Release Note¶

Add paddle.io.random_split and paddle.io.Subset. (#32090)

Fix the issue that the stride and padding of paddle.nn.MaxPool3D and paddle.nn.AvgPool3D do not have default values. (#32014)

Fix the issue that when RNN supporting cudnn creates parameters, repeated creations are reported. (#31916)

Fix the issue that when the soft_label of paddle.nn.functional.cross_entropy is True, and the weight parameter is specified, an error will be reported; add the use_softmax parameter to paddle.nn.functional.cross_entropy, which controls whether to perform softmax operation before calculating the cross entropy; add the deprecated mark to paddle.nn.functional.softmax_with_cross_entropy, for this API will be deprecated in the future version. (#31953, #32105, #32035)

Fix the issue of paddle.nn.ClipByNorm generating NaN values as the gradients are all zero, which will lead to non-convergence when using mixed precision training. (#32038)

Fix the issue of accessing array out of bounds in paddle.stack. (#32005)

Fix the issue that in parameter server mode the calculation graph segmentation supports GradClip strategy.(#31945)

Fix the initialization of truncated gaussian distribution in parameter server mode.(#31945)

Fix the issue of incorrectly printing the Profiler’s multi-threaded information in parameter server mode.(#31945)

Fix the Python3 incompatibility issue when data are read by Dataset and output by zip.(#31945)

Clean up redundant log information and optimize the output format of exe.train_from_dataset.(#32009)

Paddle-TRT adapts to the ERNIE/BERT model trained and saved by PaddlePaddle 2.0.(#31959)

Upgrade onednn to version 2.2, which has improved many models inference performance. (#31270)

Add hard_swish oneDNN support and conv + hard_swish fusion, which has improved ocr_det model inference performance by 18% on SkyLake. (#31870)

Add aligned in roi_align, and pixel_offset in generate_proposals、distribute_fpn_proposals to improve performance.

paddle.nn.functional.cross_entropy supports float type label in XPU accelerator.

Add label error checks and optimized error message of paddle.nn.functional.softmax_with_cross_entropy.

paddle.nn.LayerList supports paddle.nn.LayerList([None]) .

Add the support of tuple as loop variable in for-loop.

Add Tensor support to be indexed by unspecific start and stop variables, such as x[:], x[2:].

Now Tensor supports slicing with lvalue in static graph. In dynamic graph, Tensor uses slicing can correctly turn into static graph. Tensor can be modified by indexing or slicing. Python.Int、Tensor、Python.slice can be used for indexing. The stride could be 1, greater than 1 or negative. NumPy.array, Tensor types could be used as rvalue.

Mixed precision training of dynamic graph supports paddle.nn.LayerNorm , improving efficiency by reducing the number of cast.

paddle.distributed.fleet.DistributedStrategy amp adds pure fp16 strategy.

paddle.distributed.ProbabilityEntry and paddle.distributed.CountFilterEntry are added for sparse parameters training.

Optimized the number of communications in parallel pipeline.

In parameter server mode, fields like count/unseen_day could be saved into model.

Add the elimination strategy of sparse parameters in parameter server mode.

paddle.save and paddle.static.save allow users to select the pickle version, and the default version is 2. For Python 3, users can choose Pickle version 4+. In this way, saving speed could be increased and single file could over 4G. But, please notice that models saved this way must be loaded and used under Python 3.

Add paddle.static.normalize_program to obtain the pruned computation graph.

paddle.abs supports Complex64 and Complex128 types.

Offered the new scheme of custom operators outside the framework, simplify the writing and using process of custom operators, support two installation and calling methods, and support Linux and Window at the same time; custom operators by using the new scheme can be used in dynamic graphs, static graphs, dynamic-to-static and inference scenarios; for specific instructions, please refer to the file: Customizing External Operators.

Fixed Entry Config has no effect issue in parameter server mode.

Fixed the saved parameters could not be loaded issue in parameter server mode.

Fixed Profiler abnormal issue in parameter server mode.

Fixed training abnormal issue when data type category is higher than INT32 in parameter server mode.

Fixed long stringIP cannot be bounded issue in parameter server mode.

Fixed the issue of too much log outputs issue, in distributed training caused by lower level LOG config.

Fixed the issue of inconsistent parameters of each devices, when if else control flow is used in dynamic graph distributed training.

