19. TensorFlow API changes

19.1. Release 3.0

19.1.1. Non-breaking changes

Deprecated modules

  • The tensorflow.python.ipu.horovod module has been moved to tensorflow.python.ipu.distributed and will be removed in the next release. Using tensorflow.python.ipu.horovod will still work in release 3.0, but trigger a deprecation warning.

19.2. Release 2.6

The following changes have been made to the TensorFlow API in the Poplar SDK version 2.6. This may require you to change your code.

19.2.1. Breaking changes

Warning

These will require changes to any code that uses them.

Removal of deprecated APIs

  • The tensorflow.python.ipu.ops.rnn_ops module has been removed (deprecated in Poplar SDK version 2.5). As a result, the following ops are no longer available - PopnnLSTM, PopnnDynamicLSTM, PopnnGRU, PopnnDynamicGRU, and PopnnAUGRU.

  • The tensorflow.python.ipu.keras.layers.rnn module has been removed (deprecated in Poplar SDK version 2.5). As a result, the following Keras layers are no longer available - PopnnLSTM, PopnnGRU.

19.3. Release 2.5

The following changes have been made to the TensorFlow API in the Poplar SDK version 2.5. This may require you to change your code.

19.3.1. Breaking changes

Warning

These will require changes to any code that uses them.

Removal of deprecated APIs

  • IpuConfig.floating_point_behaviour.esr may no longer be assigned a boolean value (deprecated in Poplar SDK version 2.4). Doing so will now raise a ValueError. From Poplar SDK version 2.5 onwards, StochasticRoundingBehaviour must be used. Note that serialized IpuConfig instances from an earlier Poplar SDK version will not be deserializable from Poplar SDK version 2.5 onwards.

  • extract_all_events, extract_all_strings_from_event_trace and extract_all_types_from_event_trace have been removed from tensorflow.python.ipu.utils. These items were deprecated in Poplar SDK version 2.4.

  • IPUMultiReplicaStrategy (deprecated in Poplar SDK version 2.4) has been removed. From Poplar SDK version 2.5 onwards use PopDistStrategy.

  • The save_interval_report flag provided to TF_POPLAR_FLAGS has been removed following deprecation in Poplar SDK version 2.4. From Poplar SDK version 2.5, libpva should be used.

  • tensorflow.python.ipu.ops.summary_ops has been removed as both ipu_compile_summary and get_ipu_reports were deprecated in Poplar SDK version 2.4. From Poplar SDK 2.5 onwards, the PopVision suite of tools should be used instead. Additionally, the compile_summary parameter of IPURunConfig has been removed.

Other

  • The flag save_vertex_graph from TF_POPLAR_FLAGS is now non-functional. Please use the target.saveOutputVertexGraph Poplar engine option instead.

19.3.2. Non-breaking changes

Deprecated layers

TensorFlow layers from tensorflow.python.ipu.ops.rnn_ops have been moved to the ipu_tensorflow_addons.layers namespace in IPU TensorFlow Addons.

The layers have been deprecated in TensorFlow and will be removed in a future release.

The table below lists all of the deprecated layers, and their new locations:

TensorFlow

IPU TensorFlow Addons

tensorflow.python.ipu.ops.rnn_ops.PopnnAUGRU tensorflow.python.ipu.ops.rnn_ops.PopnnDynamicGRU tensorflow.python.ipu.ops.rnn_ops.PopnnDynamicLSTM tensorflow.python.ipu.ops.rnn_ops.PopnnGRU tensorflow.python.ipu.ops.rnn_ops.PopnnLSTM

ipu_tensorflow_addons.layers.PopnnAUGRU ipu_tensorflow_addons.layers.PopnnDynamicGRU ipu_tensorflow_addons.layers.PopnnDynamicLSTM ipu_tensorflow_addons.layers.PopnnGRU ipu_tensorflow_addons.layers.PopnnLSTM

RNN available_memory_proportion_fwd/available_memory_proportion_bwd deprecated

The available_memory_proportion_fwd and available_memory_proportion_bwd arguments have been deprecated and will be removed from the following layers in a future release:

  • tensorflow.python.ipu.ops.rnn_ops.PopnnLSTM

  • tensorflow.python.ipu.ops.rnn_ops.PopnnDynamicLSTM

  • tensorflow.python.ipu.ops.rnn_ops.PopnnGRU

  • tensorflow.python.ipu.ops.rnn_ops.PopnnDynamicGRU

  • tensorflow.python.ipu.ops.rnn_ops.PopnnAUGRU

These values are now set using the 'availableMemoryProportion' key of the options and options_bwd arguments correspondingly.

