Metal PluggableDevice: op coverage tracking toward CUDA parity

Tracking issue for bringing the in-tree Metal PluggableDevice backend added in #126254 to parity with the CUDA op set.

Where things stand

Ops registered for Metal	116
Ops with a CUDA GPU kernel in tensorflow/core/kernels/	277 explicit, 534 counting macro registrations
Remaining in this list	165

The 116 already in place cover a convolutional classifier end to end: Conv2D and its gradients, pooling, FusedBatchNorm v1 to v3 with gradients, activations, the softmax cross entropies, a broadcasting arithmetic and maths library with matching gradient ops, BatchMatMul, reductions, Slice, Pad, Split, Concat, the random initialisers, and the SGD, Adam, Momentum and RMSProp updates.

Method

Every op below is implemented against MPSGraph or a Metal compute shader and checked on device against a CPU reference before it is counted. Gradients are additionally checked against central finite differences, which is how the FusedBatchNormGrad batch-coupling terms and the four unary gradient formulas were validated.

Ops TensorFlow registers for DEVICE_DEFAULT in a form a pluggable device can actually use are excluded from this list, since the backend inherits them. Ops whose DEVICE_DEFAULT registration is an int32-in-host-memory special case, which is the pattern used for Sum, Slice, Pad, Fill and ZerosLike, are not excluded, because a float version on device is genuinely missing.

Also excluded as out of scope for a device backend: collectives and NCCL, dataset and iterator ops, TensorList and TensorArray, the CSR sparse matrix family, cuDNN-specific RNN entry points, and the FFT family.

Remaining

Convolution and pooling (24)

• AvgPoolGrad
• Conv
• Conv3D
• Conv3DBackpropFilter
• Conv3DBackpropFilterV2
• Conv3DBackpropInput
• Conv3DBackpropInputV2
• DepthwiseConv2dNative
• DepthwiseConv2dNativeBackpropFilter
• DepthwiseConv2dNativeBackpropInput
• Dilation2D
• Dilation2DBackpropFilter
• Dilation2DBackpropInput
• ExtractImagePatches
• ExtractVolumePatches
• LRN
• LRNGrad
• MaxPoolGradGrad
• MaxPoolGradGradV2
• MaxPoolGradGradWithArgmax
• MaxPoolGradV2
• MaxPoolGradWithArgmax
• MaxPoolV2
• MaxPoolWithArgmax

Image (19)

• AdjustContrast
• AdjustContrastv2
• AdjustHue
• AdjustSaturation
• CropAndResize
• CropAndResizeGradBoxes
• CropAndResizeGradImage
• GenerateBoundingBoxProposals
• HSVToRGB
• ImageProjectiveTransformV2
• ImageProjectiveTransformV3
• NonMaxSuppressionV2
• NonMaxSuppressionV3
• NonMaxSuppressionV4
• RGBToHSV
• ResizeBilinear
• ResizeBilinearGrad
• ResizeNearestNeighbor
• ResizeNearestNeighborGrad

Indexing and shape (45)

• BatchMatrixBandPart
• BatchMatrixDiag
• BatchMatrixDiagPart
• BatchMatrixSetDiag
• BatchMatrixTriangularSolve
• BatchToSpace
• BatchToSpaceND
• DepthToSpace
• Diag
• DiagPart
• DynamicPartition
• DynamicStitch
• InTopK
• InTopKV2
• LowerBound
• MatrixBandPart
• MatrixDiag
• MatrixDiagPart
• MatrixDiagPartV2
• MatrixDiagPartV3
• MatrixDiagV2
• MatrixDiagV3
• MatrixSetDiag
• MatrixSetDiagV2
• MatrixSetDiagV3
• MirrorPadGrad
• OneHot
• ParallelDynamicStitch
• ResourceGather
• ResourceGatherNd
• ResourceScatterUpdate
• Reverse
• ReverseSequence
• Roll
• SpaceToBatch
• SpaceToBatchND
• SpaceToDepth
• StridedSlice
• StridedSliceGrad
• TileGrad
• TopK
• TopKV2
• Unique
• UniqueWithCounts
• UpperBound

Reduction and scan (8)

• Bucketize
• Cumprod
• Cumsum
• CumulativeLogsumexp
• DenseBincount
• EuclideanNorm
• HistogramFixedWidth
• L2Loss

Elementwise and comparison (9)

• ApproximateEqual
• CheckNumerics
• CheckNumericsV2
• ClipByValue
• Conj
• ConjugateTranspose
• Cross
• LinSpace
• PopulationCount

Normalisation (2)

• BatchNormWithGlobalNormalization
• BatchNormWithGlobalNormalizationGrad

Recurrent (10)

• BlockLSTM
• BlockLSTMGrad
• BlockLSTMGradV2
• BlockLSTMV2
• CTCLoss
• CTCLossV2
• GRUBlockCell
• GRUBlockCellGrad
• LSTMBlockCell
• LSTMBlockCellGrad

Linear algebra (4)

• Lu
• MatrixTriangularSolve
• Qr
• SelfAdjointEigV2

Random (8)

• Multinomial
• ParameterizedTruncatedNormal
• RandomGamma
• RandomUniformInt
• StatelessMultinomial
• StatelessParameterizedTruncatedNormal
• StatelessRandomGammaV2
• StatelessRandomGammaV3

Quantisation (11)

• FakeQuantWithMinMaxArgs
• FakeQuantWithMinMaxArgsGradient
• FakeQuantWithMinMaxVars
• FakeQuantWithMinMaxVarsGradient
• FakeQuantWithMinMaxVarsPerChannel
• FakeQuantWithMinMaxVarsPerChannelGradient
• QuantizeAndDequantize
• QuantizeAndDequantizeV2
• QuantizeAndDequantizeV3
• QuantizeAndDequantizeV4
• QuantizeAndDequantizeV4Grad

Misc (25)

• ApproxTopK
• BatchFFT
• BatchFFT2D
• BatchFFT3D
• BatchIFFT
• BatchIFFT2D
• BatchIFFT3D
• BiasAddV1
• FFT
• FFT2D
• FFT3D
• FFTND
• IFFT
• IFFT2D
• IFFT3D
• IFFTND
• IRFFT
• IRFFT2D
• IRFFT3D
• IRFFTND
• ParallelConcat
• RFFT
• RFFT2D
• RFFT3D
• RFFTND

---

Related: #126254 adds the backend. #126275 fixes an unrelated llvm-raw patch failure that currently blocks bazel build on a clean checkout.

Drafted with assistance from Claude Opus 5.