{
"summary": {
"Cluster": "集群",
"BaseInfo": "基本信息",
"Parallel Strategy Analysis": "并行策略分析",
"ReportFile": "文件路径",
"ReportSize": "文件大小",
"ReportCaptureTime": "文件采集时间",
"DeviceCount": "设备数",
"StepCount": "迭代数量",
"ProfilingSessionDuration": "Profiling数据采集时长",
"Algorithm": "算法",
"AlgorithmTooltip": "混合并行时卡排布算法",
"Dimension": "维度",
"DimensionTooltipContent": [
"(1) 切换并行维度,将展示不同并行组合下卡排布情况,未展示的维度将被折叠,在折叠状态下,排布图中的一个元素代表多个卡,可将鼠标悬停在具体元素上,在提示中了解包含的卡",
"(2) 在 “性能指标” 中选择相应的选项,将在下方并行排布图中展示对应指标的热力图",
"(3) 选中排布图上方的图例,会以框选的方式展示对应的并行分组信息",
"(4) 点击排布图中具体元素,会以连线的方式展示不同并行下通信域信息",
"(5) 点击 (3)、(4) 中出现的框或线,将在下方的 “计算/通信概览” 中展示相应并行分组或通信域的柱状堆叠图、PP流水并行图等信息"
],
"DPDimensionTooltip": "折叠 TP、PP、CP 维度,隐藏相应细节",
"PPDimensionTooltip": "折叠 TP、CP 维度,隐藏相应细节",
"CPDimensionTooltip": "折叠 TP 维度,隐藏相应细节",
"TPDimensionTooltip": "全展开维度,展示全量卡的集群分析结果",
"Parallel Dimension": "并行维度",
"PPSize": "PP大小",
"TPSize": "TP大小",
"DPSize": "DP大小",
"CPSize": "CP大小",
"EPSize": "EP大小",
"MoE-TP Size": "MoE-TP大小",
"Tensor Parallelism": "张量并行",
"Data Parallelism": "数据并行",
"Pipeline Parallelism": "流水线并行",
"Context Parallelism": "上下文并行",
"Expert Parallelism": "专家并行",
"MoE Tensor Parallelism": "MoE 张量并行",
"Tensor/Data Parallelism": "张量/数据并行",
"Generate": "生成",
"Target Index": "目标编号",
"Find": "查找",
"MindIE Size Validate Message": "总卡数(PP × TP × DP)必须大于设备数量",
"Megatron Size Validate Message": "总卡数(PP × TP × CP × DP)必须大于设备数量",
"Computation/CommunicationOverview": "计算/通信概览",
"Computation/CommunicationDescribe": [
"计算时间 = NPU上的内核时间总和",
"计算时间(未被覆盖) = 计算时间 – 计算通信间覆盖时间",
"预处理时间 = 每 Step 起始时刻到第一个计算/通信算子的时间,在 Overlap Analysis 中也被视为 Free",
"空闲时间 = NPU上既不在计算也不在通信的时间,此处并未包括预处理时间",
"总时间 = 计算时间(未被覆盖) + 计算通信间覆盖时间 + 通信时间(未被覆盖)+ 空闲时间,或者 计算时间 + 通信时间(未被覆盖) + 空闲时间",
"总计算占比 = 计算时间 / 总时间",
"通信占比 = 通信时间 / 总时间 =(通信时间(未被覆盖)+ 计算通信间覆盖时间)/ 总时间"
],
"Computation/CommunicationDescribeWithPreparing": [
"计算时间 = NPU上的内核时间总和",
"计算时间(未被覆盖) = 计算时间 – 计算通信间覆盖时间",
"预处理时间 = 每Step起始时刻到第一个计算/通信算子的时间,在Overlap Analysis中也被视为Free",
"空闲时间 = NPU上既不在计算也不在通信的时间,此处并未包括预处理时间",
"总时间 = 预处理时间 + 计算时间(未被覆盖) + 计算通信间覆盖时间 + 通信时间(未被覆盖) + 空闲时间,或者 预处理时间 + 计算时间 + 通信时间(未被覆盖) + 空闲时间",
"计算占比 = 计算时间 / 总时间",
"通信占比 = 通信时间 / 总时间 =(通信时间(未被覆盖)+ 计算通信间覆盖时间)/ 总时间"
],
"Computation/CommunicationLastDescribe": "点击柱状图显示单节点计算/通信详情",
"Index": "序号",
"Name": "名称",
"Step": "迭代ID",
"Baseline Step": "基线迭代ID",
"RankGroup": "通信域",
"OrderBy": "排序方式",
"Top": "前",
"preparing": "预处理",
"Preparing": "预处理时间",
"Total Computing": "计算时间",
"Computing": "计算时间",
"Max Computing": "最大计算时间",
"Min Computing": "最小计算时间",
"Computing Range": "计算时间极差(最大与最小间的差值)",
"Computing(Not Overlapped)": "计算时间(未重叠部分)",
"Sum of Max Computing": "最大计算时间总和",
"Computing/Communication Overlapped": "计算通信间重叠时间",
"Max Computing/Communication Overlapped": "最大计算通信间重叠时间",
"Min Computing/Communication Overlapped": "最小计算通信间重叠时间",
"Communication(Overlapped)": "通信时间(重叠部分)",
"Communication(Not Overlapped)": "通信时间(未重叠部分)",
"Max Communication(Not Overlapped)": "最大通信时间(未重叠部分)",
"Min Communication(Not Overlapped)": "最小通信时间(未重叠部分)",
"Sum of Max Communication(Not Overlapped)": "最大通信时间(未重叠部分)的总和",
"Sum of Min Communication(Not Overlapped)": "最小通信时间(未重叠部分)的总和",
