fp16, y = var_850_cast_fp16)[name = string("x_93_cast_fp16")]; | |
| string hidden_states_29_pad_type_0 = const()[name = string("hidden_states_29_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> hidden_states_29_strides_0 = const()[name = string("hidden_states_29_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> hidden_states_29_pad_0 = const()[name = string("hidden_states_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> hidden_states_29_dilations_0 = const()[name = string("hidden_states_29_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 hidden_states_29_groups_0 = const()[name = string("hidden_states_29_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 4096, 1, 1]> var_708_to_fp16 = const()[name = string("op_708_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141214080)))]; | |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_29_cast_fp16 = conv(dilations = hidden_states_29_dilations_0, groups = hidden_states_29_groups_0, pad = hidden_states_29_pad_0, pad_type = hidden_states_29_pad_type_0, strides = hidden_states_29_strides_0, weight = var_708_to_fp16, x = x_93_cast_fp16)[name = string("hidden_states_29_cast_fp16")]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_95_cast_fp16 = add(x = x_85_cast_fp16, y = hidden_states_29_cast_fp16)[name = string("x_95_cast_fp16")]; | |
| int32 var_868 = const()[name = string("op_868"), val = int32(-2)]; | |
| int32 var_872 = const()[name = string("op_872"), val = int32(1)]; | |
| int32 var_877 = const()[name = string("op_877"), val = int32(2)]; | |
| fp16 const_21_promoted_to_fp16 = const()[name = string("const_21_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_882_cast_fp16 = mul(x = x_95_cast_fp16, y = const_21_promoted_to_fp16)[name = string("op_882_cast_fp16")]; | |
| bool x_97_interleave_0 = const()[name = string("x_97_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2048, 1, 5]> x_97_cast_fp16 = concat(axis = var_872, interleave = x_97_interleave_0, values = (x_95_cast_fp16, var_882_cast_fp16))[name = string("x_97_cast_fp16")]; | |
| tensor<int32, [1]> out_61_axes_0 = const()[name = string("out_61_axes_0"), val = tensor<int32, [1]>([1])]; | |
| fp16 var_892_to_fp16 = const()[name = string("op_892_to_fp16"), val = fp16(0x1.5p-17)]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_61_cast_fp16 = layer_norm(axes = out_61_axes_0, epsilon = var_892_to_fp16, x = x_97_cast_fp16)[name = string("out_61_cast_fp16")]; | |
| tensor<fp16, [1, 2048, 1, 1]> layer_encoder_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149602752)))]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_63_cast_fp16 = mul(x = out_61_cast_fp16, y = layer_encoder_layers_5_input_layernorm_weight_to_fp16)[name = string("out_63_cast_fp16")]; | |
| tensor<int32, [2]> var_898_split_sizes_0 = const()[name = string("op_898_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; | |
| int32 var_898_axis_0 = const()[name = string("op_898_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_898_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_898_cast_fp16_1 = split(axis = var_898_axis_0, split_sizes = var_898_split_sizes_0, x = out_63_cast_fp16)[name = string("op_898_cast_fp16")]; | |
| string query_states_31_pad_type_0 = const()[name = string("query_states_31_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> query_states_31_strides_0 = const()[name = string("query_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> query_states_31_pad_0 = const()[name = string("query_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> query_states_31_dilations_0 = const()[name = string("query_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 query_states_31_groups_0 = const()[name = string("query_states_31_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 1024, 1, 1]> var_863_to_fp16 = const()[name = string("op_863_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149606912)))]; | |
| tensor<fp16, [?, 1024, 1, 5]> query_states_31_cast_fp16 = conv(dilations = query_states_31_dilations_0, groups = query_states_31_groups_0, pad = query_states_31_pad_0, pad_type = query_states_31_pad_type_0, strides = query_states_31_strides_0, weight = var_863_to_fp16, x = var_898_cast_fp16_0)[name = string("query_states_31_cast_fp16")]; | |
| string key_states_31_pad_type_0 = const()[name = string("key_states_31_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> key_states_31_strides_0 = const()[name = string("key_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> key_states_31_pad_0 = const()[name = string("key_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> key_states_31_dilations_0 = const()[name = string("key_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 key_states_31_groups_0 = const()[name = string("key_states_31_groups_0"), val = int32(1)]; | |
| tensor<fp16, [128, 1024, 1, 1]> var_864_to_fp16 = const()[name = string("op_864_to_fp16"), val = tensor<fp16, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151704128)))]; | |
| tensor<fp16, [?, 128, 1, 5]> key_states_31_cast_fp16 = conv(dilations = key_states_31_dilations_0, groups = key_states_31_groups_0, pad = key_states_31_pad_0, pad_type = key_states_31_pad_type_0, strides = key_states_31_strides_0, weight = var_864_to_fp16, x = var_898_cast_fp16_0)[name = string("key_states_31_cast_fp16")]; | |
| string value_states_31_pad_type_0 = const()[name = string("value_states_31_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> value_states_31_strides_0 = const()[name = string("value_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> value_states_31_pad_0 = const()[name = string("value_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> value_states_31_dilations_0 = const()[name = string("value_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 value_states_31_groups_0 = const()[name = string("value_states_31_groups_0"), val = int32(1)]; | |
| tensor<fp16, [128, 1024, 1, 1]> var_865_to_fp16 = const()[name = string("op_865_to_fp16"), val = tensor<fp16, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151966336)))]; | |
| tensor<fp16, [?, 128, 1, 5]> value_states_31_cast_fp16 = conv(dilations = value_states_31_dilations_0, groups = value_states_31_groups_0, pad = value_states_31_pad_0, pad_type = value_states_31_pad_type_0, strides = value_states_31_strides_0, weight = var_865_to_fp16, x = var_898_cast_fp16_0)[name = string("value_states_31_cast_fp16")]; | |
| tensor<int32, [4]> concat_20x = const()[name = string("concat_20x"), val = tensor<int32, [4]>([-1, 16, 64, 5])]; | |
| tensor<fp16, [?, 16, 64, 5]> embed_21_cast_fp16 = reshape(shape = concat_20x, x = query_states_31_cast_fp16)[name = string("embed_21_cast_fp16")]; | |
| tensor<int32, [4]> concat_21x = const()[name = string("concat_21x"), val = tensor<int32, [4]>([-1, 2, 64, 5])]; | |
| tensor<fp16, [?, 2, 64, 5]> embed_23_cast_fp16 = reshape(shape = concat_21x, x = key_states_31_cast_fp16)[name = string("embed_23_cast_fp16")]; | |
| tensor<int32, [4]> concat_22x = const()[name = string("concat_22x"), val = tensor<int32, [4]>([-1, 2, 64, 5])]; | |
| tensor<fp16, [?, 2, 64, 5]> value_states_33_cast_fp16 = reshape(shape = concat_22x, x = value_states_31_cast_fp16)[name = string("value_states_33_cast_fp16")]; | |
| tensor<fp16, [?, 16, 64, 5]> var_924_cast_fp16 = mul(x = embed_21_cast_fp16, y = cos_to_fp16)[name = string("op_924_cast_fp16")]; | |
| tensor<int32, [2]> var_925_split_sizes_0 = const()[name = string("op_925_split_sizes_0"), val = tensor<int32, [2]>([32, 32])]; | |
| int32 var_925_axis_0 = const()[name = string("op_925_axis_0"), val = int32(-2)]; | |
| tensor<fp16, [?, 16, 32, 5]> var_925_cast_fp16_0, tensor<fp16, [?, 16, 32, 5]> var_925_cast_fp16_1 = split(axis = var_925_axis_0, split_sizes = var_925_split_sizes_0, x = embed_21_cast_fp16)[name = string("op_925_cast_fp16")]; | |
| fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 16, 32, 5]> var_927_cast_fp16 = mul(x = var_925_cast_fp16_1, y = const_22_promoted_to_fp16)[name = string("op_927_cast_fp16")]; | |
| bool var_929_interleave_0 = const()[name = string("op_929_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 16, 64, 5]> var_929_cast_fp16 = concat(axis = var_868, interleave = var_929_interleave_0, values = (var_927_cast_fp16, var_925_cast_fp16_0))[name = string("op_929_cast_fp16")]; | |
| tensor<fp16, [?, 16, 64, 5]> var_930_cast_fp16 = mul(x = var_929_cast_fp16, y = sin_to_fp16)[name = string("op_930_cast_fp16")]; | |
| tensor<fp16, [?, 16, 64, 5]> query_states_33_cast_fp16 = add(x = var_924_cast_fp16, y = var_930_cast_fp16)[name = string("query_states_33_cast_fp16")]; | |
| tensor<fp16, [?, 2, 64, 5]> var_932_cast_fp16 = mul(x = embed_23_cast_fp16, y = cos_to_fp16)[name = string("op_932_cast_fp16")]; | |
| tensor<int32, [2]> var_933_split_sizes_0 = const()[name = string("op_933_split_sizes_0"), val = tensor<int32, [2]>([32, 32])]; | |
| int32 var_933_axis_0 = const()[name = string("op_933_axis_0"), val = int32(-2)]; | |
| tensor<fp16, [?, 2, 32, 5]> var_933_cast_fp16_0, tensor<fp16, [?, 2, 32, 5]> var_933_cast_fp16_1 = split(axis = var_933_axis_0, split_sizes = var_933_split_sizes_0, x = embed_23_cast_fp16)[name = string("op_933_cast_fp16")]; | |
| fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 2, 32, 5]> var_935_cast_fp16 = mul(x = var_933_cast_fp16_1, y = const_23_promoted_to_fp16)[name = string("op_935_cast_fp16")]; | |
| bool var_937_interleave_0 = const()[name = string("op_937_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2, 64, 5]> var_937_cast_fp16 = concat(axis = var_868, interleave = var_937_interleave_0, values = (var_935_cast_fp16, var_933_cast_fp16_0))[name = string("op_937_cast_fp16")]; | |
| tensor<fp16, [?, 2, 64, 5]> var_938_cast_fp16 = mul(x = var_937_cast_fp16, y = sin_to_fp16)[name = string("op_938_cast_fp16")]; | |
| tensor<fp16, [?, 2, 64, 5]> key_states_33_cast_fp16 = add(x = var_932_cast_fp16, y = var_938_cast_fp16)[name = string("key_states_33_cast_fp16")]; | |
| tensor<int32, [2]> var_943_split_sizes_0 = const()[name = string("op_943_split_sizes_0"), val = tensor<int32, [2]>([8, 8])]; | |
| int32 var_943_axis_0 = const()[name = string("op_943_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 8, 64, 5]> var_943_cast_fp16_0, tensor<fp16, [?, 8, 64, 5]> var_943_cast_fp16_1 = split(axis = var_943_axis_0, split_sizes = var_943_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_943_cast_fp16")]; | |
| tensor<int32, [2]> var_945_split_sizes_0 = const()[name = string("op_945_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_945_axis_0 = const()[name = string("op_945_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1, 64, 5]> var_945_cast_fp16_0, tensor<fp16, [?, 1, 64, 5]> var_945_cast_fp16_1 = split(axis = var_945_axis_0, split_sizes = var_945_split_sizes_0, x = key_states_33_cast_fp16)[name = string("op_945_cast_fp16")]; | |
| tensor<int32, [2]> var_947_split_sizes_0 = const()[name = string("op_947_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_947_axis_0 = const()[name = string("op_947_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1, 64, 5]> var_947_cast_fp16_0, tensor<fp16, [?, 1, 64, 5]> var_947_cast_fp16_1 = split(axis = var_947_axis_0, split_sizes = var_947_split_sizes_0, x = value_states_33_cast_fp16)[name = string("op_947_cast_fp16")]; | |
| bool attn_weights_61_transpose_x_1 = const()[name = string("attn_weights_61_transpose_x_1"), val = bool(true)]; | |
| bool attn_weights_61_transpose_y_1 = const()[name = string("attn_weights_61_transpose_y_1"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_61_cast_fp16 = matmul(transpose_x = attn_weights_61_transpose_x_1, transpose_y = attn_weights_61_transpose_y_1, x = var_945_cast_fp16_0, y = var_943_cast_fp16_0)[name = string("attn_weights_61_cast_fp16")]; | |
| fp16 _inversed_attn_weights_63_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_63_y_0_to_fp16"), val = fp16(0x1p-3)]; | |
| tensor<fp16, [?, 8, 5, 5]> _inversed_attn_weights_63_cast_fp16 = mul(x = attn_weights_61_cast_fp16, y = _inversed_attn_weights_63_y_0_to_fp16)[name = string("_inversed_attn_weights_63_cast_fp16")]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_65_cast_fp16 = softmax(axis = var_877, x = _inversed_attn_weights_63_cast_fp16)[name = string("attn_weights_65_cast_fp16")]; | |
| bool var_954_transpose_x_0 = const()[name = string("op_954_transpose_x_0"), val = bool(false)]; | |
| bool var_954_transpose_y_0 = const()[name = string("op_954_transpose_y_0"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 64, 5]> var_954_cast_fp16 = matmul(transpose_x = var_954_transpose_x_0, transpose_y = var_954_transpose_y_0, x = var_947_cast_fp16_0, y = attn_weights_65_cast_fp16)[name = string("op_954_cast_fp16")]; | |
| bool attn_weights_67_transpose_x_1 = const()[name = string("attn_weights_67_transpose_x_1"), val = bool(true)]; | |
| bool attn_weights_67_transpose_y_1 = const()[name = string("attn_weights_67_transpose_y_1"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_67_cast_fp16 = matmul(transpose_x = attn_weights_67_transpose_x_1, transpose_y = attn_weights_67_transpose_y_1, x = var_945_cast_fp16_1, y = var_943_cast_fp16_1)[name = string("attn_weights_67_cast_fp16")]; | |
| fp16 _inversed_attn_weights_69_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_69_y_0_to_fp16"), val = fp16(0x1p-3)]; | |
| tensor<fp16, [?, 8, 5, 5]> _inversed_attn_weights_69_cast_fp16 = mul(x = attn_weights_67_cast_fp16, y = _inversed_attn_weights_69_y_0_to_fp16)[name = string("_inversed_attn_weights_69_cast_fp16")]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_71_cast_fp16 = softmax(axis = var_877, x = _inversed_attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")]; | |
| bool attn_output_21_transpose_x_0 = const()[name = string("attn_output_21_transpose_x_0"), val = bool(false)]; | |
| bool attn_output_21_transpose_y_0 = const()[name = string("attn_output_21_transpose_y_0"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 64, 5]> attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = var_947_cast_fp16_1, y = attn_weights_71_cast_fp16)[name = string("attn_output_21_cast_fp16")]; | |
| bool attn_output_23_interleave_0 = const()[name = string("attn_output_23_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 16, 64, 5]> attn_output_23_cast_fp16 = concat(axis = var_872, interleave = attn_output_23_interleave_0, values = (var_954_cast_fp16, attn_output_21_cast_fp16))[name = string("attn_output_23_cast_fp16")]; | |
| tensor<int32, [4]> concat_23x = const()[name = string("concat_23x"), val = tensor<int32, [4]>([-1, 1024, 1, 5])]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_101_cast_fp16 = reshape(shape = concat_23x, x = attn_output_23_cast_fp16)[name = string("x_101_cast_fp16")]; | |
| string hidden_states_33_pad_type_0 = const()[name = string("hidden_states_33_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> hidden_states_33_strides_0 = const()[name = string("hidden_states_33_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> hidden_states_33_pad_0 = const()[name = string("hidden_states_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> hidden_states_33_dilations_0 = const()[name = string("hidden_states_33_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 hidden_states_33_groups_0 = const()[name = string("hidden_states_33_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 1024, 1, 1]> var_871_to_fp16 = const()[name = string("op_871_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152228544)))]; | |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_33_cast_fp16 = conv(dilations = hidden_states_33_dilations_0, groups = hidden_states_33_groups_0, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = hidden_states_33_strides_0, weight = var_871_to_fp16, x = x_101_cast_fp16)[name = string("hidden_states_33_cast_fp16")]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_103_cast_fp16 = add(x = x_95_cast_fp16, y = hidden_states_33_cast_fp16)[name = string("x_103_cast_fp16")]; | |
| fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_973_cast_fp16 = mul(x = x_103_cast_fp16, y = const_24_promoted_to_fp16)[name = string("op_973_cast_fp16")]; | |
| bool x_105_interleave_0 = const()[name = string("x_105_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2048, 1, 5]> x_105_cast_fp16 = concat(axis = var_872, interleave = x_105_interleave_0, values = (x_103_cast_fp16, var_973_cast_fp16))[name = string("x_105_cast_fp16")]; | |
| tensor<int32, [1]> out_67_axes_0 = const()[name = string("out_67_axes_0"), val = tensor<int32, [1]>([1])]; | |
| fp16 var_983_to_fp16 = const()[name = string("op_983_to_fp16"), val = fp16(0x1.5p-17)]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_67_cast_fp16 = layer_norm(axes = out_67_axes_0, epsilon = var_983_to_fp16, x = x_105_cast_fp16)[name = string("out_67_cast_fp16")]; | |
| tensor<fp16, [1, 2048, 1, 1]> layer_encoder_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154325760)))]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_69_cast_fp16 = mul(x = out_67_cast_fp16, y = layer_encoder_layers_5_post_attention_layernorm_weight_to_fp16)[name = string("out_69_cast_fp16")]; | |
| tensor<int32, [2]> var_989_split_sizes_0 = const()[name = string("op_989_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; | |
| int32 var_989_axis_0 = const()[name = string("op_989_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_989_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_989_cast_fp16_1 = split(axis = var_989_axis_0, split_sizes = var_989_split_sizes_0, x = out_69_cast_fp16)[name = string("op_989_cast_fp16")]; | |
| string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)]; | |
| tensor<fp16, [4096, 1024, 1, 1]> var_858_to_fp16 = const()[name = string("op_858_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154329920)))]; | |
| tensor<fp16, [?, 4096, 1, 5]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = var_858_to_fp16, x = var_989_cast_fp16_0)[name = string("input_11_cast_fp16")]; | |
| tensor<fp16, [?, 4096, 1, 5]> var_997_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_997_cast_fp16")]; | |
| string var_1002_pad_type_0 = const()[name = string("op_1002_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> var_1002_strides_0 = const()[name = string("op_1002_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> var_1002_pad_0 = const()[name = string("op_1002_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> var_1002_dilations_0 = const()[name = string("op_1002_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_1002_groups_0 = const()[name = string("op_1002_groups_0"), val = int32(1)]; | |
| tensor<fp16, [4096, 1024, 1, 1]> var_859_to_fp16 = const()[name = string("op_859_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162718592)))]; | |
| tensor<fp16, [?, 4096, 1, 5]> var_1002_cast_fp16 = conv(dilations = var_1002_dilations_0, groups = var_1002_groups_0, pad = var_1002_pad_0, pad_type = var_1002_pad_type_0, strides = var_1002_strides_0, weight = var_859_to_fp16, x = var_989_cast_fp16_0)[name = string("op_1002_cast_fp16")]; | |
| tensor<fp16, [?, 4096, 1, 5]> x_111_cast_fp16 = mul(x = var_997_cast_fp16, y = var_1002_cast_fp16)[name = string("x_111_cast_fp16")]; | |
| string hidden_states_35_pad_type_0 = const()[name = string("hidden_states_35_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> hidden_states_35_strides_0 = const()[name = string("hidden_states_35_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> hidden_states_35_pad_0 = const()[name = string("hidden_states_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> hidden_states_35_dilations_0 = const()[name = string("hidden_states_35_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 hidden_states_35_groups_0 = const()[name = string("hidden_states_35_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 4096, 1, 1]> var_860_to_fp16 = const()[name = string("op_860_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171107264)))]; | |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_35_cast_fp16 = conv(dilations = hidden_states_35_dilations_0, groups = hidden_states_35_groups_0, pad = hidden_states_35_pad_0, pad_type = hidden_states_35_pad_type_0, strides = hidden_states_35_strides_0, weight = var_860_to_fp16, x = x_111_cast_fp16)[name = string("hidden_states_35_cast_fp16")]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_113_cast_fp16 = add(x = x_103_cast_fp16, y = hidden_states_35_cast_fp16)[name = string("x_113_cast_fp16")]; | |
| int32 var_1020 = const()[name = string("op_1020"), val = int32(-2)]; | |
| int32 var_1024 = const()[name = string("op_1024"), val = int32(1)]; | |
| int32 var_1029 = const()[name = string("op_1029"), val = int32(2)]; | |
| fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_1034_cast_fp16 = mul(x = x_113_cast_fp16, y = const_25_promoted_to_fp16)[name = string("op_1034_cast_fp16")]; | |
| bool x_115_interleave_0 = const()[name = string("x_115_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2048, 1, 5]> x_115_cast_fp16 = concat(axis = var_1024, interleave = x_115_interleave_0, values = (x_113_cast_fp16, var_1034_cast_fp16))[name = string("x_115_cast_fp16")]; | |
| tensor<int32, [1]> out_73_axes_0 = const()[name = string("out_73_axes_0"), val = tensor<int32, [1]>([1])]; | |
| fp16 var_1044_to_fp16 = const()[name = string("op_1044_to_fp16"), val = fp16(0x1.5p-17)]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_73_cast_fp16 = layer_norm(axes = out_73_axes_0, epsilon = var_1044_to_fp16, x = x_115_cast_fp16)[name = string("out_73_cast_fp16")]; | |
| tensor<fp16, [1, 2048, 1, 1]> layer_encoder_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179495936)))]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_75_cast_fp16 = mul(x = out_73_cast_fp16, y = layer_encoder_layers_6_input_layernorm_weight_to_fp16)[name = string("out_75_cast_fp16")]; | |
| tensor<int32, [2]> var_1050_split_sizes_0 = const()[name = string("op_1050_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; | |
| int32 var_1050_axis_0 = const()[name = string("op_1050_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_1050_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_1050_cast_fp16_1 = split(axis = var_1050_axis_0, split_sizes = var_1050_split_sizes_0, x = out_75_cast_fp16)[name = string("op_1050_cast_fp16")]; | |
| string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> query_states_37_pad_0 = const()[name = string("query_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> query_states_37_dilations_0 = const()[name = string("query_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 query_states_37_groups_0 = const()[name = string("query_states_37_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 1024, 1, 1]> var_1015_to_fp16 = const()[name = string("op_1015_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179500096)))]; | |
| tensor<fp16, [?, 1024, 1, 5]> query_states_37_cast_fp16 = conv(dilations = query_states_37_dilations_0, groups = query_states_37_groups_0, pad = query_states_37_pad_0, pad_type = query_states_37_pad_type_0, strides = query_states_37_strides_0, weight = var_1015_to_fp16, x = var_1050_cast_fp16_0)[name = string("query_states_37_cast_fp16")]; | |
| string key_states_37_pad_type_0 = const()[name = string("key_states_37_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> key_states_37_strides_0 = const()[name = string("key_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> key_states_37_pad_0 = const()[name = string("key_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> key_states_37_dilations_0 = const()[name = string("key_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 key_states_37_groups_0 = const()[name = string("key_states_37_groups_0"), val = int32(1)]; | |
