Targeting the IPU from TensorFlow 2

• 1. Introduction
  • 1.1. Document overview
• 2. Targeting the Poplar XLA device
  • 2.1. Supported types
  • 2.2. Device selection
  • 2.3. Configuring system options
    • 2.3.1. TF_POPLAR_FLAGS environment variable
  • 2.4. Supported operations
  • 2.5. Unsupported operations
  • 2.6. Error Handling
    • 2.6.1. Construction and compilation errors
    • 2.6.2. Runtime errors
• 3. Support for TensorFlow 2
  • 3.1. IPUStrategy
  • 3.2. Execution modes
    • 3.2.1. Graph mode with @tf.function
    • 3.2.2. Eager mode
  • 3.3. On-device loops
• 4. Keras with IPUs
  • 4.1. Single IPU models
  • 4.2. Using steps_per_execution
  • 4.3. Gradient accumulation
  • 4.4. Model parallelism
    • 4.4.1. Sequential model
    • 4.4.2. Functional model
      • Pipelining a model you are writing yourself
      • Pipelining an existing functional model
    • 4.4.3. Model subclass
      • Pipelining a model you are writing yourself
      • Pipelining an existing model
    • 4.4.4. Pipelining options
  • 4.5. Automatic data parallelism
  • 4.6. Asynchronous callbacks
  • 4.7. Configuring Infeeds and Outfeed
  • 4.8. Saving and loading Keras models
  • 4.9. Implementation details
• 5. Compiling and pre-compiling executables
  • 5.1. Caching of compiled executables
  • 5.2. Pre-compiling executables
    • 5.2.1. Unsupported Operations
• 6. Training a model
  • 6.1. Training loops, data sets and feed queues
  • 6.2. Optional simplification of infeeds and outfeeds
  • 6.3. Accessing outfeed queue results during execution
  • 6.4. Replicated graphs
    • 6.4.1. Selecting the number of replicas
    • 6.4.2. Performing parameter updates
  • 6.5. Pipelined training
    • 6.5.1. Grouped scheduling
    • 6.5.2. Interleaved scheduling
    • 6.5.3. Sequential scheduling
    • 6.5.4. Pipeline stage inputs and outputs
    • 6.5.5. Applying an optimiser to the graph
    • 6.5.6. Device mapping
    • 6.5.7. Concurrent pipeline stages
  • 6.6. Gradient accumulation
    • 6.6.1. Optimizers
    • 6.6.2. Pipelining
    • 6.6.3. Accumulation data type
  • 6.7. Optimizer state offloading
  • 6.8. Dataset benchmarking
    • 6.8.1. Accessing the JSON data
• 7. Efficient IPU I/O
  • 7.1. Prefetch elements
  • 7.2. I/O Tiles
• 8. Example using IPUEstimator
• 9. Example using IPUPipelineEstimator
• 10. Distributed training
  • 10.1. Example using IPUMultiWorkerStrategy
    • 10.1.1. The input function
    • 10.1.2. The model function
    • 10.1.3. Cluster definition
    • 10.1.4. Complete example
  • 10.2. Distributed training with Horovod
  • 10.3. Launching Horovod training
  • 10.4. Complete Horovod example
• 11. Half-precision floating point and stochastic rounding
  • 11.1. Controlling the half-precision floating-point unit
  • 11.2. Resetting the global random number seed
  • 11.3. Debugging numerical issues
• 12. IPU-optimised operations
  • 12.1. Image operations
  • 12.2. Matmul serialisation
  • 12.3. LSTM and GRU
  • 12.4. Dropout
  • 12.5. Embedding lookup
  • 12.6. Group normalisation
  • 12.7. Instance normalisation
  • 12.8. Layer normalisation
  • 12.9. GeLU activation
  • 12.10. Sequence slice
  • 12.11. Histogram
• 13. IPU Outlined Functions
  • 13.1. Usage
  • 13.2. Examples
    • 13.2.1. Models with common structures
    • 13.2.2. Serializing large operations
• 14. Writing custom operations
  • 14.1. Custom operation on the IPU
    • 14.1.1. Building the Poplar graph
    • 14.1.2. Gradient builders
    • 14.1.3. Metadata
    • 14.1.4. Compiling the IPU code
      • API level
      • PopLibs library code
      • Compiling the library file
    • 14.1.5. Using the custom op in TensorFlow
    • 14.1.6. Tensor allocation
    • 14.1.7. Examples
      • In-place operations
      • Operation attributes
      • Custom codelet
  • 14.2. Custom host CPU operations
    • 14.2.1. Gradient callback
• 15. IPU host embeddings
  • 15.1. Usage
  • 15.2. Example
  • 15.3. Experimental functionality: IPU embeddings in remote buffers
    • 15.3.1. Partitioning strategies
      • Token strategy
      • Encoding strategy
      • Choosing a strategy for your application
• 16. IPU embedded application runtime
  • 16.1. Usage
  • 16.2. Pipelining and I/O tiles
    • 16.2.1. Parallel requests
    • 16.2.2. Timeout
    • 16.2.3. Engine restarts
  • 16.3. Example
  • 16.4. Error Handling
    • 16.4.1. Runtime errors
• 17. Retrieving information about compilation and execution
  • 17.1. TensorFlow options for reporting
  • 17.2. XLA graph file naming
• 18. IPU TensorFlow Addons
  • 18.1. Introduction
  • 18.2. Keras layers
    • 18.2.1. IPU implementations of standard Keras layers
    • 18.2.2. Layers without upstream equivalents
    • 18.2.3. Code example
  • 18.3. Optimizers
