15. Changelog

15.1. v3.0 (Poplar SDK 3.0)

15.1.1. New features

Support for tensor.new_full
Improved performance of index_put when the indices are a one dimensional vector.

15.1.2. API changes

Use the dispatcher instead of torch.jit.trace() by default to capture the graph on supported platforms (see Section 11.1, Dispatcher support). Use poptorch.Options.Jit.traceModel(True) to revert to previous behaviour.

15.1.3. Bug Fixes

Fixed BatchNorm running statistics, now using the unbiased estimator to update running_var at training time.
Fixed scalar tensor inputs to the graph not working.
Fixed expand, when the desired shape contains both added dimensions and -1.

15.2. v2.6 (Poplar SDK 2.6)

15.2.1. New features

Improved performance of torch.gather in some cases where the index tensor has come from an expand or expand_as.
Improved error message when trying to apply bitwise ops to unsupported scalar types.
Support for upsample bicubic mode.
Support for zero_infinity in torch.nn.CTCLoss.
Experimental support for Torch’s dispatcher as an alternative to torch.jit.trace() (see Section 11.1, Dispatcher support).
Improved performance by compiling built-in custom ops at install time.

15.2.2. API changes

Data loader: default process creation strategy for worker process changed from “spawn” to “forkserver”.

15.2.3. Bug Fixes

Fixed remaining in-place operations on slices
Fixed einsum transpose error
Fixed floating point exception in torch.Tensor.exponential_ and torch.distributions.Exponential
Improved support for torch.int16 tensors.

15.3. v2.5 (Poplar SDK 2.5)

15.3.1. New features

Ignore missing values when reloading an Optimizer state.
Support saving Optimizer states when compiling offline.
Support for the following functions:
- torch.var
- torch.std
- torch.var_mean
- torch.std_mean
Also save the random number generator’s state and the seed when saving a model.
Support for col2im (used by torch.nn.Fold).
Improve error message of aten::index, aten::index_put_ when indexing with boolean tensor masks.
Support for torch.argsort.
Support for torch.nn.RNN.
Add support for __repr__ in PoplarExecutor.
For models annotated with BeginBlock, show the IPU blocks in repr(model).
Improve implementation of torch.scatter_add
Support for torch.nn.utils.weight_norm
Support for torch.randperm
Support for torch.nn.functional.cosine_similarity and torch.nn.CosineSimilarity
Support for torch.all, torch.any, torch.Tensor.all and torch.Tensor.any
Support for torch.Tensor.exponential_ and torch.distributions.Exponential

15.3.2. API changes

Removed poptorch.AnchorMode, poptorch.Options.anchorMode which were deprecated in favour of poptorch.OutputMode and poptorch.Options.outputMode respectively.

15.3.3. Bug Fixes

Fix thread safety issue in LogContext.
Fix torch.clamp with integer tensors
Fix in-place modification of slices
Fix torch.index_put_ when operating on slices
Fix torch.chunk when dim size is indivisible by the specified number of chunks
Fix cases where tensor.half() was in-place
Fix tracing with half buffers
Fix for loops with in-place ops
Fix torch.flip with negative indices
Fix masked assign when using tensor indexing syntax
Fix some cases where use of serializedMatMul was ignored or resulted in errors.

15.4. v2.4 (Poplar SDK 2.4)

15.4.1. New features

Support for deepcopy functionality in poptorch.Options class
Added functionality to add a name scope for each operator present in the module. This function is enabled by default. It can be disabled using poptorch.Options.disableModuleNamescope.
Support for a greater number of convolution and transpose convolution parameters including those which result in input/kernel/output truncation, either for inference (transpose) or gradient calculation.
Migrated to PyTorch version 1.10.0
Support for gradient clipping by norm in poptorch.optim optimizers
Support saving and restoring internal optimiser state with PopTorch optimisers via optimizer.state_dict() and optimizer.load_state_dict()
Add removeBlocks function to remove block annotations from a Model / Layer.
Support for CPU ops using poptorch.CPU.
Support for im2col (used by torch.nn.Unfold).
Make optimizers work with LR schedulers.
Switched to gold linker by default.