| tensor<fp16, [128, 1024, 1, 1]> var_1016_to_fp16 = const()[name = string("op_1016_to_fp16"), val = tensor<fp16, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181597312)))]; | |
| tensor<fp16, [?, 128, 1, 5]> key_states_37_cast_fp16 = conv(dilations = key_states_37_dilations_0, groups = key_states_37_groups_0, pad = key_states_37_pad_0, pad_type = key_states_37_pad_type_0, strides = key_states_37_strides_0, weight = var_1016_to_fp16, x = var_1050_cast_fp16_0)[name = string("key_states_37_cast_fp16")]; | |
| string value_states_37_pad_type_0 = const()[name = string("value_states_37_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> value_states_37_strides_0 = const()[name = string("value_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> value_states_37_pad_0 = const()[name = string("value_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> value_states_37_dilations_0 = const()[name = string("value_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 value_states_37_groups_0 = const()[name = string("value_states_37_groups_0"), val = int32(1)]; | |
| tensor<fp16, [128, 1024, 1, 1]> var_1017_to_fp16 = const()[name = string("op_1017_to_fp16"), val = tensor<fp16, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181859520)))]; | |
| tensor<fp16, [?, 128, 1, 5]> value_states_37_cast_fp16 = conv(dilations = value_states_37_dilations_0, groups = value_states_37_groups_0, pad = value_states_37_pad_0, pad_type = value_states_37_pad_type_0, strides = value_states_37_strides_0, weight = var_1017_to_fp16, x = var_1050_cast_fp16_0)[name = string("value_states_37_cast_fp16")]; | |
| tensor<int32, [4]> concat_24x = const()[name = string("concat_24x"), val = tensor<int32, [4]>([-1, 16, 64, 5])]; | |
| tensor<fp16, [?, 16, 64, 5]> embed_25_cast_fp16 = reshape(shape = concat_24x, x = query_states_37_cast_fp16)[name = string("embed_25_cast_fp16")]; | |
| tensor<int32, [4]> concat_25x = const()[name = string("concat_25x"), val = tensor<int32, [4]>([-1, 2, 64, 5])]; | |
| tensor<fp16, [?, 2, 64, 5]> embed_27_cast_fp16 = reshape(shape = concat_25x, x = key_states_37_cast_fp16)[name = string("embed_27_cast_fp16")]; | |
| tensor<int32, [4]> concat_26x = const()[name = string("concat_26x"), val = tensor<int32, [4]>([-1, 2, 64, 5])]; | |
| tensor<fp16, [?, 2, 64, 5]> value_states_39_cast_fp16 = reshape(shape = concat_26x, x = value_states_37_cast_fp16)[name = string("value_states_39_cast_fp16")]; | |
| tensor<fp16, [?, 16, 64, 5]> var_1076_cast_fp16 = mul(x = embed_25_cast_fp16, y = cos_to_fp16)[name = string("op_1076_cast_fp16")]; | |
| tensor<int32, [2]> var_1077_split_sizes_0 = const()[name = string("op_1077_split_sizes_0"), val = tensor<int32, [2]>([32, 32])]; | |
| int32 var_1077_axis_0 = const()[name = string("op_1077_axis_0"), val = int32(-2)]; | |
| tensor<fp16, [?, 16, 32, 5]> var_1077_cast_fp16_0, tensor<fp16, [?, 16, 32, 5]> var_1077_cast_fp16_1 = split(axis = var_1077_axis_0, split_sizes = var_1077_split_sizes_0, x = embed_25_cast_fp16)[name = string("op_1077_cast_fp16")]; | |
| fp16 const_26_promoted_to_fp16 = const()[name = string("const_26_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 16, 32, 5]> var_1079_cast_fp16 = mul(x = var_1077_cast_fp16_1, y = const_26_promoted_to_fp16)[name = string("op_1079_cast_fp16")]; | |
| bool var_1081_interleave_0 = const()[name = string("op_1081_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 16, 64, 5]> var_1081_cast_fp16 = concat(axis = var_1020, interleave = var_1081_interleave_0, values = (var_1079_cast_fp16, var_1077_cast_fp16_0))[name = string("op_1081_cast_fp16")]; | |
| tensor<fp16, [?, 16, 64, 5]> var_1082_cast_fp16 = mul(x = var_1081_cast_fp16, y = sin_to_fp16)[name = string("op_1082_cast_fp16")]; | |
| tensor<fp16, [?, 16, 64, 5]> query_states_39_cast_fp16 = add(x = var_1076_cast_fp16, y = var_1082_cast_fp16)[name = string("query_states_39_cast_fp16")]; | |
| tensor<fp16, [?, 2, 64, 5]> var_1084_cast_fp16 = mul(x = embed_27_cast_fp16, y = cos_to_fp16)[name = string("op_1084_cast_fp16")]; | |
| tensor<int32, [2]> var_1085_split_sizes_0 = const()[name = string("op_1085_split_sizes_0"), val = tensor<int32, [2]>([32, 32])]; | |
| int32 var_1085_axis_0 = const()[name = string("op_1085_axis_0"), val = int32(-2)]; | |
| tensor<fp16, [?, 2, 32, 5]> var_1085_cast_fp16_0, tensor<fp16, [?, 2, 32, 5]> var_1085_cast_fp16_1 = split(axis = var_1085_axis_0, split_sizes = var_1085_split_sizes_0, x = embed_27_cast_fp16)[name = string("op_1085_cast_fp16")]; | |
| fp16 const_27_promoted_to_fp16 = const()[name = string("const_27_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 2, 32, 5]> var_1087_cast_fp16 = mul(x = var_1085_cast_fp16_1, y = const_27_promoted_to_fp16)[name = string("op_1087_cast_fp16")]; | |
| bool var_1089_interleave_0 = const()[name = string("op_1089_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2, 64, 5]> var_1089_cast_fp16 = concat(axis = var_1020, interleave = var_1089_interleave_0, values = (var_1087_cast_fp16, var_1085_cast_fp16_0))[name = string("op_1089_cast_fp16")]; | |
| tensor<fp16, [?, 2, 64, 5]> var_1090_cast_fp16 = mul(x = var_1089_cast_fp16, y = sin_to_fp16)[name = string("op_1090_cast_fp16")]; | |
| tensor<fp16, [?, 2, 64, 5]> key_states_39_cast_fp16 = add(x = var_1084_cast_fp16, y = var_1090_cast_fp16)[name = string("key_states_39_cast_fp16")]; | |
| tensor<int32, [2]> var_1095_split_sizes_0 = const()[name = string("op_1095_split_sizes_0"), val = tensor<int32, [2]>([8, 8])]; | |
| int32 var_1095_axis_0 = const()[name = string("op_1095_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 8, 64, 5]> var_1095_cast_fp16_0, tensor<fp16, [?, 8, 64, 5]> var_1095_cast_fp16_1 = split(axis = var_1095_axis_0, split_sizes = var_1095_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_1095_cast_fp16")]; | |
| tensor<int32, [2]> var_1097_split_sizes_0 = const()[name = string("op_1097_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_1097_axis_0 = const()[name = string("op_1097_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1, 64, 5]> var_1097_cast_fp16_0, tensor<fp16, [?, 1, 64, 5]> var_1097_cast_fp16_1 = split(axis = var_1097_axis_0, split_sizes = var_1097_split_sizes_0, x = key_states_39_cast_fp16)[name = string("op_1097_cast_fp16")]; | |
| tensor<int32, [2]> var_1099_split_sizes_0 = const()[name = string("op_1099_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_1099_axis_0 = const()[name = string("op_1099_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1, 64, 5]> var_1099_cast_fp16_0, tensor<fp16, [?, 1, 64, 5]> var_1099_cast_fp16_1 = split(axis = var_1099_axis_0, split_sizes = var_1099_split_sizes_0, x = value_states_39_cast_fp16)[name = string("op_1099_cast_fp16")]; | |
| bool attn_weights_73_transpose_x_1 = const()[name = string("attn_weights_73_transpose_x_1"), val = bool(true)]; | |
| bool attn_weights_73_transpose_y_1 = const()[name = string("attn_weights_73_transpose_y_1"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_73_cast_fp16 = matmul(transpose_x = attn_weights_73_transpose_x_1, transpose_y = attn_weights_73_transpose_y_1, x = var_1097_cast_fp16_0, y = var_1095_cast_fp16_0)[name = string("attn_weights_73_cast_fp16")]; | |
| fp16 _inversed_attn_weights_75_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_75_y_0_to_fp16"), val = fp16(0x1p-3)]; | |
| tensor<fp16, [?, 8, 5, 5]> _inversed_attn_weights_75_cast_fp16 = mul(x = attn_weights_73_cast_fp16, y = _inversed_attn_weights_75_y_0_to_fp16)[name = string("_inversed_attn_weights_75_cast_fp16")]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_77_cast_fp16 = softmax(axis = var_1029, x = _inversed_attn_weights_75_cast_fp16)[name = string("attn_weights_77_cast_fp16")]; | |
