Release Notes

MindSpore 2.5.0 Release Notes

Major Features and Improvements

Distributed Startup Component msrun

• [STABLE] msrun supports passing in the hostname of the node (e.g. localhost) as -master_addr, which improves the ease of use of msrun.

• [STABLE] msrun supports printing training logs to standard output. Users can control which ranks to print with the -tail_worker_log parameter.

• [STABLE] After setting export VLOG_v=12500, the scheduler log can output cluster information, which helps users to quickly count cluster data.

• [STABLE] msrun supports formatting the log file name with the --worker_log_name parameter to help users quickly locate the problem node.

For details, refer to msrun Launching.

Profiler

• [STABLE] New interfaces mindspore.profiler.schedule and mindspore.profiler.tensor_board_trace_handler are added to support the acquisition and rendering of PyNative scenarios by step, which improves the ease of use of PyNative scenarios.

• [STABLE] Dynamic Profiling supports customized for loops to enhance the ease of use of dynamic graphical scenarios.

• [STABLE] Profiler initialization parameters and deliverables directory structure aligned to PTA to reduce user migration difficulty.

• [STABLE] A new lightweight interface, mindspore.profiler.mstx, has been added to provide users with a low-overhead performance data collection method.

• [STABLE] Timeline supports displaying hardware utilization data to help users locate downclocking issues.

For details, refer to Ascend Performance Tuning.

PyNative

• [Beta] PyNative mode supports the inplace operator process, introduces the inplace operators. Taking the mindspore.mint.nn.functional.relu as an example, if you want to use the inplace updated version of the relu operator, you can call the mindspore.mint.nn.functional.relu_.

• [STABLE] Enable the environment variable MS_SIMULATION_LEVEL=1 to enable the PyNative dryrun, the multi-device process can be simulated without occupying the device, and the display memory usage can be viewed in logs. For details, refer to Environment Variables。

FrontEnd

• [STABLE] Added mindspore.nn.utils.no_init_parameters API, which supports delayed initialization of network parameters and reduces model startup time in inference scenarios.

API Change

New APIs

• [DEMO] mindspore.mint API provides more functional, nn interfaces. The mint interface is currently an experimental interface and performs better than ops in O0/O1 and pynative mode. Currently, the O2 compilation mode (graph sinking mode) and CPU/GPU backend are not supported, and it will be gradually improved in the future.

  mindspore.mint
  mindspore.mint.bernoulli	mindspore.mint.bincount	mindspore.mint.clone	mindspore.mint.einsum
  mindspore.mint.empty	mindspore.mint.empty_like	mindspore.mint.full_like	mindspore.mint.randint
  mindspore.mint.randint_like	mindspore.mint.randn	mindspore.mint.randn_like	mindspore.mint.randperm
  mindspore.mint.chunk	mindspore.mint.concat	mindspore.mint.count_nonzero	mindspore.mint.scatter
  mindspore.mint.select	mindspore.mint.squeeze	mindspore.mint.swapaxes	mindspore.mint.transpose
  mindspore.mint.triu	mindspore.mint.unbind	mindspore.mint.unique_consecutive	mindspore.mint.multinomial
  mindspore.mint.addmv	mindspore.mint.diff	mindspore.mint.exp2	mindspore.mint.float_power
  mindspore.mint.fix	mindspore.mint.fmod	mindspore.mint.frac	mindspore.mint.lerp
  mindspore.mint.log2	mindspore.mint.log10	mindspore.mint.logaddexp	mindspore.mint.mv
  mindspore.mint.nansum	mindspore.mint.nan_to_num	mindspore.mint.polar	mindspore.mint.ravel
  mindspore.mint.outer	mindspore.mint.softmax	mindspore.mint.t	mindspore.mint.cdist
  mindspore.mint.amax	mindspore.mint.amin	mindspore.mint.cumprod	mindspore.mint.histc
  mindspore.mint.logsumexp	mindspore.mint.norm	mindspore.mint.std	mindspore.mint.std_mean
  mindspore.mint.var	mindspore.mint.var_mean	mindspore.mint.allclose	mindspore.mint.argsort
  mindspore.mint.equal	mindspore.mint.isinf	mindspore.mint.isneginf	mindspore.mint.not_equal
  mindspore.mint.addbmm	mindspore.mint.addmm	mindspore.mint.baddbmm	mindspore.mint.dot
  mindspore.mint.meshgrid	mindspore.mint.mm