19.4. Release 2.4

The following changes have been made to the TensorFlow API in the Poplar SDK version 2.4. This may require you to change your code.

19.4.1. Breaking changes

Warning

These will require changes to any code that uses them.

Summary ops

The following items related to summary ops have been deprecated, are no longer functional and will be removed in a future release. To profile IPU programs, use the PopVision suite of analysis tools. Trying to use these items will raise a NotImplementedError:

  • tensorflow.python.ipu.ops.summary_ops:

    • ipu_compile_summary

  • IPUEstimator:

    • The compile_summary argument to IPURunConfig

    • Passing a IPURunConfig with compile_summary set to True to an IPUEstimator

Removal of deprecated members

The following have been removed, as they were deprecated in a previous release:

  • The following TF_POPLAR_FLAGS:

    • dump_text_reports_to_stdio

    • add_all_reduce_copies

    • force_replicated_mode

    • save_oom_profiler

  • The following constructor arguments for IPUInfeedQueue:

    • replication_factor

    • data_to_prefetch

    • feed_name

  • The following constructor arguments for IPUOutfeedQueue:

    • replication_factor

    • io_batch_size

    • feed_name

  • The following constructor arguments for IPULoggingTensorHook:

    • replication_factor

    • feed_name

  • The following functions from tensorflow.python.ipu.utils:

    • create_ipu_config

    • set_serialization_options

    • set_optimization_options

    • set_norm_options

    • set_compilation_options

    • set_convolution_options

    • set_matmul_options

    • set_pooling_options

    • set_report_options

    • set_ipu_model_options

    • set_recomputation_options

    • set_floating_point_behaviour_options

    • set_io_tile_options

    • set_gcl_options

    • auto_select_ipus

    • select_ipus

    • set_ipu_connection_type

    • set_experimental_multi_replica_distribution_options

    • extract_compile_reports

    • extract_poplar_serialized_graphs

    • extract_execute_reports

  • The following functions from tensorflow.python.ipu.ops.nn_ops:

    • ctc_loss

    • ctc_loss_with_logits

  • The following functions from tensorflow.python.ipu.ops.internal_ops:

    • recompute

    • block_recompute

Additionally, the documentation section on profiling through the deprecated TensorFlow profiling APIs has been removed and an IpuOptions configuration protobuf can no longer be passed to the IPURunConfig constructor.

19.4.2. Non-breaking changes

  • The following functions from tensorflow.python.ipu.utils are now considered internal-only tools and have correspondingly been moved to tensorflow.compiler.plugin.poplar.tests.test_utils. They can still be accessed from their previous location, but not in future releases:

    • extract_all_events

    • extract_all_strings_from_event_trace

    • extract_all_types_from_event_trace

  • ‘IPUConfig.floating_point_behaviour.esr’ - Assigning a bool value is deprecated and will not be supported in a future release. StochasticRoundingBehaviour should be used instead.

  • ipu_multi_replica_strategy.IPUMultiReplicaStrategy has been renamed to popdist_strategy.PopDistStrategy. Using ipu_multi_replica_strategy.IPUMultiReplicaStrategy will trigger a deprecation warning.

  • IPUMultiWorkerStrategy is deprecated. Using IPUMultiWorkerStrategy will trigger a deprecation warning.

  • The flag save_interval_report from TF_POPLAR_FLAGS is now deprecated. Please use libpva instead.

19.5. Release 2.3

The following changes have been made to the TensorFlow API in the Poplar SDK version 2.3. This may require you to change your code.

19.5.1. Breaking changes

Warning

These will require changes to any code that uses them.

Custom user op metadata interface updates

The metadata interface for custom user ops has been updated with an additional parameter.