Fixed the issue that FLAGS setting of multi-host distributed training is not consistent with single host distributed training.

Fixed the muti_precision function of paddle.optimizer.AdamW to ensure the master weights in FP32 type, which are regularized, in order to prevent possible diverge.

Fixed the issue when the input of paddle.nn.ELU is nan, the output is nan.

Fixed gradient calculation error of using Tensor.backward() for gradient accumulation, in dynamic graph mulit-card training.

Fixed the integer overflow issue when paddle.nn.functional.softmax_with_cross_entropy processes a Tensor with over 2^31 elements.

Fixed crash bug during the for-loop traversal of paddle.nn.Sequential.

Fixed Wrong error message of dynamic graph slicing.

Fixed the issue that batch_size=-1 cannot be used, when paddle.nn.functional.local_response_norm is used in static graph or dynamic graph to static graph converting.

Fixed paddle.nn.LayerNorm computation error when data type is float64.

Fixed the error message of metric_learning finetune under PaddlePaddle/models.

Fixed weight asynchrony issue caused by lack of operators, when XPU’s static graph multi-card is used.

Add API— paddle_infer::Config::EnableTensorRtDLA(). At the stage of inference, users can apply DLA of NVIDIA while using TensorRT.

Paddle-TRT will check inputs of model, If input shape is variant, the error messages are optimized and Paddle-TRT will hint users to use dynamic_shape.

Support inference and deployment models that have the operators customized by users, and provide User Documentation.

PaddlePaddle inference APIs of C++ and Python support XPU, which is aligned with training supported by PaddlePaddle to XPU.

Paddle-TRT supports group_norm op, and speed up solov2_r50_fpn_1x as following: compared with TRT v2.0.0, on T4, CUDA11, cuDNN8.1 and TRT7.1.3, the performance of TRT FP32 improves by 13%, from 87.019ms to 75.13ms, and the performance of TRT FP16 improves by 65%, from 72.9253ms to 44.149ms.

Fix some operator problems in TensorRT v7.1+, like TensorRT’s inference of ERNIE.

Fix some issues of using Python pass_builder API.

Due to limited memory of Jetson, auto_growth is regarded as default distribution policy of memory, tackling problems that some models cannot run with limited memory.

Avoid the problem of cuDNN8.0’s memory leaks to ensure the availability, and this will not influence other versions of cuDNN.

Fixed MakeCipher’s symbol absence issue in inference dynamic library.

Fixed wrong predicting results issue of mask_rcnn_r50_1x_coco model when this static graph model is converted from dynamic graph.

Fixed the inference failure of the segmentation models, caused by adaptive pooling is not fully supported by oneDNN,

Fixed the issue that oneDNN’s OCR model inference will be incorrect when batch_size>1.

Fixed freeze_model inference failure due to ReLU CPU’s implementation error.

Fixed the incompatibility issue that BF16’s images cannot change into binary script for Python3.

Usability optimization:

• Tensor function enhancements: Added Tensor copy interface Tensor.clone(), and more than 120 Tensor computation interfaces (e.g. Tensor.cos(), etc.). Added the new function to modify the Tensor function by using index or slice. Added the new function of automatic type boost in case of Tensor and Scalar operation. Optimized the printing information of dynamic graph Tensor. The display form remains the same as Numpy.

• Layer function enhancement: Added the new Layer deep copy interface Layer.deepcopy(). Added the new Layer property and function to view interface Layer.dir(). From this version, the Trace function still records reverse operation automatically after the invoking of Layer.eval(). If you don’t need to record reverse, you need to explicitly call paddle. no_grad().

• Added a set_lr() interface for Optimizer so that users can flexibly adjust a learning rate in dynamic graph mode.

• Added a new set_global_initializer() interface to define global parameter initialization methods.

Simplified the code for multi-card operation without needing to explicitly call scale_loss and apply_collective_grads.

Performance optimization:

• Supported the function of gradient updating by using sparse parameters for APIs (for example, embedding) in case of multi-card training.

• For dynamic graph training and inference, added the support for Intel acceleration library oneDNN (former MKL-DNN). The speed of Resnet50 model in CPU training scenario can improve by 6 times.

• New dynamic graph Inplace calculation function: The Tensor storage space can be reused, reducing the occupation of video memory. Added the new View method. You can change the Tensor description in case of shared underlying storage.