15.4.2. API changes

Deprecated poptorch.Options.anchorMode in favour of poptorch.Options.outputMode
Deprecated poptorch.Options.defaultAnchorMode in favour of poptorch.Options.defaultOutputMode
Deprecated poptorch.AnchorMode in favour of poptorch.OutputMode

15.4.3. Bug Fixes

Fixed incorrect gradient when using torch.nn.Embedding with padding_idx

15.5. v2.3 (Poplar SDK 2.3)

15.5.1. New features

Support for torch.bitwise_and, torch.bitwise_or, torch.bitwise_xor
Support for torch.logical_and, torch.logical_or,
Support K-dimensional NLLLoss, K-dimensional CrossEntropyLoss
Support for non-default affine parameter flags in normalisation ops
Support for torch.Tensor.T
Support for torch.bool in torch.zeros, torch.zeros_like, torch.ones, torch.ones_like
Support for torch.scatter and its in-place variant
Support for in-place modification to buffers on IPU
Support for taking slices of scalars
Support version of bilinear upsampling specifying intended output size instead of scale factors
Add support for overlapping host IO on inputs via poptorch.set_overlap_for_input().
Add option for setting number of IO tiles via numIOTiles in poptorch.Options (required for useIOTilesToLoad() and poptorch.set_overlap_for_input().)
Add method, cycleCount(), to determine the cycle count of the last model run.
Improve PopTorch’s parity with PyTorch’s Softplus
Improve implementation of torch.SiLU by using Poplar’s Swish operator
Additional support for operation overloads
Add documentation on available memory proportion to incorporate embeddings and indexing operations
Add documentation on how users can generate debug information
Support replicated tensor sharding when running on multiple processes
Allow selection for a non-constant x input.
Support for enableConvDithering convolution option

15.5.2. Bug Fixes

Fix issue where PopTorch recalculated upsampling scales in fp16
Fix issue where the last use of poptorch.set_available_memory would be pruned

15.5.3. API changes

Default mean reduction strategies have changed from the deprecated PostAndLoss strategy to Post or Running based on optimiser accumulation type
Mean reduction strategy can now be set via poptorch.Options.Training.setMeanAccumulationAndReplicationReductionStrategy.
Add warning that IPU-specific optimiser states cannot be read from the host, when calling get_state() on poptorch.optim optimisers

15.6. v2.2 (Poplar SDK 2.2)

15.6.1. New features

Migrated to PyTorch version 1.9.0
Support for torch.roll
Support for torch.clone
Add modelName session option that can be passed to PopART
Support List inputs to a model
Tuples/Lists of constants can now be returned by a model
Add enableProfiling convenience method in poptorch.Options to enable profile report generation
Fix bug with torch.Tensor.repeat when applied to an input during training
Fix bug with aten::to when applied to a constant used as an input to another node
Improved error message when encountering untraceable types during compilation
Support for torch.gather. Please note: this operator is known to cause long compilation times. Consider using a onehot-based solution instead or torch.index_select if appropriate.
Using a convolution layer op with the value of padding greater than or equal to kernel_size` is now supported.
Support for Poplar recoverable and unrecoverable errors.
Support for torch.flip.
Support for torch.Tensor.new_ones and torch.Tensor.new_zeros

15.6.2. API changes

Removed accumulationReductionType which was deprecated in 2.1 in favour of accumulationAndReplicationReductionType in poptorch.Options.Training
Removed runningVarianceAlwaysFloat which was deprecated in 2.1 and replaced by runningStatisticsAlwaysFloat in poptorch.Options.Precision,

15.7. v2.1 (Poplar SDK 2.1)

15.7.1. New features

Support for torch.unbind
Add option to set poptorch.Options using options specified in a config file.
Add mode=poptorch.DataLoaderMode.AsyncRebatched
Support for PopART name scopes via poptorch.NameScope
Add mixed precision automatic casting
Support for torch.cross
Support for torch.functional.one_hot
Support for torch.int8 data types
Support for torch.median
Support for torch.index_select
Support for torch.scatter_add
Add poptorch.Options.Precision.enableFloatingPointExceptions to control floating point exception behavior
Support for inplace changes to inputs.
Add option to log the number of IPU cycles used in executing the main graph
Support for torch.nn.GRU
Add automatic loss scaling option which can be enabled via poptorch.Options.Training.setAutomaticLossScaling.
Add poptorch.BlockFunction decorating for assigning an existing function to a block.
Add mechanism for inspecting arbitrary tensors
Add custom operator for CTC beam search decoding: poptorch.ctc_beam_search_decoder
Add a separate tensor variant (now default) to the SGD optimiser.
Add a TensorFlow variant to the RMSProp optimiser.