Existing user ops must update their custom_op_api_level value to 5 and update their metadata function to match the following signature

1void Build_metadata(
2  std::vector<std::int64_t>& allocating_indices,
3  std::vector<std::int64_t>& replica_identical_output_indices,
4  std::map<std::int64_t, std::int64_t>& input_to_output_tensor_aliasing,
5  bool& is_elementwise, bool& is_stateless, bool& is_hashable,
6  std::uint32_t num_inputs);

The verified transfers feature has been removed

The following functions from tensorflow.python.ipu.utils have been removed:

  • set_transfer_options

  • set_verification_options

The following classes from tensorflow.python.ipu.config have been removed:

  • KeyId

  • VerificationOptions

19.5.2. Non-breaking changes

  • IPUConfig.optimizations.enable_fast_math has been moved to IPUConfig.optimizations.math.fast

19.6. Release 2.2

The following changes have been made to the TensorFlow API in the Poplar SDK version 2.2. This may require you to change your code.

19.6.1. Breaking changes

Warning

These will require changes to any code that uses them.

C++ Poplar TensorFlow libraries are private by default

Users interested in targeting the IPU from C++ are required to use the new ipu_config library. We’ve made most C++ libraries produced as part of the Poplar backend private, so dependencies on poplar:driver and other libraries will no longer be valid and should be replaced with a dependency to //tensorflow/compiler/plugin/poplar:ipu_config. This library provides a public interface for configuring IPUs in C++, all other operations should use the standard TensorFlow C++ API. No other Poplar TensorFlow libraries should be directly depended on.

Reports removed from ipu events

Following the exclusion of profiling options from the New configuration API, reports have been removed from IPU events. The following functions from tensorflow.python.ipu.utils have been deprecated and now return blank lists:

  • extract_compile_reports

  • extract_poplar_serialized_graphs

  • extract_execute_reports

See the New configuration API changes for information on profiling TensorFlow programs using the profiling tools available in the SDK.

19.6.2. Non-breaking changes

These changes are recommended.

IPULoggingTensorHook replication_factor deprecated

The replication_factor argument of IPULoggingTensorHook will be removed in release 2.3. The replication factor is now automatically set based on the model being executed.

IPUInfeedQueue/IPUOutfeedQueue/IPULoggingTensorHook feed_name deprecated

The feed_name argument of IPUInfeedQueue, IPUOutfeedQueue and IPULoggingTensorHook has been deprecated and will be removed in release 2.3. The feed_name is now generated automatically internally.

Change of output location for profiling information

By default the profile information (profile.pop & frameworks.json) will now be output to a subdirectory of the Poplar autoReport.directory. If autoReport.directory is not set, it will be output to a subdirectory of the current working directory. This change means that mutliple profiles can be captured for a single model, if it is separated into different Poplar graphs.

The subdirectories are created using the following format tf_report__<iso_date>__<pid> and the cluster name can be read from the frameworks.json file in each subdirectory.

IPU Keras Layers deprecation in TensorFlow 1.15

IPU Keras layers (AssumeEqualAcrossReplicas, Dropout, Embedding, GroupNormalization, InstanceNormalization, LayerNormalization, RecomputationCheckpoint, PopnnLSTM and PopnnGRU) are deprecated and will be removed in the next release. If you require Keras support please migrate your model to TensorFlow 2 which has full Keras support for IPU.

Warning when epsilon value is too low

When the epsilon value given to instance_norm, layer_norm or group_norm is less than 1.53e-5, a warning will show on the screen that explains the potential dangers and suggests to increase it.

19.7. Release 2.1

The following changes have been made to the TensorFlow API in the Poplar SDK version 2.1. This may require you to change your code.

19.7.1. Breaking changes

Warning

These will require changes to any code that uses them.

We have removed several items that have been deprecated for at least one release.

tensorflow.python.ipu.ops.all_to_all_op.all_gather

  • The output shape has changed to have the replication_factor as the outermost instead of innermost dimension, matching the documentation.

tensorflow.python.ipu.utils

  • Removed report_options parameter from set_report_options.

    Use graph_options and execution_options parameters instead.

    Only removed for TensorFlow 1.15. Already removed in TensorFlow 2.1.

  • Removed allow_stateful_recompute parameter from set_recomputation_options.

    Pipelining recomputation will recompute all the non-stateful operations when recomputation is enabled.