• 19. TensorFlow API changes
  • 19.1. Release 2.5
    • 19.1.1. Breaking changes
      • IPU Keras changes
      • Removal of deprecated APIs
      • Other
    • 19.1.2. Non-breaking changes
      • Deprecated layers
      • Deprecated pipeline and gradient_accumulation options
      • RNN available_memory_proportion_fwd/available_memory_proportion_bwd deprecated
  • 19.2. Release 2.4
    • 19.2.1. Breaking changes
      • Summary ops
      • Removal of deprecated members
    • 19.2.2. Non-breaking changes
  • 19.3. Release 2.3
    • 19.3.1. Breaking changes
      • Custom user op metadata interface updates
      • The verified transfers feature has been removed
    • 19.3.2. Non-breaking changes
  • 19.4. Release 2.2
    • 19.4.1. Breaking changes
      • C++ Poplar TensorFlow libraries are private by default
      • Reports removed from ipu events
      • TensorFlow 2.1 to TensorFlow 2.4 Migration
    • 19.4.2. Non-breaking changes
      • IPULoggingTensorHook replication_factor deprecated
      • IPUInfeedQueue/IPUOutfeedQueue/IPULoggingTensorHook feed_name deprecated
      • Change of output location for profiling information
      • Warning when epsilon value is too low
  • 19.5. Release 2.1
    • 19.5.1. Breaking changes
      • IPUPipelineEstimator change
      • Autosharding removed
      • Old IPU option configuration API changes
      • IPU Keras changes [TensorFlow 2]
    • 19.5.2. Non-breaking changes
      • Recompute suggestions deprecated
      • IPUInfeedQueue/IPUOutfeedQueue replication_factor deprecated
      • IPUInfeedQueue data_to_prefetch deprecated
      • IPUOutfeedQueue data_to_prefetch deprecated
      • CTC loss ops deprecated
      • New configuration API
      • Support for grouped collectives
      • Environment variable changes
  • 19.6. Release 2.0
    • 19.6.1. Breaking changes
    • 19.6.2. Non-breaking changes
      • IPUPipelineEstimator change
      • Autosharding deprecated
      • IPU config change
      • IPU Keras changes [TensorFlow 2]
• 20. TensorFlow Python API
  • 20.1. Operations and utilities related to the Graphcore IPU
  • 20.2. Distribution strategy for a single system
  • 20.3. Compiler interface
  • 20.4. Scoping contexts
  • 20.5. Infeed queue
  • 20.6. Outfeed queue
  • 20.7. General utilities
  • 20.8. Configuration utilities
  • 20.9. Looping utilities
  • 20.10. Distributed training
  • 20.11. Horovod
  • 20.12. Serving utilities
  • 20.13. Datasets
    • 20.13.1. Dataset benchmarking
    • 20.13.2. Dataset wrappers
  • 20.14. Estimators
    • 20.14.1. IPUEstimator
    • 20.14.2. IPUPipelineEstimator
    • 20.14.3. Run configs
    • 20.14.4. Session run hooks
  • 20.15. Keras
    • 20.15.1. IPU specific Keras extensions
  • 20.16. Keras layers
    • 20.16.1. Keras layer specializations for the Graphcore IPU
  • 20.17. Keras losses
    • 20.17.1. Keras loss functions for the Graphcore IPU
  • 20.18. Keras optimizers
    • 20.18.1. Keras Optimizer wrappers for the Graphcore IPU
  • 20.19. Operators
    • 20.19.1. Control flow operations.
    • 20.19.2. Custom operations
    • 20.19.3. Functional operators
    • 20.19.4. Image operations
    • 20.19.5. Graphcore utility operations
    • 20.19.6. IPU specific maths operations
    • 20.19.7. Pipelining operators
    • 20.19.8. Popnn primitive neural network operators
    • 20.19.9. Popnn normalization operators
    • 20.19.10. Popnn recurrent neural network operators
    • 20.19.11. Popops all to all and all gather operators
    • 20.19.12. Popops cross replica operators
    • 20.19.13. Popops embedding operators
    • 20.19.14. Popops reduce scatter operator
    • 20.19.15. Popops within replica operators
    • 20.19.16. Poprand operators
    • 20.19.17. Utility operations to be used in replicated mode
    • 20.19.18. Slicing operators
    • 20.19.19. Statistics operators
    • 20.19.20. Embedded application runtime
  • 20.20. Optimisers
    • 20.20.1. Optimizer classes for the Graphcore IPU
  • 20.21. Sharding
    • 20.21.1. Utility functions for sharding graphs
• 21. TensorFlow operators supported by the IPU
• 22. IPU TensorFlow Addons API changes
  • 22.1. Release 2.5
    • 22.1.1. Non-breaking changes
      • RNN available_memory_proportion_fwd/available_memory_proportion_bwd deprecated
  • 22.2. Release 2.4
• 23. IPU TensorFlow Addons Python API
  • 23.1. Keras Layers
    • 23.1.1. Keras layers made for IPU TensorFlow
  • 23.2. Keras Optimizers
    • 23.2.1. Keras optimizers made for IPU TensorFlow
  • 23.3. Legacy TensorFlow Layers
    • 23.3.1. TensorFlow layers made for IPU TensorFlow
  • 23.4. Legacy TensorFlow Optimizers
    • 23.4.1. Optimizers made for IPU TensorFlow
• 24. Resources
  • 24.1. Graphcore
  • 24.2. TensorFlow
  • 24.3. Other
• 25. Trademarks & copyright