• [Incompatible upgrade] new dynamic graph gradient accumulation function, with disguised “expand BatchSize” role. By default, the gradient of backward() interface is not clear, with needing to explicitly call optimizer.clear_grad() to clear the gradient.

Fixing bugs:

• Fixed the bug of train and eval interference with each other when switching between them in multiple models.

Added the grammar support for dynamic-to-static conversion

• Added the support for the return grammar. In the if-elif-else or loop conditions, the loop can return earlier, with return different types of tensor or None.

• Added support for the **kwargs parameter contained in the signature function.

• Added the grammar support of “for”, “for enumerate” traversing Tensor and TensorList, more flexible operation for traversing Tensor.

• Added the support for more python grammars, such as print, assert, cast, isinstance, tuple, dict.pop(), etc.

Optimized the usability of dynamic-static conversion

• Changed the return type of dynamic-to-static from callable function to Class. The code and main_program interfaces invoking the Class can obtain the converted static graph information more easily.

• The dynamic-to-static decorator to_static is added with the directly decorating model instances, such as to_static (model, input_spec).

• Added the jit.not_to_static decorator. The function is not converted in the dynamic-to-static process.

• Added set_verbosity() and set_code_level() interfaces. You can set different levels to view the log or intermediate state code of the dynamic to static process.

• Added InputSpec. You can specify the shape and data type of input Tensor variables in the dynamic to static process.

• Error message optimization: Locate the specific wrong line of code in the original dynamic graph and hide the user-unrelated error message.

• Support break point debugging by using pdb.set_trace().

Optimized deployment of model storage and loading APIs

• Added paddle.jit.save interface for storing dynamic-to-static models: The interface is compatible with and used to store both the Layer object not transcribed by paddle.jit.to_static and paddle.DataParallel models. Remove the old interface ProgramTranslator. save_ inference_model.

• Added the paddle.jit.load interface for loading prediction models stored in static graph format, including models saved by paddle.jit.save and paddle.io.save_inference_model. This can be used for model inference or model training optimization under dynamic graph after loading.

• Added the program method for opaddle.jit. TransLatedLayer for obtaining the program of the paddle.jit.load loading model. It is for understanding of the model structure.

• [Incompatible upgrade] changed the meaning of the interface parameter model_path of paddle.jit.save and paddle.jit.load: That is, changed to the prefix of storage files instead of that of directory.

Mixed precision policy upgrade: In addition to the black and white list policy (hereinafter referred to as “O1 policy”), “Almost FP16 (hereinafter referred to as O2 policy)” is added. That is, use FP16 for calculation as much as possible.

• Added the FP16 Guard function (paddle.static.amp.fp16_guard): Support users to freely control whether a single Op in the model chooses FP16 calculation type.

• User can customize custom_black_list to control a certain type of Op to keep FP32 computation.

• Using the O2 policy: Resnet50 and Bert base can be trained at 1400 images/s and 590 sequences/s, respectively, on a single card V100.

Usability optimization:

• Use the paddle.static.amp package to manage the interfaces related to static graph mixed precision training in a unified manner.

• Provide the simplified name CustomOpLists for AutoMixedPrecisionLists: That is, users can customize the AMP black and white list Op list by using CustomOpLists.

Integrated communication All Reduce

• Support mixed parallel training of 100 billion language models: support pipeline parallel training based on the executor interface, with sharding-DP strategy, GradientMerge+AMP strategy, Recompute+Offload strategy, and megatron strategy.

• Support dynamic graph: support multi-stream communication strategy, automatic rebuild group strategy, high performance sparse parameter communication, and multi-card gradient sequential consistency strategy.

Parameter server PS

• Upgraded the large-scale sparse function: Upgrade large-scale sparse PS-API, and abstract communication component/parameter table/optimizer base class. It is convenient for users to carry out secondary development in a subclass derivation mode. Meanwhile, it also supports 100 billion features streaming training, including feature access, exit, incremental training, distributed metrics prediction, etc. The communication mode switches from GRPC to BRPC.

• Open source heterogeneous parameter server: Support both traditional pure CPU machine PS, and pure GPU machine PS based on three levels of storage (SSD/memory/video memory). It also supports CPU machine + GPU machine/Kunlun machine mixing distributed PS, with completing the minute-level training of trillions of parameter hit rate prediction models

Support of new training mechanism:

• Support control flow-based multitasking distributed training: The performance is improved by more than 50% compared to the Intag-based multitasking.

