2. Release overview
2.1. Operating system support
Ubuntu 18.04 is no longer supported.
PyTorch and TensorFlow 2 are no longer supported on CentOS 7.6.
The Poplar SDK for RHEL 8 requires the Python version to be upgraded to 3.9.
These changes enable the Poplar SDK to support PyTorch version 1.13.
2.2. PopTorch
Compatible with PyTorch 1.13 (upgraded from 1.10)
Added support for dict inputs.
Removed support for the legacy tracing frontend
torch.jit.trace()
.
2.3. TensorFlow
TensorFlow 2.6.3
Simplified public API for collectives.
Improved documentation on recomputation.
TensorFlow 1.15.5 (CentOS 7.6 only)
No significant changes.
2.4. Keras
Keras 2.6.0
Improved support for pipelining of nested Keras models.
Added automatic loss scaling support for training using Keras optimizers (experimental).
2.5. PopART
Added experimental support for custom transforms.
Improved handling of cached executables.
2.6. Poplar Libraries
Support sparse-dense and dense-sparse matrix multiplication where the sparsity structure of the sparse operand does not change.
Improved addConstant performance by 10x (and 20x when broadcasting) when adding large constants.
2.7. PopRun and PopDist
On IPU-Pods, automatically detect the RNIC interfaces to use for host-to-host communication.
2.8. PopEF
Added PopEF serialisation support for float8.
Added option to print tensor and feed data values and tensor shapes with
popef_dump
.
2.9. Model Runtime
Added user help and input verification to each of the tools.
Added functionality that allows NumPy arrays to be used with Model Runtime.
2.10. Poplar Triton Backend
Introduce new configuration option
check_package_hash
to prevent running an incompatible model.
2.11. Driver and utilities
Driver level support for C600 cards.
Command line tools support for C600 cards.