    Only removed for TensorFlow 1.15. Already removed in TensorFlow 2.1.

  • Removed num_io_tiles from set_gcl_options.

    Use the set_io_tile_options instead.

    Only removed for TensorFlow 1.15. Already removed in TensorFlow 2.1.

IPUPipelineEstimator change

The definition for iterations_per_loop has changed. Previously the number of iterations was defined as the number of weight updates performed. The new definition is the number of mini-batches consumed, which makes it consistent with the IPUEstimator when using gradient accumulation.

The argument count_gradient_accumulation_as_iterations=True was previously required to use this new definition. That parameter has now been removed and the new definition is always used.

Autosharding removed

Autosharding has been removed. You should now use alternative execution modes such as pipelining instead.

Old IPU option configuration API changes

Note

These are changes to the old option configuration API. A new option configuration API has been introduced in this release and the old API is being deprecated. For more information, please see New configuration API.

The disable_graph_convolution_caching parameter for create_ipu_config (from tensorflow.python.ipu.utils) has been removed.

The disable_graph_outlining parameter must be used instead.

IPU Keras changes [TensorFlow 2]

The SequentialPipelineModel alias for PipelineSequential has been removed.

In the constructors of ipu.keras.Model and ipu.keras.Sequential, the alias accumulation_count for the gradient_accumulation_count parameter has been removed.

Similarly, the alias accumulation_dtype for gradient_accumulation_dtype has been removed.

19.7.2. Non-breaking changes

These changes are recommended.

Recompute suggestions deprecated

The recompute and block_recompute utility ops have been deprecated and will be removed in release 2.2. Automatic recomputation of casts will remain.

IPUInfeedQueue/IPUOutfeedQueue replication_factor deprecated

The replication_factor argument of IPUInfeedQueue and IPUOutfeedQueue has been deprecated and will be removed in release 2.2. The replication factor is now automatically set based on the model being executed.

IPUInfeedQueue data_to_prefetch deprecated

The data_to_prefetch argument of IPUInfeedQueue has been deprecated and will be removed in release 2.2. It is recommended to use the prefetch_depth argument instead.

IPUOutfeedQueue data_to_prefetch deprecated

The io_batch_size argument of IPUOutfeedQueue has been deprecated and will be removed in release 2.2. It is recommended to either manually accumulate results or use accumulate_outfeed when using pipelining.

CTC loss ops deprecated

The ctc_loss and ctc_loss_with_logits ops from ipu.ops.nn_ops have been deprecated and will be removed in release 2.2. They have been superseeded by ctc_loss_v2 and ctc_loss_with_log_probs.

New configuration API

A new API for configuring the IPU system has been added which is replacing the current API. The new API consists of a single class called IPUConfig with a hierarchical organisation of options as attributes. You can set options by assigning values to the attributes of an instance of this class. The class includes some usability features which should make the process of configuring the IPU system easier and with no hidden pitfalls. For more information about the new API, see Configuring system options.

Warning

The new IPUConfig API does not include the profiling options in the former configuration API, such as profiling, profile_execution, report_every_nth_execution etc. To profile a TensorFlow program, you should instead use the suite of profiling tools that have been added to the SDK. For general advice on how to enable profiling, refer to the Capturing IPU Reports chapter in the PopVision User Guide. To parse profiles, use the PopVision Analysis Python API chapter or PopVision Analysis C++ API chapter in the Poplar and PopLibs API Reference. To enable time-based profiling of events, see the Capturing Execution Information chapter of the PopVision User Guide.

Note that any Poplar engine options mentioned in the above guides can be passed to the compilation_poplar_options IPUConfig option, so it is not impossible to enable profiling using the new configuration API directly, but it is not advised, as environment variables will overwrite any values set this way.

Warning

The new IPUConfig API does not support verified transfers. This means the verified transfers feature will be removed when the old API is removed.

The new IPUConfig class is in a new namespace tensorflow.python.ipu.config. Multiple functions and classes have moved from tensorflow.python.ipu.utils to the config namespace:

  • configure_ipu_system()

  • get_ipu_config()

  • SelectionOrder

  • ExecutionProfileType

  • DeviceConnectionType

They can still be accessed from tensorflow.python.ipu.utils - along with IPUConfig - and there are currently no plans to remove this additional access route.

