5.1.6. PopART changelog

2.6.0+5997

New features

• Improvements for explicit pipelining (support for overlapped IO)

• Support half-precision tensors in scaledVarUpdate

• Allow AddArg0GradOp to change its output tensor type after construction

• Improved support for gradient clipping in accumulate outer fragment paralleliser transform

• Add execution context constraints in AliasModelGrower

• Add RoiAlign operation

• Add init_type to ops.init to allow for uninitialised tensors

• Allow AiGraphcoreOpset1::Reshape to use -1 dim

• Add AiGraphcoreOpset1::slice to mimic Slice-1

• Improved implementation of resize gradient reduceDimension

• Separate load and store landing pad tensors for remote exchanges when required

• Improved support for int16/uint16

• Use LeakyRelu output tensor instead of input to compute the gradient

• Add transform to backup inplace updated tensors when they are required for recomputation

• Add step to verify that users aren’t using modifying (inplace) operations in autodiff

• Add support for custom programs, introduce special custom program for implicit pipeline forward only (experimental)

• Add pass argument to in_sequence in PopXL, to allow transforms to add topological constraints after an operation is created

• Add shape inference to tensor remap operation

• Enable profiling of cached executables. See PopVision documentation

• Add code loading to PopXL. See Graphs

• Add custom operation support to PopXL. See Custom operations

• Add tanh, conv, averagepool, argmin, argmax, exp, histogram, sqrt, maximum, log, onehot and roialign to PopXL. See Supported operations

• Add negative log likelihood loss in PopXL

• Add per-replica variable initialisation and retrieval to PopXL. See Replication

• Add support for torch input tensors in PopXL

• Improved device management in PopXL

• Add .vscode workspace file

• Add argument type check to popxl.Session.get_tensors_data in PopXL

• Add support for enabling engine caching via POPXL_CACHE_DIR environment variable

• Avoid use of deprecated variables in GCL

• Add “zeroInfinity” option and plan option flag “enableReducedClassesInLabel” support to CTC operation

• Add ability to run CTC operation in validation mode

• Switch to new collective balanced reorder API

• Add “disableOptimizerStateTensorStreams” option to selectively disable streaming and storing of optimizer tensors

• Improved handling of setting weights from host and loading weights to the host in PopXL via context managers

Bug Fixes

• Tidy up linting issues

• Fix subgraph pruner used by autodiff

• Fix allreduce logic

• Fix autodiff bugs

• Fix bug related to change in collective balanced reorder padding behaviour

• Fix missing RTSGroup error

• Fix bug in equal in PopXL

• Fix partialTypeMatMuls support in PopXL

• Fix for torch linear mode test

• Fix for misssing pipelineStage attribute

• Reload engine and connect streams on every re-attach through popxl.Session context manager

• Fix bug where VariableSettings CommGroup is not respected by AllReduce in gradient and accumulator reduction

• Fix for random number state management when replicated graph option is set

• Fix alias zero copy setting verification

• Fix non-determinism bugs in accumulate outer fragment paralleliser and multi collective transforms

• Fix explicit recompute (annotation issue and recompute to recompute connections)

• Fix for resize gradient operation

• Fix mechanism to write variable data when using a cached binary in PopXL

• Various fixes for executable caching (added missing tensors, random seed, anchors, and various bug fixes)

• Correct incorrect state-tensor initial vector dimensions

• Revert inplace WhereOp to outplace when it is not parallel writable

• Don’t overwrite the number of tiles to 4 when a custom IPU model config is used

• Print before erase to avoid memory error

Optimisations

• Reduce number of dummy graph objects constructed for lowering MatMul operations

• Remove cases of a clone followed by a mapTensorLinearly

• Remove unnecessary uses of mapTensorLinearly

• Optimisations in parsing ONNX protobuf files

• Speed up ignore applyInplacePattern by ignoring graph.isSchedulable

• Faster & more memory efficient implementation of cubic resize

• Take compilation-affecting engine options into account when calculating hashes for the purpose of executable caching

Logging and documentation

• General improvement of PopXL user guide (including sessions, remote variables, replica grouping, code loading, custom operations, links to related PyTorch Numpy and ONNX operations, MNIST example)

• General improvements to PopART API documentation

• Error message improvements

2.5.1

New features

• Add PopXL API (experimental)

• Add support for RNN operator (preview)

• Improvements to automatic loss scaling (experimental)

• Add improved ability to manage PRNG behaviour across replicas (experimental)

• Add ability to retrieve random seed

• Add an overload of Builder.setAvailableMemoryProportion which can target multi-output nodes

• Ensure initial inputs of gradient graphs match any user-specified provided grads

• Ensure outputs of gradient graphs match any user-specified required grads

• Add ability to run exported models using the Poplar Triton Backend via PopEF integration

• Add visualisations for inplace modified and aliased tensors and graph inputs and outputs to Dot visualizer

• DynamicSliceOp and DynamicUpdateOp can drop the first dimension of the slice if it is 1

• Support AnchorReturnType::Final in MainLoops transform

• Improved replicated tensor sharding (RTS) compatibility for operations

• Make gradient clipping compatible with replicated tensor sharding (RTS)

• Improved linter support

• Add ability to show ONNX model proto in human readable text

• Various improvements to executable caching

• Add ability to perform per-replica reads and writes of variable values

• Improved quality of debug information

• Use slice plan in SparseAccumulateOpx

• Add ability to merge collective operations

• Add ability to dynamically switch off the backwards pass when using implicit pipelining

• Add ability to refresh engine cache on-the-fly

Bug Fixes

• Fix the logic that replaces DropoutOp with IdentityOp

• Improve device handling in tests

• Fix for potential deadlock condition in test runner

• Fix in lowering logic for trailing subgraph parts that contain only calls to child subgraph parts

• IdentityLossOpx will no longer attempt to unwind (resulting in an error) when there is a reduction

• Fix subgraph autodiff logic

• Allow CallOp to not have outputs connected for all of its callee outputs

• Fix Python binding for DeviceManager::tryAttachUntilTimeout

• Correctly promote inplace aliased and modified tensors through the Loop operation

• Fix unwinding through multiple consecutive slice operations

• Fix unwinding issue in MaxOpx

• Enable bufferingDepth to be used when SessionOptions::enablePrefetchDatastreams isn’t set

• Fix dtype clone in SparseAccumulateOpx::createInputTensor

• Fix bug in ReplicatedTensorShardingTracer

• Fix compile error if accl2 type is not FLOAT

• Fix PowArg0GradOpPattern for fp16

Optimisations

• Allow non-broadcasted indices as an input to the scatterreduce operation

• Add ExpandCast pattern to reverse the order of an expand followed by cast to reduce memory footprint

• Add inplace versions of WhereOp

• Allow IdentityInplaceOp to unwind, reducing memory use when it cannot be made inplace

• Split operators_test in two

• Add TensorRemapOp for point-fixes of bad tensor layouts

• Explicit recomputation support for pipelining

• Alias zero copy tracks variables and multi-context tensors less conservatively

• Improve graph traversal through loop-carried tensors

Logging and documentation

• Add compile-time option to log device access events to a file

• Improved CommGroupType::None comments

• Fix code listings

• Update to internal documentation build system

• Various small user guide and API improvements

• Add how to execute imported model to documentation