Optimization of the distributed startup method

• Supported distributed low-order APIs such as all_gather using the paddle.distibuted.spawn interface

• Upgraded the paddle.distributed.launch interface: Support specifying the number of processes in a single node with simplifying as fleetrun.

• Optimized gen_nccl_id: Removed the grpc dependency, added some fault tolerance, and improved the stability of starting distributed tasks.

• Supported the startup of multi-CPU in the integrated communication in the Gloo method

Standardized the set_dict method name of APIs such as Layer and Optimzier: That is, changed to set_state_dict in a unified manner.

Enhanced paddle.load compatibility: support the loading of Layer’s state_dict from storage results of interfaces such as fluid.io.save_inference_model and fluid.io.save_params/persistables.

Modified the paddle. save/load interface behavior: For the paddle.save, A suffix is not added to the storage results. In each loading, paddle.load returns only one result. Standardize the interface semantics.

Removed paddle.SaveLoadConfig: For the interface compatibility loading scenarios of paddle.jit.save, paddle.jit.load, and paddle.load, use **kwargs to pass in additional configuration to simplify the use of the interface.

Moved the original static graph APIs such as paddle.io.save, paddle.io.load, paddle.io.save_inference_model, and paddle.io.load_inference_model to the paddle.static module.

Optimized the paddle.static.load_program_state interface experience. In the scenarios without specifying the loading var_list, only a warning (instead of error report) is given when there is an interference file in the loading of a directory.

Extended the dynamic static graph execution engine: Support the plural neural network training and plural gradient accumulation.

Added Op such as mul, div, matmul, kron, and abs for supporting the plural computation.

Added API: paddle.onnx.export for supporting the conversion from Paddle2.0 dynamic graph to ONNX protocol.

Added PPOCR, PPYOLO, FasterRCNN, and ERNIE for model conversion.

Richer Paddle op coverage: Support 88 Paddle OP operators. Support the export as different versions of ONNX 1~12 operator sets.

dynamic graph performance optimization:

• Optimized the data read performance: Simplify the DataLoader underlying implementation logic in dynamic graph mode, reduce the thread reading overhead, and further improve the data reading efficiency and the overall model training speed. The overall training speed of MobileNetV1 in a scenario of single card V100 and BatchSize = 128 is improved by 34%.

• Upgraded and performance optimization of dynamic graph networking API: A large number of dynamic graph APIs directly call an automatically generated Pybind API. As a result, the performance is improved significantly.

• Improved the training performance of Resnet50 oneDNN dynamic graph. The dynamic graph training speed of the current CPU scenario Resnet50 oneDNN is improved by 6.4 times.

OP performance optimization:

• argsort: The number of elements of the input Tensor is optimized as the number equal to its axis dimensional length. In this way, the forward speed is improved by 34 times, and the reverse speed is improved by 10 times.

• dropout: Optimized GPU performance. The FP32 performance is improved by 20%. The FP16 performance is improved by 50%.

• cast: Optimized GPU performance. The performance is improved by 10% to 20%.

• softmax: Optimized GPU performance in case of axis=-1. The performance is improved by 3 times to 96 times for different shapes.

• Performance optimization of other OPs: Significantly improved the performance of other OPs such as cumsum, reshape, Flatten, IndexSelect, Roll, elementwise_add, AdamW and RNN class (LSTM, GRU, SimpleRNN).

Optimization strategy:

• Added fused_bn_add_act fusion strategy: Performed the automatic fusion acceleration for the combined pattern of batch_norm+elementwise_add+activation.

• Added inplace addto strategy for gradient aggregation: Support in-situ gradient accumulation. Improve the performance by 6.3% in ResNet-50 mixed precision training.

Optimized FastThreadedSSAGraphExecutor scheduling: Fixed the bug that the communication calculation does not overlap in the communication synchronization scenario. The performance of 4 machines and 32 cards resnet50 is improved by about 0.3%.

Distributed performance optimization:

• Optimized lars strategy: The time2train index of 16k batch size in the ResNet50 distributed multi-card training is smaller than 10 minutes.

• Optimized the paddle.fleet amp distributed performance: Fixed the bug that the last communication and calculation are not overlapping. The performance of the 4-machine 32-card FP16 is improved by about 0.5%.

