13. Changelog

13.1. v2.1 (Poplar SDK 2.1)

13.1.1. 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.

13.1.2. 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.

13.1.3. 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.

13.2. v2.0 (Poplar SDK 2.0)

13.2.1. 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

13.2.2. API changes

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

13.3. v1.0 (Poplar SDK 1.4)

13.3.1. 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 float 16 and float 32 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"}'

13.3.2. 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]
```

13.4. v0.1 (Poplar SDK 1.3)

13.4.1. 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