To help in converting from the old configuration API to the new API, the following table shows which attribute of IPUConfig each function argument in the old API corresponds to and how:

Table 19.1 Configuration API conversion

Old API function

Function argument

IPUConfig attribute equivalent

create_ipu_config()

profiling

Not supported in IPUConfig. Use the autoReport.outputGraphProfile or autoReport.all Poplar engine options.

enable_ipu_events

Not supported in IPUConfig. Use the PopVision System Analyser to inspect compilation, transfer and execution events.

use_poplar_text_report

Not supported in IPUConfig. Use the PopVision Graph Analyser for manual inspection of reports.

use_poplar_cbor_report

Not supported in IPUConfig. You can set the profiler.format Poplar engine option to the deprecated “v1” value for CBOR reports.

profile_execution

Not supported in IPUConfig. Use the autoReport.all and debug.computeInstrumentationLevel Poplar engine options.

enable_poplar_serialized_graph

Not supported in IPUConfig. Use the autoReport.outputSerializedGraph or autoReport.all Poplar engine options instead.

report_every_nth_execution

Not supported in IPUConfig. This feature will be removed when the former configuration API is removed.

max_report_size

Not supported in IPUConfig. The Poplar profiling format’s storage size has been significantly improved.

report_directory

Not supported in IPUConfig. To make module profiling files go into their own sub-directories, do not set autoReport.directory.

scheduler_selection

scheduling.algorithm 1

always_rearrange_copies_on_the_host

experimental.always_rearrange_copies_on_the_host

merge_infeed_io_copies

optimizations.merge_infeed_io_copies

disable_graph_outlining

optimizations.enable_graph_outlining 2

max_scheduler_lookahead_depth

scheduling.maximum_scheduler_lookahead_depth

max_scheduler_search_space_size

scheduling.maximum_scheduler_search_space_size

prefetch_data_streams

optimizations.prefetch_data_streams

selection_order

selection_order

enable_experimental_remote_buffer_embedding

experimental.enable_remote_buffer_embedding

set_serialization_options()

output_folder

serialization_output_folder

set_optimization_options()

combine_embedding_lookups

optimizations.combine_embedding_lookups

combine_matmuls

optimizations.combine_matmuls

max_cross_replica_sum_buffer_size

optimizations.maximum_cross_replica_sum_buffer_size

max_reduce_scatter_buffer_size

optimizations.maximum_reduce_scatter_buffer_size

max_inter_ipu_copies_buffer_size

optimizations.maximum_inter_ipu_copies_buffer_size

max_send_recv_cluster_size

optimizations.maximum_send_recv_cluster_size

minimum_remote_tensor_size

optimizations.minimum_remote_tensor_size

merge_remote_buffers

optimizations.merge_remote_buffers 3

gather_simplifier

optimizations.enable_gather_simplifier

triangular_solve_expander_block_size

optimizations.triangular_solve_expander_block_size

cholesky_block_size

optimizations.cholesky_block_size

enable_fast_math

optimizations.enable_fast_math

set_norm_options()

use_stable_statistics

norms.use_stable_statistics

experimental_distributed_batch_norm_replica_group_size

norms.experimental.distributed_batch_norm_replica_group_size

set_transfer_options()

use_verified_transfers

Not supported with IPUConfig. Verified transfers will be removed when the former configuration API is removed.

set_verification_options()

verification_options

set_compilation_options()

compilation_options 7

compilation_poplar_options

set_convolution_options()

convolution_options 7

convolutions.poplar_options

set_matmul_options()

matmul_options 7

matmuls.poplar_options

clear_pass_type

matmuls.clear_pass_type

set_pooling_options()

pooling_options 7

pooling.poplar_options

set_report_options()

graph_options

Not supported in IPUConfig. All graph report options have equivalents in the PopVision Graph Analyser or PopVision Analysis APIs

execution_options

Not supported in IPUConfig. All execution report options have equivalents in the PopVision Graph Analyser

set_ipu_model_options()