15.7.2. API changes

Removed Options.Popart which was deprecated in v2.0 and replaced with Options._Popart
Removed MultiConvPartialsType which was deprecated in v2.0
Deprecated poptorch.Options.Training.accumulationReductionType in favour of poptorch.Options.Training.accumulationAndReplicationReductionType
Deprecated runningVarianceAlwaysFloat in favour of runningStatisticsAlwaysFloat in poptorch.Options.Precision, as this new option computes both the running mean and variance in FP32 when this option is set to True.
Use of SGD via PyTorch’s or PopTorch’s API now results in use of the new separate tensor variant by default. To revert to the previous default variant, use poptorch.optim.SGD with use_combined_accum=True.

15.7.3. Known issues

Using a convolution layer op with the value of padding greater than or equal to kernel_size` results in an error when training. Use a constant pad layer instead of the excess padding prior to the convolution.

15.8. v2.0 (Poplar SDK 2.0)

15.8.1. New features

Support for the following activation functions:
- torch.nn.acosh
- torch.nn.asinh
- torch.nn.atanh
- torch.nn.Hardshrink
- torch.nn.SiLU
- torch.nn.Softplus
- torch.nn.Softshrink
- torch.nn.Threshold
Support for the following random sampling operations:
- torch.bernoulli
- torch.distributions.Bernoulli
Experimental support for torch.nn.CTCLoss
Add Adam optimizer
Support for torch.nn.AdaptiveAvgPool1d, torch.nn.AdaptiveAvgPool3d
Migrated to PyTorch version 1.7.1
Support for aten::index, aten::index_put_
Support for torch.zeros_like, torch.ones_like
Allow the user to specify which Optimizer attributes are constant or not.
Allow the user to specify mode=poptorch.DataLoaderMode.Async in poptorch.DataLoader constructor instead of explicitly creating an AsynchronousDataAccessor
Support for torch.nn.EmbeddingBag
Support for torch.clamp_max and torch.clamp_min
Support for torch.min(tensor, dim=.*, keepdim=.*) and torch.max(tensor, dim=.*, keepdim=.*) overloads.
Support for poptorch.isRunningOnIpu. This function returns True when executing on IPU and False when executing the model outside IPU scope.
Support for torch.amax and torch.amin
Support for attributes in custom ops.
Support for precompilation and reloading exported executables (poptorch.PoplarExecutor.compileAndExport and poptorch.load)
Support for slices with variable start index (slice size must be constant).
Add ipuHardwareVersion function to read the version of the IPU hardware present on the system.
Changed default targetd Ipu version for the model and offline compilation to 2.
Changed accumulationReductionType(reduction) option to now apply to replication reduction as well
Add environment variable POPTORCH_CACHE_DIR
Support for torch.fmod, and torch.remainder
Support for torch.addcdiv
Support for torch.bitwise_not

15.8.2. API changes

Deprecated Options.Popart, Options._Popart may be used experimentally.

15.9. v1.0 (Poplar SDK 1.4)

15.9.1. New features

Support for torch.nn.InstanceNorm1d, torch.nn.InstanceNorm2d and torch.nn.InstanceNorm3d
Fixed issue with torch.nn.GroupNorm where only 4-dimensional inputs could be used
Replaced Adam with AdamW optimizer.
Support for the following loss functions:
- torch.nn.KLDivLoss
- torch.nn.PoissonNLLLoss
- torch.nn.HingeEmbeddingLoss
- torch.nn.BCEWithLogitsLoss
- torch.nn.SmoothL1Loss
- torch.nn.SoftMarginLoss
- torch.nn.CosineEmbeddingLoss
- torch.nn.MarginRankingLoss
- torch.nn.TripletMarginLoss
- torch.nn.NLLLoss for aten::nll_loss2d
Support for torch.optim.RMSprop optimizer
Support for bool inputs to models
Improved support for half type models and inputs.
- Using a mix of float16 and float32 inputs is now supported. Please see the documentation for cases in which a model might use different types compared to when run natively with PyTorch.
Support for serialized matrix multiplications (poptorch.serializedMatMul)
Support for POPTORCH_IPU_MODEL_VERSION environment variable.
Support for torch.cumsum
Support for pipelined / phased / sharded execution.
Add PoplarExecutor.compile() to compile the model without executing it.
Use sphinx-build to generate the documentation.
Use Miniconda as build environment.
Support for torch.meshgrid
Support for torch.cartesian_prod
Optimized torch.matmul implementation with limitations
- Fused its input 0’s batch dimensions with the row dimension to avoid ReduceSum in its backward pass, for performance purpose
Partial support for torch.einsum
- Diagonals and ellipsis notation is unsupported
Support for executable caching: poptorch.Options.enableExecutableCaching()
Add optional title argument to poptorch.ipu_print_tensor
Add len() method to poptorch.AsynchronousDataLoader
Support for LAMB optimizer
Support for recomputationCheckpoint()
Support for torch.tensordot
Support for rounding up the number of IPU used to allow models which specify of number of IPUs which is not a power of 2: poptorch.Options.autoRoundNumIPUs(True) NB, this will reserve but not use IPUs and so it is preferable to specify the model to use a number of IPUs which is a power of two
Optimized torch.matmul implementation with limitations
- Fused its input 0’s batch dimensions with the row dimension to avoid ReduceSum in its backward pass, for performance purpose
Support for multi-convolutions with poptorch.MultiConv
Support for PopART batch serialization settings
- These can be set via poptorch.Options().Popart.set()
Support for PopVision System Analyser added: tracing can be enabled by setting PVTI_OPTIONS='{"enable":"true"}'