  mindspore.mint.nn
  mindspore.mint.nn.Conv3d	mindspore.mint.nn.ConstantPad1d
  mindspore.mint.nn.ConvTranspose2d	mindspore.mint.nn.ConstantPad2d
  mindspore.mint.nn.BatchNorm1d	mindspore.mint.nn.ConstantPad3d
  mindspore.mint.nn.BatchNorm2d	mindspore.mint.nn.ReflectionPad1d
  mindspore.mint.nn.BatchNorm3d	mindspore.mint.nn.ReflectionPad2d
  mindspore.mint.nn.LayerNorm	mindspore.mint.nn.ReflectionPad3d
  mindspore.mint.nn.SyncBatchNorm	mindspore.mint.nn.ReplicationPad1d
  mindspore.mint.nn.ELU	mindspore.mint.nn.ZeroPad1d
  mindspore.mint.nn.GELU	mindspore.mint.nn.ZeroPad2d
  mindspore.mint.nn.LogSigmoid	mindspore.mint.nn.ZeroPad3d
  mindspore.mint.nn.ReLU6	mindspore.mint.nn.BCELoss
  mindspore.mint.nn.SiLU	mindspore.mint.nn.CrossEntropyLoss
  mindspore.mint.nn.Tanh	mindspore.mint.nn.NLLLoss
  mindspore.mint.nn.Embedding	mindspore.mint.nn.SmoothL1Loss
  mindspore.mint.nn.Dropout2d	mindspore.mint.nn.Upsample
  mindspore.mint.nn.AdaptiveAvgPool1d	mindspore.mint.nn.MaxUnpool2d
  mindspore.mint.nn.AdaptiveAvgPool2d

  mindspore.mint.nn.functional
  mindspore.mint.nn.functional.adaptive_avg_pool1d
  mindspore.mint.nn.functional.adaptive_avg_pool2d
  mindspore.mint.nn.functional.avg_pool1d
  mindspore.mint.nn.functional.max_unpool2d
  mindspore.mint.nn.functional.logsigmoid
  mindspore.mint.nn.functional.relu6
  mindspore.mint.nn.functional.relu_
  mindspore.mint.nn.functional.normalize
  mindspore.mint.nn.functional.dropout2d
  mindspore.mint.nn.functional.nll_loss
  mindspore.mint.nn.functional.smooth_l1_loss
  mindspore.mint.nn.functional.interpolate
  mindspore.mint.nn.functional.conv3d

  mindspore.mint.distributed
  mindspore.mint.distributed.all_gather	mindspore.mint.distributed.get_global_rank
  mindspore.mint.distributed.all_gather_into_tensor	mindspore.mint.distributed.get_group_rank
  mindspore.mint.distributed.all_gather_object	mindspore.mint.distributed.get_process_group_ranks
  mindspore.mint.distributed.all_reduce	mindspore.mint.distributed.init_process_group
  mindspore.mint.distributed.all_to_all	mindspore.mint.distributed.irecv
  mindspore.mint.distributed.all_to_all_single	mindspore.mint.distributed.isend
  mindspore.mint.distributed.barrier	mindspore.mint.distributed.new_group
  mindspore.mint.distributed.batch_isend_irecv	mindspore.mint.distributed.P2POp
  mindspore.mint.distributed.broadcast	mindspore.mint.distributed.recv
  mindspore.mint.distributed.broadcast_object_list	mindspore.mint.distributed.reduce
  mindspore.mint.distributed.gather	mindspore.mint.distributed.reduce_scatter
  mindspore.mint.distributed.gather_object	mindspore.mint.distributed.reduce_scatter_tensor
  mindspore.mint.distributed.get_backend	mindspore.mint.distributed.scatter
  mindspore.mint.distributed.scatter_object_list	mindspore.mint.distributed.send

  others
  mindspore.mint.optim.Adam
  mindspore.mint.linalg.matrix_norm
  mindspore.mint.linalg.norm
  mindspore.mint.linalg.vector_norm
  mindspore.mint.special.exp2

• [STABLE] Two inference API mindspore.ops.incre_flash_attention and mindspore.ops.prompt_flash_attention are added. Currently, only Ascend backend is supported.

• [STABLE] mindspore.runtime replaces the original mindspore.hal interfaces and provides interfaces related to runtime resources such as stream, memory, and event.

• [STABLE] mindspore.device_context replaces some parameters of the original set_context interface and provides setting interfaces related to hardware platform.

• [DEMO] mindspore.Tensor API provides more Tensor methods. Currently the Tensor interfaces are still the experimental interfaces, and do not support the graph sink mode and CPU, GPU back-end, will be gradually improved. In addition, a large number of existing Tensor methods, including operators like +=, -=, *= and /=, have been adapted with Aclnn kernels on the Ascend backend through overloading. Details can be found in API list in official website.

Backwards Incompatible Change

• For API mindspore.Tensor.new_ones, the input argument "size" no longer supports the data type of Tensor.

• mindspore.Profiler removes timeline_limit, rank_id, analyze_only, env_enable parameters.

• Interface name: mindspore.Profiler

  Changes: profile_communication is deprecated and the communication matrix data is collected by setting profiler_level=ProfilerLevel.Level1 or profiler_level=ProfilerLevel.Level2.

  Clarification: The default value of profiler_level is ProfilerLevel.Level0.

  original interface	v2.5.0 interface
  Profiler(profile_communication=True)	Profiler(profiler_level=ProfilerLevel.Level1) or Profiler(profiler_level=ProfilerLevel.Level2)

• Interface name: mindspore.Profiler

  Changes: op_time is deprecated, set to collect NPU-side operator performance data by setting activaties=[mindspore.profiler.ProfilerActivity.NPU].