• Optimized paddle. fleet.gradient_merge distributed performance: Aggregate gradients before communication. The multi-machine performance can be improved by 20%-40% to achieve linear acceleration ratio.

• Optimized the performance of the parameter server communication component Communicator. In case of GEO-400batch communication once, the W2V model throughput rate and Simnet-Bow model performance are significantly improved. In the Async mode, compared to the PaddlePaddle Framework 1.8, the throughput rate of W2V model is improved by 11% and the performance of CTR-DNN model is improved by 14%

Uniformly changed the wording of LOG(FATAL) throw exception at just 100 points to PADDLE_THROW: Optimize the error format and content caused by non-support of a framework behavior.

Improved the Signal Handler implementation within the framework. Optimized the error format and content when system signal error occurs during the execution.

Optimized the framework error stack format: In the compiling, the python error stack is moved below the native error stack to improve error message reading experience.

An accumulative total of about 1500 error type and prompt copywritings of check errors within the framework. This enhances the overall debugging usability of the framework.

Enhanced dynamic graph error messages: Error messages on the Pybind layer under a dynamic graph are systematically enhanced to improve user experience.

Optimized exception types of Paddle Python side error report: Align with Python native error report types.

Hide the C++ error stack by default: Optimized the error format after hiding the C++ stack, removed the demarcation flag Error Message Summary, and aligned with the native Python error format.

Optimized the error prompts of APIs in non-static graph mode in some static modules, including 9 APIs such as static. append_backward, static.gradients, static.scope_guard, static.Print, static.nn.embedding, static.nn. data_norm, static.nn.multi_box_head, static.nn.nce, and static.nn.py_func.

Optimized the error message when passing in Tensor as None under dynamic graph model.

Optimized the printing information of Layers, and supported printing the relationship of each hierarchy in Layers.

The inference C++ API is upgraded fully. The new APIs are recommended. The old APIs remain temporarily. There is warning reported in the use of old APIs. The old APIs are planned to be deleted in the future. The new APIs include changes of naming standardization and simplification of usage method, including:

• A new paddle_infer namespace for the C++ interface, containing inference-related interfaces.

• Renamed ZeroCopyTensor to Tensor as the default input/output representation of the inference interface.

• Simplify CreatePaddlePredictor to CreatePredictor, with keeping the support for only AnalysisConfig. Other multiple Configs are not supported.

• Added service-related utility classes such as PredictorPool, which can be used when multiple predictors are created.

Operator-related version information

Some operators are newly added or upgraded in Paddle V2.0. Starting from this version, the forward operator version is defined with compatibility constraints. Through the alignment of operator versions between frameworks, ensure consistent definition and behavior of the same operator, thus enhancing the overall robustness of the framework.

Added the registration mechanism for inference forward operator versions and included the incompatible upgrade behavior of operators for the statistics.

Added the operator version information for the prediction models. Through the model file, the inference library is able to identify the definition of the operator corresponding to this model, so as to avoid calculation errors caused by different definitions.

Model interface

• The load_inference_model and save_inference_model APIs are migrated to paddle.static to improve the usability, with compatibility with the old interfaces.

• Added six APIs such as serialize_program, deserialize_program, serialize_persistables, deserialize_persistables, save_to_file, and load_from_file for users to perform serialize/deserialize programs, serialize/deserialize params, and saved models/parameters to file, or loaded models/parameters from files.

Inference-related NV GPU

• Added the adaptive support for TRT 7.1.

• Added the adaptive support for Jetson Nx hardware.

• Paddle-TensorRT enhances the support for the PaddleSlim quantitative model. Cover multiple tasks such as detection, classification, and segmentation on CV.

• Paddle-TRT supports clip op, and supports the classification model GhostNet running on the Paddle-TRT.

• Paddle-TRT supports mul op models with channelwise quantification, and supports the PaddleOCR detection. Identified the quantitative models running in the Paddle-TRT int8.

• Paddle-TRT dynamic shape function supports PaddleSlim quantification Int8 models.

X86 CPU-related inference

• Added the support for oneDNN BF16: support the computation of conv2d and gru bf16. It currently supports BF16 prediction for resnet50, googlenet, mobilenetv1 and mobilenetv2 models.

• Added support for quantification and inverse quantification of scales with bias in oneDNN INT8 quantification strategy.