compile_ipu_code

ipu_model.compile_ipu_code

tiles_per_ipu

ipu_model.tiles_per_ipu

ipu_model_version

ipu_model.version

set_recomputation_options() 4

allow_recompute

allow_recompute

set_floating_point_behaviour_options() 5

inv

floating_point_behaviour.inv

div0

floating_point_behaviour.div0

oflo

floating_point_behaviour.oflo

esr

floating_point_behaviour.esr

nanoo

floating_point_behaviour.nanoo

set_io_tile_options()

num_io_tiles

io_tiles.num_io_tiles

place_ops_on_io_tiles

io_tiles.place_ops_on_io_tiles

io_tile_available_memory_proportion

io_tiles.available_memory_proportion

set_gcl_options()

gcl_options 7

gcl_poplar_options

auto_select_ipus()

num_ipus

auto_select_ipus

select_ipus()

indices

select_ipus

set_ipu_connection_type()

connection_type

device_connection.type 6

ipu_version

device_connection.version

enable_remote_buffers

device_connection.enable_remote_buffers

set_experimental_multi_replica_distribution_options()

process_count

experimental.multi_replica_distribution.process_count

process_index

experimental.multi_replica_distribution.process_index
1

IPUConfig.scheduling.algorithm takes a value from the new SchedulingAlgorithm enumeration, whereas the former configuration API took a string. The old string values map to the enumeration as follows:

  • “”: SchedulingAlgorithm.CHOOSE_BEST

  • “Clustering”: SchedulingAlgorithm.CLUSTERING

  • “PostOrder”: SchedulingAlgorithm.POST_ORDER

  • “LookAhead”: SchedulingAlgorithm.LOOK_AHEAD

  • “ShortestPath”: SchedulingAlgorithm.SHORTEST_PATH

2

IPUConfig.optimizations.enable_graph_outlining takes a boolean value that specifies whether or not graph outlining should be enabled. A value of True means that graph outlining is enabled. This is different to the old configuration API, which took a boolean value that specifies whether or not graph outlining should be disabled. Therefore, you should invert the boolean you gave to the old configuration API when passing it to an IPUConfig.

3

IPUConfig.optimizations.merge_remote_buffers takes a value from the new MergeRemoteBuffersBehaviour enumeration, whereas the former configuration API took a boolean or None value. The old values map to the enumeration as follows:

  • True: MergeRemoteBuffersBehaviour.MERGE

  • False: MergeRemoteBuffersBehaviour.NO_MERGING

  • None: MergeRemoteBuffersBehaviour.IF_BENEFICIAL

The IPUConfig also sets the default value to IF_BENEFICIAL, whereas the old configuration API sets the default value to NO_MERGING.

4

In the old configuration API, a call to set_recomputation_options would make the allow_recompute argument True by default, therefore merely calling set_recomputation_options(opts) would turn recomputation on. Please bear this in mind when moving to IPUConfig.

5

In the old configuration API, a call to set_floating_point_behaviour_options would make all of the arguments True by default, therefore merely calling set_floating_point_behaviour_options(opts) would turn all of inv, oflo, nanoo, div0 and esr on. Please bear this in mind when moving to IPUConfig. Note that there is the floating_point_behaviour.set_all option to unconditionally set all of these options on provided for convenience.

6

IPUConfig.device_connection.version takes a string, whereas the former configuration API took an integer. The old values map to the string values as follows:

  • 1: “ipu1”

  • 2: “ipu2”

7(1,2,3,4,5)

In the old configuration API, all options dictionaries are accumulative each time their function is called. For example, doing:

opts = set_compilation_options(opts, {"option1": "true"})
...
opts = set_compilation_options(opts, {"option2": "5"})

would mean that Poplar compilation is given both options {"option1": "true", "option2": "5"}.

In the IPUConfig API, this is not the case, as these options dictionaries are like any other Python dictionary: assigning to them again will overwrite them:

opts.compilation_poplar_options = {"option1": "true"}
...
opts.compilation_poplar_options = {"option2": "5"}

would mean that Poplar compilation is given only {"option2": "5"}. To achieve behaviour like the old configuration API, use the following:

opts.compilation_poplar_options = {"option1": "true"}
...
opts.compilation_poplar_options = {**{"option2", "5"},
                                   **opts.compilation_poplar_options}

Support for grouped collectives

tensorflow.python.ipu.ops.all_to_all_op.all_gather tensorflow.python.ipu.ops.reduce_scatter_op.reduce_scatter

  • The replication_factor can now be set to a value smaller than the total number of replicas in the model, in which case the collective operation will be performed within groups of the given size.

tensorflow.python.ipu.ops.cross_replica_ops.cross_replica_sum

  • A new optional argument replica_group_size is added for specifying the number of replicas in each collective group. If not specified, there is a single group containing all the replicas.

