tensorflow-neuron 2.x Release Notes¶
This document lists the release notes for the tensorflow-neuron 2.x packages.
Support on serialized TensorFlow 2.x custom operators is currently limited. Serializing some operators registered from tensorflow-text through TensorFlow Hub is going to cause failure in tensorflow.neuron.trace.
Memory leak exists on latest releases of TensorFlow Neuron for versions 2.1, 2.2, 2.3, and 2.4.
Issue: When compiling large models, user might run out of memory and encounter this fatal error.
terminate called after throwing an instance of 'std::bad_alloc'
Solution: run compilation on a c5.4xlarge instance type or larger.
Issue: When upgrading
pip install tensorflow-neuron --upgrade, the following error message may appear, which is caused by
pipversion being too low.
Could not find a version that satisfies the requirement tensorflow<1.16.0,>=1.15.0 (from tensorflow-neuron)
Solution: run a
pip install pip --upgrade before upgrading
Issue: Some Keras routines throws the following error:
AttributeError: 'str' object has no attribute 'decode'.
Solution: Please downgrade h5py by pip install ‘h5py<3’. This is caused by https://github.com/TensorFlow/TensorFlow/issues/44467.
Added support for Tensorflow 2.8.0.
Added support for Slice operator
The graph partitioner now prefers to place less compute intensive operators on CPU if the model already contains a large amount of compute intensive operators.
Fixed Github issue #408, the fix solves data type handling bug in
tfn.tracewhen the model contains Conv2D operators.
Updated TensorFlow 2.5 to version 2.5.3.
Added support for TensorFlow 2.6 and 2.7.
Added a warning message when calling
tfn.saved_model.compileAPI. In tensorflow-neuron 2.x you should call tensorflow.neuron.trace.
tfn.saved_model.compileAPI supports only partial functionality of tensorflow.neuron.trace and will be deprecated in the future.
Fixed a bug in TensorFlow Neuron versions 2.1, 2.2. 2.3 and 2.4. The fixed bug was causing a memory leak of 128 bytes for each inference.
Improved warning message when calling deprecated compilation API under tensorflow-neuron 2.x.
Fixed a bug that caused a memory leak. The memory leak was approximately 128b for each inference and exists in all versions of Neuron TensorFlow versions part of Neuron 1.16.0 to Neuron 1.17.0 releases. see Previous Releases Content for exact versions included in each release. This release only addresses the leak in TensorFlow Neuron 2.5. Future release of TensorFlow Neuron will fix the leak in other versions as well (2.1, 2.2, 2.3, 2.4).
Updated TensorFlow 2.5 to version 2.5.2.
Enhanced auto data parallel (e.g. when using NEURONCORE_GROUP_SIZES=X,Y,Z,W) to support edge cases.
Fixed a bug that may cause tensorflow-neuron to generate in some cases scalar gather instruction with incorrect arguments.
Updated Neuron Runtime (which is integrated within this package) to
libnrt 22.214.171.124to fix a container issue that was preventing the use of containers when /dev/neuron0 was not present. See details here Neuron Runtime 2.x Release Notes.
New in this release¶
TensorFlow Neuron 2.x now support Neuron Runtime 2.x (
libnrt.soshared library) only.
You must update to the latest Neuron Driver (
aws-neuron-dkmsversion 2.1 or newer) for proper functionality of the new runtime library.
Read Migrate your application to Neuron Runtime 2.x (libnrt.so) for detailed information of how to migrate your application.
Updated TensorFlow 2.3.x from TensorFlow 2.3.3 to TensorFlow 2.3.4.
Updated TensorFlow 2.4.x from TensorFlow 2.4.2 to TensorFlow 2.4.3.
Updated TensorFlow 2.5.x from TensorFlow 2.5.0 to TensorFlow 2.5.1.
Fix bug that can cause illegal compiler optimizations
Fix bug that can cause dynamic-shape operators be placed on Neuron
New in this release¶
First release of TensorFlow 2.x integration, Neuron support now TensorFlow versions 2.1.4, 2.2.3, 2.3.3, 2.4.2, and 2.5.0.
New public API tensorflow.neuron.trace: trace a TensorFlow 2.x keras.Model or a Python callable that can be decorated by tf.function, and return an AWS-Neuron-optimized keras.Model that can execute on AWS Machine Learning Accelerators.
Please note that TensorFlow 1.x SavedModel compilation API tensorflow.neuron.saved_model.compile is not supported in tensorflow-neuron 2.x . It continues to function in tensorflow-neuron 1.15.x .