• Added version compatibility support for some oneDNN operators.

• Added the kernel support for elementwise_add and elementwise_mul INT8 oneDNN on the CPU side.

• Improved the usability of CPU-side test quantification models. Supported the comparative test of original models and quantitative models at the same time.

Custom OP

• Added the support for user-defined Ops on Python-side inference.

Memory/GPU memory correlation

Added the TryShrinkMemory interface. Reduced the occupation of application’s memory/video memory by releasing temporary tensors. For the demo, see Paddle-Inference-Demo.

dynamic graph quantitative model support

• X86 inference supports dynamic graph quantitative models.

• NVIDIA GPU inference supports dynamic graph quantitative model.

Error message:

• In the Compiling, when enabling ON_INFER, FLAGS_call_stack_level is on by default. The error message indicates that the stack is invoked.

Improved the transformation and optimization of quantitative models.

NV GPU correlation

• Optimized the ArgMin and ArgMax OP of CUDA so that the binary system size of the OP is decreased from 60 M to 1.3 M.

• For the ERNIE model on T4 with using the Paddle-TRT FP16 inference, the performance is improved by 15%.

• The ERNIE model adds the support for variable-length inputs when TenorRT is enabled. The performance is improved by 147%.In software versions cuda10.1, cudnn 7.6, tensorrt 6.0, OSS 7.2.1, model ernie- base-2.0, dataset QNLI, the performance on Nvidia Telsa T4 improves from 905 sentences/s to 2237 sentences/s when input BatchSize = 32.Sample code: Paddle-Inference-Demo/c++.

X86 CPU related

• Added the conv + affine_op pass. The MASK-RCNN fp32 single-threaded performance is improved by 26% on machine 6248.

• Added the fc + gru pass and enable oneDNN (former MKL-DNN) GRU fp32 kernel, speeding up GRU fp32 model inference on 4 CRU threads by 20% on machine Intel Xeon 6248.

• By supporting oneDNN INT8 GRU, the GRU INT8 model is about 1.65 times faster compared to NativeConfig inference (threads = 1, batch_size = 50).

• Added the fuse support for oneDNN batchnorm + activation. The pvanet_ocr model performance is improved by 2.8% as a result.

• Added the oneDNN FC + Gelu, FC + Sigmoid and FC + tanh operator fusion. The BERT inference model is improved by 4.5%.

• Added oneDNN inplace support for partial Op

• Optimized oneDNN LRN op (speedup 1% for the GoogleNet fp32 model).

• With oneDNN upgraded to 1.6, Ernie Large oneDNN inference on Skylake (Intel Core 6148) is about 2.7x faster (i.e. unit test test_analyzer_ernie_large).

• Added the interpolate oneDNN forward operator support. Now ocr_det model inference performance improved by 2.04x compared to CPU Native inference alone.

Released the installation package supporting the use of x86 CPUs and the use of Kunlun chips under the FT CPU.

Added the support for python3.8 in the installation package.

Added the installation package for cuda10.1 and cuda 10.2.

(experimental) Released the installation package for cuda11.

Upgraded the Paddle image of cuda 10.1 and later, and the NCCL version in the CI system image to V2.7.8

Upgraded oneDNN (former MKL-DNN) from V1.3 to V1.5.

Added the pre-installed openssl-dev dependencies to the image.

Removed installed dependencies: nltk, opencv, scipy, rarfile, prettytable, pathlib, matplotlib, graphviz, objgraph.

Paddle’s avx and no_avx are released separately. whl package is reduced by 40%. avx version is installed by default. Optimized installation error message. The system checks the user’s CPU type and Paddle version, automatically prompting the corresponding installation error.

Improved the pypi installation user experience for the Paddle develop version. Reduced the user installation path. You can run pip –pre for installation.

The inference library supports cuda10.2-cudnn8-trt7.1 version.

Release the installation package supporting jetpack and C++ inference library supporting nv_jetson.

Newly release the joint compilation of two wheel packages for tensorrt, that is, cuda10.0-cudnn7.6-trt6.0.1.5-python36 and cuda10.0-cudnn7.6-trt6.0.1.5-python36.

Fixed the joint compilation strategy, released the gpu package containing tensorrt separately to avoid the error of no tensorrt when users install the packages of other GPU versions.

Fixed a bug of duplicate in the inference library packages.