15.9.2. Known issues

Race condition in poptorch.DataLoader when using several workers resulting in the iteration sometimes finishing one element early.
- Workaround: set num_workers to 0 or 1.
poptorch.custom_op() doesn’t allow the user to set attributes.
- Workaround: hardcode the attributes in the custom operation or pass them as regular inputs.
Graphs containing block annotations (poptorch.Block or poptorch.BeginBlock) cannot be exported using torch.save()
- Workaround: Make a soft copy of the model that doesn’t contain Blocks and use it to save /load the weights. (The weights should be shared between the two models).
Lists of tensors are not supported as inputs.
- Workaround: Use tuples instead.
```
# Use a tuple
assert inference_model((t1, t2)) # instead of [t1, t2]
```

15.10. v0.1 (Poplar SDK 1.3)

15.10.1. New features

PopTorch now exposes PopART anchor options to choose how much data to return from a model. These are passed into the model wrapper via anchor_mode. options are Sum, All, Final and EveryN.
Support for batched LSTM and batch first
An Options object can now be passed to poptorch.trainingModel / poptorch.inferenceModel to configure the session and select IPUs
The ‘profile’ option has been removed, instead profiling can be enabled by setting the environment variable POPLAR_ENGINE_OPTIONS='{autoReport.all:true, autoReport.directory:.}'
Support for POPTORCH_IPU_MODEL and POPTORCH_WAIT_FOR_IPU environment variables.
Support for the torch comparisons operations:
- torch.eq
- torch.ge
- torch.gt
- torch.le
- torch.lt
- torch.max
- torch.min
- torch.ne
- torch.isnan
- torch.topk
- torch.min and torch.max only support (tensor, tensor) and (tensor) overloads. They do not support the (tensor, dim=, keepdim=) overload.
- torch.topk only supports sorted=False and Largest=True
Automatically synchronise the weights back to the Host after using the IPU for training. (i.e no need to explicitly call copyWeightsToHost() anymore)
Support for non-linear activations torch.nn.PReLU and torch.nn.Hardtanh
Support for Adam optimizer.
Support for half type models and inputs.
- Models that require operations on input tensors of mixed precision are not currently supported. For example:
```
def forward(self, x, y):
  x # Half
  y # Float32
  return x + y # Not supported.
```
Support for tensor.fill_, torch.full, torch.full_like
Support for user provided custom operations. See PopART documentation for information on how to write them. They are exposed by poptorch.custom_op this takes in a list of input tensors, strings for the PopART op name and domain, the domain version, and a list of tensors the same shape and size as the expected output tensors. This is to ensure the pytorch trace remains valid as it traces on CPU so won’t actually execute the operation when building the graph.
Support for torch.nn.Conv1D / torch.nn.Conv2D / torch.nn.Conv3D
Support for torch.nn.Upsample (‘nearest’ mode only)
Support for tensor.size
Support for the following random sampling operations.
- torch.rand
- torch.uniform_
- torch.distributions.Uniform
- torch.randn
- torch.normal
- torch.normal_
For repeatable random number generation use the randomSeed method of poptorch.Options
Support for torch.clamp
Adds poptorch.DataLoader
Adds optimized poptorch.AsynchronousDataAccessor which allows for a dataloader to be offloaded to a background thread asynchronously.
Support for torch.norm
Upgraded from torch 1.5.0 to torch 1.6.0
Experimental support for single host distributed execution
Add torch.where and tensor.masked_fill