| bool var_1106_transpose_x_0 = const()[name = string("op_1106_transpose_x_0"), val = bool(false)]; | |
| bool var_1106_transpose_y_0 = const()[name = string("op_1106_transpose_y_0"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 64, 5]> var_1106_cast_fp16 = matmul(transpose_x = var_1106_transpose_x_0, transpose_y = var_1106_transpose_y_0, x = var_1099_cast_fp16_0, y = attn_weights_77_cast_fp16)[name = string("op_1106_cast_fp16")]; | |
| bool attn_weights_79_transpose_x_1 = const()[name = string("attn_weights_79_transpose_x_1"), val = bool(true)]; | |
| bool attn_weights_79_transpose_y_1 = const()[name = string("attn_weights_79_transpose_y_1"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_79_cast_fp16 = matmul(transpose_x = attn_weights_79_transpose_x_1, transpose_y = attn_weights_79_transpose_y_1, x = var_1097_cast_fp16_1, y = var_1095_cast_fp16_1)[name = string("attn_weights_79_cast_fp16")]; | |
| fp16 _inversed_attn_weights_81_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_81_y_0_to_fp16"), val = fp16(0x1p-3)]; | |
| tensor<fp16, [?, 8, 5, 5]> _inversed_attn_weights_81_cast_fp16 = mul(x = attn_weights_79_cast_fp16, y = _inversed_attn_weights_81_y_0_to_fp16)[name = string("_inversed_attn_weights_81_cast_fp16")]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_83_cast_fp16 = softmax(axis = var_1029, x = _inversed_attn_weights_81_cast_fp16)[name = string("attn_weights_83_cast_fp16")]; | |
| bool attn_output_25_transpose_x_0 = const()[name = string("attn_output_25_transpose_x_0"), val = bool(false)]; | |
| bool attn_output_25_transpose_y_0 = const()[name = string("attn_output_25_transpose_y_0"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 64, 5]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = var_1099_cast_fp16_1, y = attn_weights_83_cast_fp16)[name = string("attn_output_25_cast_fp16")]; | |
| bool attn_output_27_interleave_0 = const()[name = string("attn_output_27_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 16, 64, 5]> attn_output_27_cast_fp16 = concat(axis = var_1024, interleave = attn_output_27_interleave_0, values = (var_1106_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")]; | |
| tensor<int32, [4]> concat_27x = const()[name = string("concat_27x"), val = tensor<int32, [4]>([-1, 1024, 1, 5])]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_119_cast_fp16 = reshape(shape = concat_27x, x = attn_output_27_cast_fp16)[name = string("x_119_cast_fp16")]; | |
| string hidden_states_39_pad_type_0 = const()[name = string("hidden_states_39_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> hidden_states_39_strides_0 = const()[name = string("hidden_states_39_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> hidden_states_39_pad_0 = const()[name = string("hidden_states_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> hidden_states_39_dilations_0 = const()[name = string("hidden_states_39_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 hidden_states_39_groups_0 = const()[name = string("hidden_states_39_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 1024, 1, 1]> var_1023_to_fp16 = const()[name = string("op_1023_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182121728)))]; | |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_39_cast_fp16 = conv(dilations = hidden_states_39_dilations_0, groups = hidden_states_39_groups_0, pad = hidden_states_39_pad_0, pad_type = hidden_states_39_pad_type_0, strides = hidden_states_39_strides_0, weight = var_1023_to_fp16, x = x_119_cast_fp16)[name = string("hidden_states_39_cast_fp16")]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_121_cast_fp16 = add(x = x_113_cast_fp16, y = hidden_states_39_cast_fp16)[name = string("x_121_cast_fp16")]; | |
| fp16 const_28_promoted_to_fp16 = const()[name = string("const_28_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_1125_cast_fp16 = mul(x = x_121_cast_fp16, y = const_28_promoted_to_fp16)[name = string("op_1125_cast_fp16")]; | |
| bool x_123_interleave_0 = const()[name = string("x_123_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2048, 1, 5]> x_123_cast_fp16 = concat(axis = var_1024, interleave = x_123_interleave_0, values = (x_121_cast_fp16, var_1125_cast_fp16))[name = string("x_123_cast_fp16")]; | |
| tensor<int32, [1]> out_79_axes_0 = const()[name = string("out_79_axes_0"), val = tensor<int32, [1]>([1])]; | |
| fp16 var_1135_to_fp16 = const()[name = string("op_1135_to_fp16"), val = fp16(0x1.5p-17)]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_79_cast_fp16 = layer_norm(axes = out_79_axes_0, epsilon = var_1135_to_fp16, x = x_123_cast_fp16)[name = string("out_79_cast_fp16")]; | |
| tensor<fp16, [1, 2048, 1, 1]> layer_encoder_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184218944)))]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_81_cast_fp16 = mul(x = out_79_cast_fp16, y = layer_encoder_layers_6_post_attention_layernorm_weight_to_fp16)[name = string("out_81_cast_fp16")]; | |
| tensor<int32, [2]> var_1141_split_sizes_0 = const()[name = string("op_1141_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; | |
| int32 var_1141_axis_0 = const()[name = string("op_1141_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_1141_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_1141_cast_fp16_1 = split(axis = var_1141_axis_0, split_sizes = var_1141_split_sizes_0, x = out_81_cast_fp16)[name = string("op_1141_cast_fp16")]; | |
| string input_13_pad_type_0 = const()[name = string("input_13_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> input_13_strides_0 = const()[name = string("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> input_13_pad_0 = const()[name = string("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> input_13_dilations_0 = const()[name = string("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 input_13_groups_0 = const()[name = string("input_13_groups_0"), val = int32(1)]; | |
| tensor<fp16, [4096, 1024, 1, 1]> var_1010_to_fp16 = const()[name = string("op_1010_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184223104)))]; | |
| tensor<fp16, [?, 4096, 1, 5]> input_13_cast_fp16 = conv(dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = var_1010_to_fp16, x = var_1141_cast_fp16_0)[name = string("input_13_cast_fp16")]; | |
| tensor<fp16, [?, 4096, 1, 5]> var_1149_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_1149_cast_fp16")]; | |
| string var_1154_pad_type_0 = const()[name = string("op_1154_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> var_1154_strides_0 = const()[name = string("op_1154_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> var_1154_pad_0 = const()[name = string("op_1154_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> var_1154_dilations_0 = const()[name = string("op_1154_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_1154_groups_0 = const()[name = string("op_1154_groups_0"), val = int32(1)]; | |
| tensor<fp16, [4096, 1024, 1, 1]> var_1011_to_fp16 = const()[name = string("op_1011_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192611776)))]; | |
| tensor<fp16, [?, 4096, 1, 5]> var_1154_cast_fp16 = conv(dilations = var_1154_dilations_0, groups = var_1154_groups_0, pad = var_1154_pad_0, pad_type = var_1154_pad_type_0, strides = var_1154_strides_0, weight = var_1011_to_fp16, x = var_1141_cast_fp16_0)[name = string("op_1154_cast_fp16")]; | |
| tensor<fp16, [?, 4096, 1, 5]> x_129_cast_fp16 = mul(x = var_1149_cast_fp16, y = var_1154_cast_fp16)[name = string("x_129_cast_fp16")]; | |
