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 compilation options
    • 2.3.1. TF_POPLAR_FLAGS environment variable
  • 2.4. Caching of compiled executables
  • 2.5. Supported operations
  • 2.6. Unsupported operations
• 3. Support for TensorFlow 2
  • 3.1. Function annotation with @tf.function
  • 3.2. IPUStrategy
  • 3.3. Keras
    • 3.3.1. Model class
    • 3.3.2. Sequential class
    • 3.3.3. PipelineModel class
    • 3.3.4. SequentialPipelineModel class
    • 3.3.5. Custom training loops
• 4. TensorFlow 2 examples
  • 4.1. Training on the IPU
  • 4.2. Custom training function
  • 4.3. Pipelined model
• 5. Training a model
  • 5.1. Training loops, data sets and feed queues
  • 5.2. Accessing outfeed queue results during execution
  • 5.3. Replicated graphs
    • 5.3.1. Selecting the number of replicas
    • 5.3.2. Data feeds
    • 5.3.3. Performing parameter updates
  • 5.4. Pipelined training
    • 5.4.1. Sequential scheduling
    • 5.4.2. Interleaved scheduling
    • 5.4.3. Grouped scheduling
    • 5.4.4. Pipeline stage inputs and outputs
    • 5.4.5. Applying an optimiser to the graph
    • 5.4.6. Device mapping
  • 5.5. Gradient accumulation
    • 5.5.1. Optimizers
    • 5.5.2. Pipelining
    • 5.5.3. Accumulation data type
  • 5.6. Optimizer state offloading
  • 5.7. Dataset benchmarking
    • 5.7.1. Accessing the JSON data
• 6. Example using IPUEstimator
• 7. Example using IPUPipelineEstimator
• 8. Distributed training
  • 8.1. The input function
  • 8.2. The model function
  • 8.3. Cluster definition
  • 8.4. Complete example
  • 8.5. Distributed training with Horovod
  • 8.6. Horovod Open MPI dependency
  • 8.7. Launching Horovod training
  • 8.8. Complete Horovod example
• 9. Half-precision floating point and stochastic rounding
  • 9.1. Controlling the half-precision floating-point unit
  • 9.2. Resetting the global random number seed
  • 9.3. Debugging numerical issues
• 10. IPU-optimised operations
  • 10.1. LSTM and GRU
  • 10.2. Dropout
  • 10.3. Embedding lookup
  • 10.4. Group normalisation
  • 10.5. Instance normalisation
  • 10.6. Layer normalisation
  • 10.7. GeLU activation
• 11. IPU Outlined Functions
  • 11.1. Usage
  • 11.2. Examples
    • 11.2.1. Models with common structures
    • 11.2.2. Serializing large operations
• 12. Custom IPU operations
  • 12.1. Fully customised IPU operations
    • 12.1.1. Metadata
    • 12.1.2. In-place operations
    • 12.1.3. Elementwise operations
    • 12.1.4. Tensor allocation
    • 12.1.5. Gradient operations
    • 12.1.6. Stateless operations
    • 12.1.7. Operation attributes
    • 12.1.8. Example
  • 12.2. Fully customised CPU operations
  • 12.3. Custom elementwise expressions
  • 12.4. API Level Versioning
    • 12.4.1. Changes in API Level
• 13. IPU host embeddings
  • 13.1. Usage
  • 13.2. Example
  • 13.3. Experimental functionality: IPU embeddings in remote buffers
    • 13.3.1. Partitioning strategies
      • Token strategy
      • Encoding strategy
      • Choosing a strategy for your application
• 14. Retrieving information about compilation and execution
  • 14.1. Adding an operation to get compilation and execution events
    • 14.1.1. ipu_event_trace()
    • 14.1.2. ipu_compile_summary(name, [op list])
  • 14.2. Enabling tracing in the hardware configuration options
  • 14.3. Extract the reports from the returned events
  • 14.4. Producing reports for use with the PopVision Graph Analyser
    • 14.4.1. COMPILE_BEGIN
    • 14.4.2. COMPILE_END
      • Tensor map
    • 14.4.3. EXECUTE
  • 14.5. Using the IPU Model device for debugging
  • 14.6. TensorFlow options for reporting
  • 14.7. Reading the Poplar textual summary report
    • 14.7.1. Target
    • 14.7.2. Graph
    • 14.7.3. Memory usage
  • 14.8. Producing an ELF image of the compilation
  • 14.9. Dumping auxiliary Poplar information
    • 14.9.1. Poplar vertex graph
    • 14.9.2. Poplar interval report
  • 14.10. XLA graph file naming
• 15. Python API
  • 15.1. Operations and utilities related to the Graphcore IPU
  • 15.2. Distribution strategy for a single system
  • 15.3. Compiler interface
  • 15.4. Scoping contexts
  • 15.5. Infeed queue
  • 15.6. Outfeed queue
  • 15.7. General utilities
  • 15.8. Looping utilities
  • 15.9. Distributed training
  • 15.10. Datasets
    • 15.10.1. Dataset benchmarking
    • 15.10.2. Dataset wrappers
  • 15.11. Estimators
    • 15.11.1. IPUEstimator
    • 15.11.2. IPUPipelineEstimator
    • 15.11.3. Run configs
    • 15.11.4. Session run hooks
  • 15.12. Keras
    • 15.12.1. Keras API
    • 15.12.2. Keras Model interfaces for IPU
  • 15.13. Keras layers
    • 15.13.1. Keras layer specializations for the Graphcore IPU
  • 15.14. Operators
    • 15.14.1. Custom operations
    • 15.14.2. Functional operators
    • 15.14.3. Graphcore utility operations
    • 15.14.4. IPU specific maths operations
    • 15.14.5. Pipelining operators
    • 15.14.6. Popnn primitive neural network operators
    • 15.14.7. Popnn normalization operators
    • 15.14.8. Popnn recurrent neural network operators
    • 15.14.9. Popops all to all and all gather operators
    • 15.14.10. Popops cross replica operators
    • 15.14.11. Popops embedding operators
    • 15.14.12. Popops reduce scatter operator
    • 15.14.13. Poprand operators
    • 15.14.14. Utility operations to be used in replicated mode
    • 15.14.15. Summary operations for IPUs
  • 15.15. Optimisers
    • 15.15.1. Optimizer wrapper for replicated graphs
    • 15.15.2. Optimizer wrappers which perform local gradient accumulation
    • 15.15.3. Optimizer wrapper that modifies gradients before application
    • 15.15.4. Optimizer wrapper for sharded graphs
  • 15.16. Sharding
    • 15.16.1. Automatic graph sharding
    • 15.16.2. Utility functions for sharding graphs
• 16. TensorFlow operators supported by the IPU
• 17. Resources
  • 17.1. Graphcore
  • 17.2. TensorFlow
  • 17.3. Other
• 18. Index
• 19. Trademarks & copyright