Environment variable changes

The dump_text_reports_to_stdio flag passed to TF_POPLAR_OPTIONS has been deprecated and has no effect. Use the PopVision Graph Analyser to manually inspect profiles.

19.8. Release 2.0

The following changes have been made to the TensorFlow API in the Poplar SDK version 2.0. This may require you to change your code.

19.8.1. Breaking changes

Warning

These will require changes to any code that uses them.

We have removed several items that have been deprecated for at least one release.

tensorflow.python.ipu.ipu_outfeed_queue

  • Removed outfeed_all parameter from IPUOutfeedQueue.

    Use outfeed_mode parameter instead.

tensorflow.python.ipu.ipu_pipeline_estimator

  • Removed pipeline_depth parameter from IPUPipelineEstimatorSpec.

    Use gradient_accumulation_count parameter instead.

tensorflow.python.ipu.utils

  • Removed retain_control_dependencies parameter from create_ipu_config.

    Only removed in TensorFlow 2.1.

  • Removed max_cross_replica_sum_buffer_size, and max_inter_ipu_copies_buffer_size parameters from create_ipu_config.

    Use set_optimization_options instead.

  • Removed report_options parameter from set_report_options.

    Use graph_options and execution_options parameters instead.

  • Removed allow_stateful_recompute parameter from set_recomputation_options.

    Pipelining recomputation will recompute all the non-stateful operations when recomputation is enabled.

  • Removed num_io_tiles from set_gcl_options.

    Use the set_io_tile_options instead.

tensorflow.python.ipu.ops.embedding_ops.embedding_lookup

  • Removed one_hot_threshold and min_encoding_size parameters from embedding_lookup.

  • Removed count parameter from HostEmbeddingScope.lookup.

tensorflow.python.ipu.ops.functional_ops

  • Removed function.

    Use outlined_function instead.

tensorflow.python.ipu.ops.normalization_ops

  • Removed reduction_axes parameter from group_norm, layer_norm, and instance_norm.

tensorflow.python.ipu.ops.pipelining_ops

  • Removed pipeline_depth parameter from pipeline.

    Use gradient_accumulation_count instead.

tensorflow.python.ipu.ops.rnn_ops

  • Removed support for passing a tuple as the initial_state argument for PopnnLSTM.call.

    This must be an LSTMStateTuple now.

The following deprecated namespace has been removed:

  • tensorflow.python.ipu.ipu_optimizer

Use the tensorflow.python.ipu.optimizers namespace instead.

19.8.2. Non-breaking changes

These changes are recommended.

IPUPipelineEstimator change

The definition for iterations_per_loop has changed. Previously the number of iterations was defined as the number of weight updates performed. The new definition is the number of mini-batches consumed, which makes it consistent with the IPUEstimator when using gradient accumulation. The old definition is still used by default, but it will be removed in a future release.

Use the argument count_gradient_accumulation_as_iterations=True to use the new definition.

Autosharding deprecated

Autosharding has been deprecated, and will be removed in a future release. You should now use alternative execution modes such as pipelining instead.

IPU config change

The disable_graph_convolution_caching parameter for create_ipu_config (from tensorflow.python.ipu.utils) has been deprecated as it has no effect. It will be removed in a future release.

The disable_graph_outlining parameter should be used instead.

IPU Keras changes [TensorFlow 2]

SequentialPipelineModel has been renamed to PipelineSequential for consistency with its Model counterpart. The old name can still be used, but is deprecated and will be removed in a future release.

The accumulation_count argument in the constructors of the ipu.keras.Model and ipu.keras.Sequential has been renamed to gradient_accumulation_count to be consistent with the rest of the code base. The old name can still be used, but is deprecated and will be removed in a future release.

Similarly, accumulation_dtype has been renamed to gradient_accumulation_dtype.