DIPU v0.2.0 Release Note
Release Note
DIPU v0.2.0 Release Notes
大家好！我们很高兴的宣布 DIPU v0.2.0版本正式发布！
DIPU 是一套能够将训练设备快速接入训练框架的 Pytorch 扩展包。
版本号
v0.2.0
核心功能
- 多硬件接入机制：实现一组函数接口即可低成本、快捷接入新硬件Runtime，包含设备接口（
dipu/torch_dipu/csrc_dipu/runtime/device/deviceapis.h）和通信接口（dipu/torch_dipu/csrc_dipu/runtime/device/diclapis.h）； - 高效率算子接入机制：可根据算子 schema 等信息自动生成调用 DIOPI 函数的 wrapper 代码和函数注册代码；
- 算子精度自动对比 autocompare 功能: 打开该功能后算子调用时可同时调用 CPU 算子实现，并对设备计算结果进行精度对比，可快速定位算子精度问题；
- 算子参数获取功能：设置环境变量即可查看算子所有参数的信息（如tensor的shape，stride，dtype，memformat，tensor值等）和函数调用信息；
- 强大的 fallback 功能：增强了原生 Pytorch fallback 能力，支持了运行时控制算子强制 fallback 的能力，支持了未适配函数自动 fallback 的能力；
- 适配了191个 DIOPI 函数，支持了229个 Aten 函数（实际注册了的函数）；
- 提供了在各设备上质量保障的基本能力；
- 提供了多机多卡训练相关接口和能力。
文档
新支持以下 DIOPI 函数
- diopiAbs
- diopiAbsInp
- diopiAdaptiveAvgPool2d
- diopiAdaptiveAvgPool2dBackward
- diopiAdd
- diopiAddInp
- diopiAddInpScalar
- diopiAddScalar
- diopiAddcdiv
- diopiAddcdivInp
- diopiAddcmul
- diopiAddcmulInp
- diopiAddmm
- diopiAll
- diopiAny
- diopiArange
- diopiArgmax
- diopiBCELoss
- diopiBCELossBackward
- diopiBaddbmm
- diopiBaddbmmInp
- diopiBatchNorm
- diopiBatchNormBackward
- diopiBitwiseAnd
- diopiBitwiseAndInp
- diopiBitwiseAndInpScalar
- diopiBitwiseAndScalar
- diopiBitwiseNot
- diopiBitwiseNotInp
- diopiBmm
- diopiCTCLoss
- diopiCTCLossBackward
- diopiCat
- diopiCdist
- diopiCdistBackward
- diopiClamp
- diopiClampInp
- diopiClampInpScalar
- diopiClampMaxScalar
- diopiClampMinScalar
- diopiClampScalar
- diopiConvolution2d
- diopiConvolution2dBackward
- diopiCos
- diopiCosInp
- diopiCrossEntropyLoss
- diopiCrossEntropyLossBackward
- diopiCumsum
- diopiDiv
- diopiDivInp
- diopiDivInpScalar
- diopiDivScalar
- diopiDropout
- diopiDropoutInp
- diopiEq
- diopiEqInp
- diopiEqInpScalar
- diopiEqScalar
- diopiErf
- diopiErfInp
- diopiExp
- diopiExpInp
- diopiFill
- diopiFlip
- diopiFloor
- diopiFloorInp
- diopiGather
- diopiGe
- diopiGeInp
- diopiGeInpScalar
- diopiGeScalar
- diopiGelu
- diopiGeluBackward
- diopiGt
- diopiGtInp
- diopiGtInpScalar
- diopiGtScalar
- diopiHardswish
- diopiHardswishBackward
- diopiHardswishInp
- diopiHardtanh
- diopiHardtanhBackward
- diopiHardtanhInp
- diopiIndex
- diopiIndexPut
- diopiIndexSelect
- diopiLayerNorm
- diopiLayerNormBackward
- diopiLe
- diopiLeInp
- diopiLeInpScalar
- diopiLeScalar
- diopiLeakyRelu
- diopiLeakyReluBackward
- diopiLeakyReluInp
- diopiLinear
- diopiLinearBackward
- diopiLinspace
- diopiLog
- diopiLog2
- diopiLog2Inp
- diopiLogInp
- diopiLogSoftmax
- diopiLogSoftmaxBackward
- diopiLogicalAnd
- diopiLogicalAndInp
- diopiLt
- diopiLtInp
- diopiLtInpScalar
- diopiLtScalar
- diopiMSELoss
- diopiMSELossBackward
- diopiMaskedFill
- diopiMaskedFillInp
- diopiMaskedFillInpScalar
- diopiMaskedFillScalar
- diopiMaskedSelect
- diopiMatmul
- diopiMax
- diopiMaxAll
- diopiMaxPool2d
- diopiMaxPool2dBackward
- diopiMaxPool2dWithIndices
- diopiMaximum
- diopiMean
- diopiMin
- diopiMinAll
- diopiMinimum
- diopiMm
- diopiMul
- diopiMulInp
- diopiMulInpScalar
- diopiMulScalar
- diopiMultinomial
- diopiNLLLoss
- diopiNLLLossBackward
- diopiNe
- diopiNeInp
- diopiNeInpScalar
- diopiNeScalar
- diopiNeg
- diopiNegInp
- diopiNonzero
- diopiNorm
- diopiNormalInp
- diopiPolar
- diopiPow
- diopiPowInp
- diopiPowInpTensor
- diopiPowScalar
- diopiPowTensor
- diopiRandomInp
- diopiRandperm
- diopiReciprocal
- diopiRelu
- diopiReluInp
- diopiRemainderTensor
- diopiRepeat
- diopiRoll
- diopiRsqrt
- diopiScatter
- diopiScatterScalar
- diopiSigmoid
- diopiSigmoidBackward
- diopiSilu
- diopiSin
- diopiSinInp
- diopiSoftmax
- diopiSoftmaxBackward
- diopiSort
- diopiSqrt
- diopiSqrtInp
- diopiStack
- diopiStd
- diopiSub
- diopiSubScalar
- diopiSum
- diopiTanh
- diopiTanhBackward
- diopiTanhInp
- diopiThresholdBackward
- diopiTopk
- diopiTranspose
- diopiTril
- diopiUniformInp
- diopiUnique
- diopiUpsampleLinear
- diopiUpsampleLinearBackward
- diopiUpsampleNearest
- diopiUpsampleNearestBackward
- diopiWhere