Release Notes

MindSpore 2.8.0 Release Notes

Major Features and Improvements

Dataset

• [STABLE] The mindspore.dataset.dataloader.DataLoader API is added, which is consistent with the mainstream API usage in the industry in terms of usage and functions, facilitating user reference. In addition, various commonly used methods for datasets, samplers, collate functions, and tools are added.

• [STABLE] The .send(...) and .recv(...) APIs are added for communication between nodes of the dataset object. These APIs enable the dataset data processing results to be transmitted between different nodes.

• [STABLE] When .map(...) is used for data augmentation and a custom PyFunc augmentation function is used in multi-process mode, if the PyFunc function execution is slow or hung, a warning will be printed to notify the user: ** worker subprocess stack:....

Executor

• [STABLE] MindSpore now supports custom backends, allowing users to adapt to third-party backends.

Compiler

• [STABLE] In graph mode, augmented assignment statements are parsed into corresponding in-place operators, improving graph mode performance and unifying programming experience for pynative and graph mode. Currently only the Ascend backend is supported.

PyNative

• [STABLE] The recompute interface supports setting use_reentrant=False to enable gradient computation for complex type inputs. It also supports setting output_recompute to determine whether outputs should be re-computed.

• [STABLE] Added support for CPU Tensor conversion with DLPack.

• [STABLE] Storage supports shared memory between processes.

API Change

• [STABLE] mindspore.ops API has added three Native Sparse Attention(NSA) interfaces:

  mindspore.ops
  mindspore.ops.nsa_compress
  mindspore.ops.nsa_compress_attention
  mindspore.ops.nsa_select_attention

• [STABLE] The mindspore.mint API has added the mindspore.mint.nn.functional.cosine_embedding_loss and mindspore.mint.nn.CosineEmbeddingLoss interfaces. Most mint interfaces are currently still experimental and outperform ops in terms of performance under Graph Mode O0/O1 and PyNative Mode. O2 compilation mode (Graph Sinking), as well as CPU and GPU backends, are not currently supported and will be improved incrementally.

• [STABLE] The mindspore.mint.nn.functional.adaptive_max_pool2d and mindspore.mint.nn.AdaptiveMaxPool2d interfaces have been promoted from demo to stable status.

• [STABLE] The mindspore.Tensor.view interface now supports dtype as an input argument.

• [STABLE] The mindspore.mint.nn.functional.interpolate interface now supports setting scale_factor in bilinear/bicubic modes. The restriction preventing scale_factor from being set when align_corners=False in linear mode has been lifted.

• [STABLE] The mindspore.ops.grad and mindspore.ops.value_and_grad interfaces add the sens_param parameter, which is used to specify whether to configure sensitivity (gradient with respect to the output) in the input.

• [STABLE] When iterating over a dataset, if the output data contains the string type, the default output type will change from Tensor to numpy.ndarray.

• [BETA] Added mindspore.Tensor.to and mindspore.Tensor.to_, which convert a tensor's device and data type to the specified device and dtype.

• [BETA] Added mindspore.Tensor.delete_ for actively releasing the memory of the tensor on the device or host.

• [BETA] Added mindspore.Tensor.data, providing access to the raw data without tracking its computational history for autograd.

Backwards Incompatible Change

Dataset

• [STABLE] The following obsolete import methods are completely removed: import mindspore.dataset.vision.c_transforms as c_vision, import mindspore.dataset.vision.py_transforms as py_vision, import mindspore.dataset.transforms.c_transforms as c_transforms, import mindspore.dataset.transforms.py_transforms as py_transforms.

• [STABLE] The column_order parameter in the .map(...) operation is completely removed. You can use .project(...) to adjust the column order.

Bug Fixes

Dataset

• ID6JRL: Fixed an issue where an incorrect data dimension was caused when GeneratorDataset (when the custom dataset __getitem__ returns the self member variable of the dict type) and .batch(batch_size=1) was used.

Contributors

anyrenwei,Bellatan,caifubi,Carey,chaijinwei,changzherui,chengbin,chenshan2623,chujinjin,DavidFFFan,DeshiChen,dingjinshan,fangwenyi,fary86,fengyixing,fuchao,Gaoxiong,gaoyong10,guangpengz,guozhijian,haozhang,hedongdong,hhz886,HighCloud,huangbingjian,huangfuxin,huda,jiangna,jiangshanfeng,jiaorui,jijiarong,laoyu,leida,liangchenghui,LiangZhibo,lichen,lijiajie1234,limingqi107,LiNuohang,litingyu,liubuyu,liuchao,liuchuting,liuluobin,liuyanwei,lizhitong,looop5,luochao60,machangwei,maoyuanpeng1,Margaret_wangrui,mengxian,mwt,NaCN,panzhihui,Qiao_Fu,qiuleilei,qqqhhhbbb,r1chardf1d0,rainyhorse,rogeryu11,shaoshengqi,shen_haochen,shenwei41,shuqian0,tanghuikang,Tianci Xiao,tianxiaodong,wang_ziqi,wangjialin,wangyin,wujueying,wusuqin4,wuyanernuo,XianglongZeng,xiaopeng,xiaotianci,xuzhen,yanghaoran,yangting,yao_yf,yide12,yiguangzheng,yuanqi,yuchaojie,Yuheng Wang,YuJianfeng,Yule100,yuliangbin,YzLi,zhangbuxue,Zhanghanbo,zhanghanLeo,zhangyihui,zhangyinxia,zhangzhen,zhaochenjie,zhengzuohe,zhongmin,ZPaC,zyb,zyli2020,阿琛,曹彤,胡彬,宦晓玲,黄勇,李良灿,李林杰,刘飞扬,刘力力,宋佳琪,王振邦,熊攀,徐子康,严珞珈,杨晓春,张峻源,张栩浩