  Clarification: The activaties parameter is of type list, and as long as it contains the mindspore.profiler.ProfilerActivity.NPU parameter, it means that the collection of NPU-side operator performance data is enabled, and the collection is turned on by default.

  original interface	v2.5.0 interface
  Profiler(op_time=True)	Profiler(activaties=[mindspore.profiler.ProfilerActivity.NPU])

• Interface name: mindspore.Profiler

  Changes: The type of aicore_metrics changed from int to mindspore.profiler.AicoreMetrics enum value.

  Clarification: The default value of aicore_metrics is mindspore.profiler.AicoreMetric.AiCoreNone.

  original interface	v2.5.0 interface
  Profiler(aicore_metrics=0)	Profiler(aicore_metrics=mindspore.profiler.AicoreMetric.AiCoreNone)

• Interface name: mindspore.Profiler

  Changes: profile_framework is deprecated and the frame measurements are collected by setting activaties=[mindspore.profiler.ProfilerActivity.CPU].

  Clarification: activaties parameter type is list, as long as it contains mindspore.profiler.ProfilerActivity.CPU parameter means enable to collect the framework performance data, the default on the collection.

  original interface	v2.5.0 interface
  Profiler(profile_framework="all")	Profiler(activaties=[mindspore.profiler.ProfilerActivity.CPU])

Contributors

baishanyang ,bantao ,Bellatan ,biangelin ,BigSkySea ,caifubi ,candanzg ,candyhong ,Carey ,cccc1111 ,chaijinwei ,changzherui ,chengbin ,chengfeng27 ,chengxb7532 ,chujinjin ,coder2237 ,czrz ,dairenjie ,DavidFFFan ,DeshiChen ,dingjinshan ,ehaleva ,Erpim ,fary86 ,fengyixing ,ffmh ,fuchao ,fuhouyu ,gaoyong10 ,geyuhong ,guoyuzhe ,GuoZhibin ,guozhijian ,halo ,hangq ,haozhang ,hedongdong ,hehongzhe ,hhz886 ,HighCloud ,huangbingjian ,HuangLe02 ,huangziling ,huda ,Huilan Li ,hujiahui8 ,jiahaochen666 ,jiangchao_j ,jiangchenglin3 ,jiangshanfeng ,jiaorui ,jiaxueyu ,jizewei ,jjfeing ,JoeyLin ,jshawjc ,kakyo82 ,kingxian ,kisnwang ,leida ,liangchenghui ,lianghongrui ,LiangZhibo ,lichen ,limingqi107 ,LINH ,linux ,lionelchang ,lishanni ,liubuyu ,liujunzhu ,liuluobin ,liuxu ,liuyanwei ,liyan2022 ,LLLRT ,looop5 ,luochao60 ,luoxuewei ,luoyang ,lyk ,machenggui ,maoyuanpeng1 ,Margaret_wangrui ,master,mengxian ,MengXiangyu ,mengyuanli ,Mrtutu ,mylinchi ,NaCN ,Nikanuo ,niujunhao ,panzhihui ,pengqi ,PingqiLi ,pipecat ,qiuleilei ,qiuyufeng ,qiuzhongya ,r1chardf1d0 ,shaoshengqi ,shen_haochen ,shenhaojing ,shenwei41 ,shilishan ,shiro-zzz ,shuqian0 ,St.Universe ,stavewu ,superxf ,suteng ,TAJh ,tanghuikang ,tangmengcheng ,tan-wei-cheng ,tianxiaodong ,TuDouNi ,TYWZ22259 ,user_0145 ,VectorSL ,vincen45 ,wang_ziqi ,wangshaocong ,wangwensheng4 ,weiyang ,wtcheng ,wtobill ,wujiangming ,wujueying ,wuweikang ,wwwbby ,XianglongZeng ,xiaopeng ,xiaotianci ,xiaoyao ,xiedejin1 ,XinDu ,xuxinglei ,yang guodong ,yangben ,yanghaoran ,yanglong ,yanx ,Yanzhi_YI ,yao_yf ,yefeng ,Yi_zhang95 ,yide12 ,yihangchen ,YijieChen ,YingtongHu ,ylw ,yonibaehr ,yuanqi ,yuchaojie ,yuezenglin ,YuJianfeng ,yyuse ,Zhang QR ,zhangbuxue ,zhangdanyang ,zhanghaibo ,zhangminli ,zhangyinxia ,ZhangZGC ,zhangzhen ,zhengzuohe ,zhouyaqiang0 ,zhuguodong ,zichun_ye ,zlq2020 ,zong_shuai ,ZPaC ,zyli2020 ,陈一 ,程超 ,冯一航 ,胡彬 ,宦晓玲 ,黄勇 ,简云超 ,康伟 ,李栋 ,李良灿 ,李林杰 ,李寅杰,刘崇鸣 ,刘力力 ,刘思铭 ,刘涛Liu ,刘勇琪 ,刘子涵 ,吕浩宇 ,吕凯盟 ,梅飞要 ,倪轩 ,任新 ,十一雷 ,孙昊辰 ,王禹程 ,王振邦 ,熊攀 ,俞涵 ,虞良斌 ,张栩浩 ,赵文璇 ,周莉莉 ,周一航 ,邹文祥