| string hidden_states_41_pad_type_0 = const()[name = string("hidden_states_41_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> hidden_states_41_strides_0 = const()[name = string("hidden_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> hidden_states_41_pad_0 = const()[name = string("hidden_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> hidden_states_41_dilations_0 = const()[name = string("hidden_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 hidden_states_41_groups_0 = const()[name = string("hidden_states_41_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 4096, 1, 1]> var_1012_to_fp16 = const()[name = string("op_1012_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201000448)))]; | |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_41_cast_fp16 = conv(dilations = hidden_states_41_dilations_0, groups = hidden_states_41_groups_0, pad = hidden_states_41_pad_0, pad_type = hidden_states_41_pad_type_0, strides = hidden_states_41_strides_0, weight = var_1012_to_fp16, x = x_129_cast_fp16)[name = string("hidden_states_41_cast_fp16")]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_131_cast_fp16 = add(x = x_121_cast_fp16, y = hidden_states_41_cast_fp16)[name = string("x_131_cast_fp16")]; | |
| int32 var_1172 = const()[name = string("op_1172"), val = int32(-2)]; | |
| int32 var_1176 = const()[name = string("op_1176"), val = int32(1)]; | |
| int32 var_1181 = const()[name = string("op_1181"), val = int32(2)]; | |
| fp16 const_29_promoted_to_fp16 = const()[name = string("const_29_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_1186_cast_fp16 = mul(x = x_131_cast_fp16, y = const_29_promoted_to_fp16)[name = string("op_1186_cast_fp16")]; | |
| bool x_133_interleave_0 = const()[name = string("x_133_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2048, 1, 5]> x_133_cast_fp16 = concat(axis = var_1176, interleave = x_133_interleave_0, values = (x_131_cast_fp16, var_1186_cast_fp16))[name = string("x_133_cast_fp16")]; | |
| tensor<int32, [1]> out_85_axes_0 = const()[name = string("out_85_axes_0"), val = tensor<int32, [1]>([1])]; | |
| fp16 var_1196_to_fp16 = const()[name = string("op_1196_to_fp16"), val = fp16(0x1.5p-17)]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_85_cast_fp16 = layer_norm(axes = out_85_axes_0, epsilon = var_1196_to_fp16, x = x_133_cast_fp16)[name = string("out_85_cast_fp16")]; | |
| tensor<fp16, [1, 2048, 1, 1]> layer_encoder_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209389120)))]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_87_cast_fp16 = mul(x = out_85_cast_fp16, y = layer_encoder_layers_7_input_layernorm_weight_to_fp16)[name = string("out_87_cast_fp16")]; | |
| tensor<int32, [2]> var_1202_split_sizes_0 = const()[name = string("op_1202_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; | |
| int32 var_1202_axis_0 = const()[name = string("op_1202_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_1202_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_1202_cast_fp16_1 = split(axis = var_1202_axis_0, split_sizes = var_1202_split_sizes_0, x = out_87_cast_fp16)[name = string("op_1202_cast_fp16")]; | |
| string query_states_43_pad_type_0 = const()[name = string("query_states_43_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> query_states_43_strides_0 = const()[name = string("query_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> query_states_43_pad_0 = const()[name = string("query_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> query_states_43_dilations_0 = const()[name = string("query_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 query_states_43_groups_0 = const()[name = string("query_states_43_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 1024, 1, 1]> var_1167_to_fp16 = const()[name = string("op_1167_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209393280)))]; | |
| tensor<fp16, [?, 1024, 1, 5]> query_states_43_cast_fp16 = conv(dilations = query_states_43_dilations_0, groups = query_states_43_groups_0, pad = query_states_43_pad_0, pad_type = query_states_43_pad_type_0, strides = query_states_43_strides_0, weight = var_1167_to_fp16, x = var_1202_cast_fp16_0)[name = string("query_states_43_cast_fp16")]; | |
| string key_states_43_pad_type_0 = const()[name = string("key_states_43_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> key_states_43_strides_0 = const()[name = string("key_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> key_states_43_pad_0 = const()[name = string("key_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> key_states_43_dilations_0 = const()[name = string("key_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 key_states_43_groups_0 = const()[name = string("key_states_43_groups_0"), val = int32(1)]; | |
| tensor<fp16, [128, 1024, 1, 1]> var_1168_to_fp16 = const()[name = string("op_1168_to_fp16"), val = tensor<fp16, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211490496)))]; | |
| tensor<fp16, [?, 128, 1, 5]> key_states_43_cast_fp16 = conv(dilations = key_states_43_dilations_0, groups = key_states_43_groups_0, pad = key_states_43_pad_0, pad_type = key_states_43_pad_type_0, strides = key_states_43_strides_0, weight = var_1168_to_fp16, x = var_1202_cast_fp16_0)[name = string("key_states_43_cast_fp16")]; | |
| string value_states_43_pad_type_0 = const()[name = string("value_states_43_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> value_states_43_strides_0 = const()[name = string("value_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> value_states_43_pad_0 = const()[name = string("value_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> value_states_43_dilations_0 = const()[name = string("value_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 value_states_43_groups_0 = const()[name = string("value_states_43_groups_0"), val = int32(1)]; | |
| tensor<fp16, [128, 1024, 1, 1]> var_1169_to_fp16 = const()[name = string("op_1169_to_fp16"), val = tensor<fp16, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211752704)))]; | |
| tensor<fp16, [?, 128, 1, 5]> value_states_43_cast_fp16 = conv(dilations = value_states_43_dilations_0, groups = value_states_43_groups_0, pad = value_states_43_pad_0, pad_type = value_states_43_pad_type_0, strides = value_states_43_strides_0, weight = var_1169_to_fp16, x = var_1202_cast_fp16_0)[name = string("value_states_43_cast_fp16")]; | |
| tensor<int32, [4]> concat_28x = const()[name = string("concat_28x"), val = tensor<int32, [4]>([-1, 16, 64, 5])]; | |
| tensor<fp16, [?, 16, 64, 5]> embed_29_cast_fp16 = reshape(shape = concat_28x, x = query_states_43_cast_fp16)[name = string("embed_29_cast_fp16")]; | |
| tensor<int32, [4]> concat_29x = const()[name = string("concat_29x"), val = tensor<int32, [4]>([-1, 2, 64, 5])]; | |
| tensor<fp16, [?, 2, 64, 5]> embed_cast_fp16 = reshape(shape = concat_29x, x = key_states_43_cast_fp16)[name = string("embed_cast_fp16")]; | |
| tensor<int32, [4]> concat_30x = const()[name = string("concat_30x"), val = tensor<int32, [4]>([-1, 2, 64, 5])]; | |
| tensor<fp16, [?, 2, 64, 5]> value_states_45_cast_fp16 = reshape(shape = concat_30x, x = value_states_43_cast_fp16)[name = string("value_states_45_cast_fp16")]; | |
| tensor<fp16, [?, 16, 64, 5]> var_1228_cast_fp16 = mul(x = embed_29_cast_fp16, y = cos_to_fp16)[name = string("op_1228_cast_fp16")]; | |
| tensor<int32, [2]> var_1229_split_sizes_0 = const()[name = string("op_1229_split_sizes_0"), val = tensor<int32, [2]>([32, 32])]; | |
| int32 var_1229_axis_0 = const()[name = string("op_1229_axis_0"), val = int32(-2)]; | |
| tensor<fp16, [?, 16, 32, 5]> var_1229_cast_fp16_0, tensor<fp16, [?, 16, 32, 5]> var_1229_cast_fp16_1 = split(axis = var_1229_axis_0, split_sizes = var_1229_split_sizes_0, x = embed_29_cast_fp16)[name = string("op_1229_cast_fp16")]; | |
| fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 16, 32, 5]> var_1231_cast_fp16 = mul(x = var_1229_cast_fp16_1, y = const_30_promoted_to_fp16)[name = string("op_1231_cast_fp16")]; | |
| bool var_1233_interleave_0 = const()[name = string("op_1233_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 16, 64, 5]> var_1233_cast_fp16 = concat(axis = var_1172, interleave = var_1233_interleave_0, values = (var_1231_cast_fp16, var_1229_cast_fp16_0))[name = string("op_1233_cast_fp16")]; | |