"Communication(Not Overlapped and Exclude Receive)": "通信时间(未重叠部分且不包含Receive算子)",
"Communication": "通信时间",
"Avg Communication": "平均通信时间",
"Max Communication": "最大通信时间",
"Min Communication": "最小通信时间",
"Communication Range": "通信时间极差(最大与最小间的差值)",
"Sum of Max Communication": "最大通信时间的总和",
"Free": "空闲时间",
"Max Free": "最大空闲时间",
"Min Free": "最小空闲时间",
"Free Range": "空闲时间极差(最大与最小间的差值)",
"Sum of Max Free": "最大空闲时间的总和",
"Computing + Communication(Not Overlapped) + Free": "总时间(计算、通信(未重叠部分)、空闲时间的总和)",
"Max Computing + Communication(Not Overlapped) + Free": "最大总时间(计算、通信(未重叠部分)、空闲时间的总和)",
"Min Computing + Communication(Not Overlapped) + Free": "最小总时间(计算、通信(未重叠部分)、空闲时间的总和)",
"Computing + Communication(Not Overlapped) + Free Range": "总时间(计算、通信(未重叠部分)、空闲时间的总和)极差",
"Rank ID": "卡序号",
"Included Ranks": "包含的卡",
"Total Computing Ratio": "总计算占比",
"Computing Ratio": "计算占比",
"Communication Ratio": "通信占比",
"ComputingDetail": "计算详情",
"AcceleratorCore": "加速器核",
"AcceleratorCoreDurations": "加速器核时长",
"Detail": "详情",
"Details": "详情",
"CommunicationDetail": "通信详情",
"CommunicationDurationsOverlapped": "通信时长(重叠部分)",
"CommunicationDurationsNotOverlapped": "通信时长(未重叠部分)",
"OverlappedDurations": "重叠部分时长",
"NotOverlappedDurations": "未重叠部分时长",
"Advice": "专家建议",
"SummaryAdvice": "此集群中{{type}}存在一些问题,因为{{type}}的最大差值达到了{{time}}微秒。",
"Stage": "非空泡时间",
"Bubble": "空泡时间",
"StageTimeAndBubbleTimeDescribe": [
"Stage Time: PP并行时,stage时间代表除receive算子时间外的其他耗时",
"Bubble Time: Bubble时间,指receive时间的总和"
],
"Timeline not fully parsed": "时间线未完全解析",
"GenerateConfirm": "该配置与系统采集的配置不一致,确定要更改吗?",
"NoDataTip": "原始采集的数据中未包含并行策略配置参数,请手动配置",
"Data Type": "数据类型",
"Performance Metric": "性能指标",
"None": "无",
"VisibleRange": "筛选范围",
"Minimum": "最小值",
"Maximum": "最大值",
"Pipeline Parallelism Chart": "流水并行图",
"MoE Expert Load Balancing Analysis": "MoE大模型专家负载均衡分析",
"MoE Expert Load Balancing Analysis Tooltip": "该功能专为大模型推理性能优化设计。系统自动识别Profiling数据中的GroupedMatmul算子,统计其耗时以可视化热力图,助力分析MoE负载均衡情况。此外,支持导入MindIE采集的Dump数据,用于专家热点深度分析。",
"Profiling Expert Data Parsing": "Profiling 专家数据解析中……",
"Model Layer Num": "模型层数",
"Dense Layer List": "非MoE层列表",
"Expert Num": "专家数",
"Model Stage": "模型阶段",
"Data Version": "数据类型",
"Dump unbalanced": "Dump-均衡前",
"Dump balanced": "Dump-均衡后",
"Import data": "导入{{version}} dump 数据",
"Search": "查询",
"slowRanksTopN": "应关注Top3慢卡",
"slowGroupsTopN": "应关注Top3分组",
"dpSynchronizeTime": "DP-平均通信时间差值(us)",
"cpSynchronizeTime": "CP-平均通信时间差值(us)",
"tpSynchronizeTime": "TP-通信时间差值(us)",
"number": "序号",
"slow rank advice": "请关注如下分组/卡是否拖累了整网。同一通信域内,通信时间差值越大,说明集群快慢卡不同步问题越严重。此部分差值可认为,存在一定优化空间。可右键慢卡所在通信域连线,跳转至“通信页签->通信耗时分析” ,进一步确认快慢卡差异来源。",
"No problem": "当前维度无明显快慢卡问题",
"contextMenu": {
"Copy attributes": "复制属性",
"Expand": "展开",
"Collapse": "折叠",
"View Communication Duration Analysis": "查看通信耗时分析",
"Set as Upper Bound": "设为最大筛选值",
"Set as Lower Bound": "设为最小筛选值"
},
"SlowRank ToolTip": {
"slowRanksTopN": "应关注Top3慢卡",
"slowGroupsTopN": "应关注Top3分组",
"dpSynchronizeTime": "所在DP通信域最大'DP-平均通信时间' 减去 当前元素'DP-通信时间'",
"cpSynchronizeTime": "所在CP通信域最大'CP-平均通信时间' 减去 当前元素'CP-通信时间'",
"tpSynchronizeTime": "所在TP通信域最大'TP-通信时间' 减去 当前元素'TP-通信时间'"
}
}
}