| tensor<fp16, [?, 16, 64, 5]> var_1234_cast_fp16 = mul(x = var_1233_cast_fp16, y = sin_to_fp16)[name = string("op_1234_cast_fp16")]; | |
| tensor<fp16, [?, 16, 64, 5]> query_states_45_cast_fp16 = add(x = var_1228_cast_fp16, y = var_1234_cast_fp16)[name = string("query_states_45_cast_fp16")]; | |
| tensor<fp16, [?, 2, 64, 5]> var_1236_cast_fp16 = mul(x = embed_cast_fp16, y = cos_to_fp16)[name = string("op_1236_cast_fp16")]; | |
| tensor<int32, [2]> var_1237_split_sizes_0 = const()[name = string("op_1237_split_sizes_0"), val = tensor<int32, [2]>([32, 32])]; | |
| int32 var_1237_axis_0 = const()[name = string("op_1237_axis_0"), val = int32(-2)]; | |
| tensor<fp16, [?, 2, 32, 5]> var_1237_cast_fp16_0, tensor<fp16, [?, 2, 32, 5]> var_1237_cast_fp16_1 = split(axis = var_1237_axis_0, split_sizes = var_1237_split_sizes_0, x = embed_cast_fp16)[name = string("op_1237_cast_fp16")]; | |
| fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 2, 32, 5]> var_1239_cast_fp16 = mul(x = var_1237_cast_fp16_1, y = const_31_promoted_to_fp16)[name = string("op_1239_cast_fp16")]; | |
| bool var_1241_interleave_0 = const()[name = string("op_1241_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2, 64, 5]> var_1241_cast_fp16 = concat(axis = var_1172, interleave = var_1241_interleave_0, values = (var_1239_cast_fp16, var_1237_cast_fp16_0))[name = string("op_1241_cast_fp16")]; | |
| tensor<fp16, [?, 2, 64, 5]> var_1242_cast_fp16 = mul(x = var_1241_cast_fp16, y = sin_to_fp16)[name = string("op_1242_cast_fp16")]; | |
| tensor<fp16, [?, 2, 64, 5]> key_states_45_cast_fp16 = add(x = var_1236_cast_fp16, y = var_1242_cast_fp16)[name = string("key_states_45_cast_fp16")]; | |
| tensor<int32, [2]> var_1247_split_sizes_0 = const()[name = string("op_1247_split_sizes_0"), val = tensor<int32, [2]>([8, 8])]; | |
| int32 var_1247_axis_0 = const()[name = string("op_1247_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 8, 64, 5]> var_1247_cast_fp16_0, tensor<fp16, [?, 8, 64, 5]> var_1247_cast_fp16_1 = split(axis = var_1247_axis_0, split_sizes = var_1247_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_1247_cast_fp16")]; | |
| tensor<int32, [2]> var_1249_split_sizes_0 = const()[name = string("op_1249_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_1249_axis_0 = const()[name = string("op_1249_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1, 64, 5]> var_1249_cast_fp16_0, tensor<fp16, [?, 1, 64, 5]> var_1249_cast_fp16_1 = split(axis = var_1249_axis_0, split_sizes = var_1249_split_sizes_0, x = key_states_45_cast_fp16)[name = string("op_1249_cast_fp16")]; | |
| tensor<int32, [2]> var_1251_split_sizes_0 = const()[name = string("op_1251_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_1251_axis_0 = const()[name = string("op_1251_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1, 64, 5]> var_1251_cast_fp16_0, tensor<fp16, [?, 1, 64, 5]> var_1251_cast_fp16_1 = split(axis = var_1251_axis_0, split_sizes = var_1251_split_sizes_0, x = value_states_45_cast_fp16)[name = string("op_1251_cast_fp16")]; | |
| bool attn_weights_85_transpose_x_1 = const()[name = string("attn_weights_85_transpose_x_1"), val = bool(true)]; | |
| bool attn_weights_85_transpose_y_1 = const()[name = string("attn_weights_85_transpose_y_1"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_85_cast_fp16 = matmul(transpose_x = attn_weights_85_transpose_x_1, transpose_y = attn_weights_85_transpose_y_1, x = var_1249_cast_fp16_0, y = var_1247_cast_fp16_0)[name = string("attn_weights_85_cast_fp16")]; | |
| fp16 _inversed_attn_weights_87_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_87_y_0_to_fp16"), val = fp16(0x1p-3)]; | |
| tensor<fp16, [?, 8, 5, 5]> _inversed_attn_weights_87_cast_fp16 = mul(x = attn_weights_85_cast_fp16, y = _inversed_attn_weights_87_y_0_to_fp16)[name = string("_inversed_attn_weights_87_cast_fp16")]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_89_cast_fp16 = softmax(axis = var_1181, x = _inversed_attn_weights_87_cast_fp16)[name = string("attn_weights_89_cast_fp16")]; | |
| bool var_1258_transpose_x_0 = const()[name = string("op_1258_transpose_x_0"), val = bool(false)]; | |
| bool var_1258_transpose_y_0 = const()[name = string("op_1258_transpose_y_0"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 64, 5]> var_1258_cast_fp16 = matmul(transpose_x = var_1258_transpose_x_0, transpose_y = var_1258_transpose_y_0, x = var_1251_cast_fp16_0, y = attn_weights_89_cast_fp16)[name = string("op_1258_cast_fp16")]; | |
| bool attn_weights_91_transpose_x_1 = const()[name = string("attn_weights_91_transpose_x_1"), val = bool(true)]; | |
| bool attn_weights_91_transpose_y_1 = const()[name = string("attn_weights_91_transpose_y_1"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_91_cast_fp16 = matmul(transpose_x = attn_weights_91_transpose_x_1, transpose_y = attn_weights_91_transpose_y_1, x = var_1249_cast_fp16_1, y = var_1247_cast_fp16_1)[name = string("attn_weights_91_cast_fp16")]; | |
| fp16 _inversed_attn_weights_93_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_93_y_0_to_fp16"), val = fp16(0x1p-3)]; | |
| tensor<fp16, [?, 8, 5, 5]> _inversed_attn_weights_93_cast_fp16 = mul(x = attn_weights_91_cast_fp16, y = _inversed_attn_weights_93_y_0_to_fp16)[name = string("_inversed_attn_weights_93_cast_fp16")]; | |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_cast_fp16 = softmax(axis = var_1181, x = _inversed_attn_weights_93_cast_fp16)[name = string("attn_weights_cast_fp16")]; | |
| bool attn_output_29_transpose_x_0 = const()[name = string("attn_output_29_transpose_x_0"), val = bool(false)]; | |
| bool attn_output_29_transpose_y_0 = const()[name = string("attn_output_29_transpose_y_0"), val = bool(false)]; | |
| tensor<fp16, [?, 8, 64, 5]> attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = var_1251_cast_fp16_1, y = attn_weights_cast_fp16)[name = string("attn_output_29_cast_fp16")]; | |
| bool attn_output_interleave_0 = const()[name = string("attn_output_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 16, 64, 5]> attn_output_cast_fp16 = concat(axis = var_1176, interleave = attn_output_interleave_0, values = (var_1258_cast_fp16, attn_output_29_cast_fp16))[name = string("attn_output_cast_fp16")]; | |
| tensor<int32, [4]> concat_31x = const()[name = string("concat_31x"), val = tensor<int32, [4]>([-1, 1024, 1, 5])]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_137_cast_fp16 = reshape(shape = concat_31x, x = attn_output_cast_fp16)[name = string("x_137_cast_fp16")]; | |
| string hidden_states_45_pad_type_0 = const()[name = string("hidden_states_45_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> hidden_states_45_strides_0 = const()[name = string("hidden_states_45_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> hidden_states_45_pad_0 = const()[name = string("hidden_states_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> hidden_states_45_dilations_0 = const()[name = string("hidden_states_45_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 hidden_states_45_groups_0 = const()[name = string("hidden_states_45_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 1024, 1, 1]> var_1175_to_fp16 = const()[name = string("op_1175_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212014912)))]; | |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_45_cast_fp16 = conv(dilations = hidden_states_45_dilations_0, groups = hidden_states_45_groups_0, pad = hidden_states_45_pad_0, pad_type = hidden_states_45_pad_type_0, strides = hidden_states_45_strides_0, weight = var_1175_to_fp16, x = x_137_cast_fp16)[name = string("hidden_states_45_cast_fp16")]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_139_cast_fp16 = add(x = x_131_cast_fp16, y = hidden_states_45_cast_fp16)[name = string("x_139_cast_fp16")]; | |
| fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_1277_cast_fp16 = mul(x = x_139_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1277_cast_fp16")]; | |
| bool x_141_interleave_0 = const()[name = string("x_141_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2048, 1, 5]> x_141_cast_fp16 = concat(axis = var_1176, interleave = x_141_interleave_0, values = (x_139_cast_fp16, var_1277_cast_fp16))[name = string("x_141_cast_fp16")]; | |
| tensor<int32, [1]> out_91_axes_0 = const()[name = string("out_91_axes_0"), val = tensor<int32, [1]>([1])]; | |
| fp16 var_1287_to_fp16 = const()[name = string("op_1287_to_fp16"), val = fp16(0x1.5p-17)]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_91_cast_fp16 = layer_norm(axes = out_91_axes_0, epsilon = var_1287_to_fp16, x = x_141_cast_fp16)[name = string("out_91_cast_fp16")]; | |
| tensor<fp16, [1, 2048, 1, 1]> layer_encoder_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214112128)))]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_93_cast_fp16 = mul(x = out_91_cast_fp16, y = layer_encoder_layers_7_post_attention_layernorm_weight_to_fp16)[name = string("out_93_cast_fp16")]; | |
| tensor<int32, [2]> var_1293_split_sizes_0 = const()[name = string("op_1293_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; | |
| int32 var_1293_axis_0 = const()[name = string("op_1293_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_1293_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_1293_cast_fp16_1 = split(axis = var_1293_axis_0, split_sizes = var_1293_split_sizes_0, x = out_93_cast_fp16)[name = string("op_1293_cast_fp16")]; | |
| string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)]; | |
| tensor<fp16, [4096, 1024, 1, 1]> var_1162_to_fp16 = const()[name = string("op_1162_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214116288)))]; | |
| tensor<fp16, [?, 4096, 1, 5]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = var_1162_to_fp16, x = var_1293_cast_fp16_0)[name = string("input_cast_fp16")]; | |
| tensor<fp16, [?, 4096, 1, 5]> var_1301_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_1301_cast_fp16")]; | |
| string var_1306_pad_type_0 = const()[name = string("op_1306_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> var_1306_strides_0 = const()[name = string("op_1306_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> var_1306_pad_0 = const()[name = string("op_1306_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> var_1306_dilations_0 = const()[name = string("op_1306_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_1306_groups_0 = const()[name = string("op_1306_groups_0"), val = int32(1)]; | |
| tensor<fp16, [4096, 1024, 1, 1]> var_1163_to_fp16 = const()[name = string("op_1163_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222504960)))]; | |
| tensor<fp16, [?, 4096, 1, 5]> var_1306_cast_fp16 = conv(dilations = var_1306_dilations_0, groups = var_1306_groups_0, pad = var_1306_pad_0, pad_type = var_1306_pad_type_0, strides = var_1306_strides_0, weight = var_1163_to_fp16, x = var_1293_cast_fp16_0)[name = string("op_1306_cast_fp16")]; | |
| tensor<fp16, [?, 4096, 1, 5]> x_147_cast_fp16 = mul(x = var_1301_cast_fp16, y = var_1306_cast_fp16)[name = string("x_147_cast_fp16")]; | |
| string hidden_states_pad_type_0 = const()[name = string("hidden_states_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> hidden_states_strides_0 = const()[name = string("hidden_states_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> hidden_states_pad_0 = const()[name = string("hidden_states_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> hidden_states_dilations_0 = const()[name = string("hidden_states_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 hidden_states_groups_0 = const()[name = string("hidden_states_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 4096, 1, 1]> var_1164_to_fp16 = const()[name = string("op_1164_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230893632)))]; | |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_cast_fp16 = conv(dilations = hidden_states_dilations_0, groups = hidden_states_groups_0, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = hidden_states_strides_0, weight = var_1164_to_fp16, x = x_147_cast_fp16)[name = string("hidden_states_cast_fp16")]; | |
| tensor<fp16, [?, 1024, 1, 5]> x_149_cast_fp16 = add(x = x_139_cast_fp16, y = hidden_states_cast_fp16)[name = string("x_149_cast_fp16")]; | |
| int32 var_1318 = const()[name = string("op_1318"), val = int32(1)]; | |
| fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_1321_cast_fp16 = mul(x = x_149_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_1321_cast_fp16")]; | |
| bool x_151_interleave_0 = const()[name = string("x_151_interleave_0"), val = bool(false)]; | |
| tensor<fp16, [?, 2048, 1, 5]> x_151_cast_fp16 = concat(axis = var_1318, interleave = x_151_interleave_0, values = (x_149_cast_fp16, var_1321_cast_fp16))[name = string("x_151_cast_fp16")]; | |
| tensor<int32, [1]> out_97_axes_0 = const()[name = string("out_97_axes_0"), val = tensor<int32, [1]>([1])]; | |
| fp16 var_1331_to_fp16 = const()[name = string("op_1331_to_fp16"), val = fp16(0x1.5p-17)]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_97_cast_fp16 = layer_norm(axes = out_97_axes_0, epsilon = var_1331_to_fp16, x = x_151_cast_fp16)[name = string("out_97_cast_fp16")]; | |
| tensor<fp16, [1, 2048, 1, 1]> layer_encoder_norm_weight_to_fp16 = const()[name = string("layer_encoder_norm_weight_to_fp16"), val = tensor<fp16, [1, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239282304)))]; | |
| tensor<fp16, [?, 2048, 1, 5]> out_99_cast_fp16 = mul(x = out_97_cast_fp16, y = layer_encoder_norm_weight_to_fp16)[name = string("out_99_cast_fp16")]; | |
| tensor<int32, [2]> var_1337_split_sizes_0 = const()[name = string("op_1337_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; | |
| int32 var_1337_axis_0 = const()[name = string("op_1337_axis_0"), val = int32(1)]; | |
| tensor<fp16, [?, 1024, 1, 5]> var_1337_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_1337_cast_fp16_1 = split(axis = var_1337_axis_0, split_sizes = var_1337_split_sizes_0, x = out_99_cast_fp16)[name = string("op_1337_cast_fp16")]; | |
| tensor<int32, [4]> x_begin_0 = const()[name = string("x_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [4]> x_end_0 = const()[name = string("x_end_0"), val = tensor<int32, [4]>([0, 1024, 1, 1])]; | |
| tensor<bool, [4]> x_end_mask_0 = const()[name = string("x_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])]; | |
| tensor<fp16, [?, 1024, 1, 1]> x_cast_fp16 = slice_by_index(begin = x_begin_0, end = x_end_0, end_mask = x_end_mask_0, x = var_1337_cast_fp16_0)[name = string("x_cast_fp16")]; | |
| string var_1355_pad_type_0 = const()[name = string("op_1355_pad_type_0"), val = string("valid")]; | |
| tensor<int32, [2]> var_1355_strides_0 = const()[name = string("op_1355_strides_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [4]> var_1355_pad_0 = const()[name = string("op_1355_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> var_1355_dilations_0 = const()[name = string("op_1355_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| int32 var_1355_groups_0 = const()[name = string("op_1355_groups_0"), val = int32(1)]; | |
| tensor<fp16, [1024, 1024, 1, 1]> var_1349_to_fp16 = const()[name = string("op_1349_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239286464)))]; | |
| tensor<fp16, [1024]> enc_to_lm_proj_bias_to_fp16 = const()[name = string("enc_to_lm_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241383680)))]; | |
| tensor<fp16, [?, 1024, 1, 1]> output = conv(bias = enc_to_lm_proj_bias_to_fp16, dilations = var_1355_dilations_0, groups = var_1355_groups_0, pad = var_1355_pad_0, pad_type = var_1355_pad_type_0, strides = var_1355_strides_0, weight = var_1349_to_fp16, x = x_cast_fp16)[name = string("op_1355_cast_fp16")]; | |
| } -> (output); | |
| } |