You can now leverage Apple’s tensorflow-metal PluggableDevice in TensorFlow v2.5 for accelerated training on Mac GPUs directly with Metal. Learn more here.
Mac-optimized TensorFlow and TensorFlow Addons
INTRODUCTION
This pre-release delivers hardware-accelerated TensorFlow and TensorFlow Addons for macOS 11.0+. Native hardware acceleration is supported on M1 Macs and Intel-based Macs through Apple’s ML Compute framework.
CURRENT RELEASE
- 0.1-alpha3
SUPPORTED VERSIONS
- TensorFlow r2.4rc0
- TensorFlow Addons 0.11.2
REQUIREMENTS
- macOS 11.0+
- Python 3.8 (required to be downloaded from Xcode Command Line Tools for M1 Macs).
INSTALLATION
An archive containing Python packages and an installation script can be downloaded from the releases.
-
To quickly try this out, copy and paste the following into Terminal:
% /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/apple/tensorflow_macos/master/scripts/download_and_install.sh)"This will verify your system, ask you for confirmation, then create a virtual environment with TensorFlow for macOS installed.
-
Alternatively, download the archive file from the releases. The archive contains an installation script, accelerated versions of TensorFlow, TensorFlow Addons, and needed dependencies.
% curl -fLO https://github.com/apple/tensorflow_macos/releases/download/v0.1alpha2/tensorflow_macos-${VERSION}.tar.gz % tar xvzf tensorflow_macos-${VERSION}.tar % cd tensorflow_macos % ./install_venv.sh --prompt
Installation on Conda
This pre-release version supports installation and testing using the Python from Xcode Command Line Tools. See #153 for more information on installation in a Conda environment.
Notes
For M1 Macs, the following packages are currently unavailable:
- SciPy and dependent packages
- Server/Client TensorBoard packages
When installing pip packages in a virtual environment, you may need to specify --target as follows:
% pip install --upgrade -t "${VIRTUAL_ENV}/lib/python3.8/site-packages/" PACKAGE_NAME
ISSUES AND FEEDBACK
Please submit feature requests or report issues via GitHub Issues.
ADDITIONAL INFORMATION
Device Selection (Optional)
It is not necessary to make any changes to your existing TensorFlow scripts to use ML Compute as a backend for TensorFlow and TensorFlow Addons.
There is an optional mlcompute.set_mlc_device(device_name='any') API for ML Compute device selection. The default value for device_name is 'any', which means ML Compute will select the best available device on your system, including multiple GPUs on multi-GPU configurations. Other available options are 'cpu' and 'gpu'. Please note that in eager mode, ML Compute will use the CPU. For example, to choose the CPU device, you may do the following:
# Import mlcompute module to use the optional set_mlc_device API for device selection with ML Compute.
from tensorflow.python.compiler.mlcompute import mlcompute
# Select CPU device.
mlcompute.set_mlc_device(device_name='cpu') # Available options are 'cpu', 'gpu', and 'any'.
Unsupported TensorFlow Features
The following TensorFlow features are currently not supported in this fork:
- tf.vectorized_map
- Higher-order gradients
- Jacobian-vector products (aka. forwardprop)
Logs and Debugging
Graph mode
Logging provides more information about what happens when a TensorFlow model is optimized by ML Compute. Turn logging on by setting the environment variable TF_MLC_LOGGING=1 when executing the model script. The following is the list of information that is logged in graph mode:
- Device used by ML Compute.
- Original TensorFlow graph without ML Compute.
- TensorFlow graph after TensorFlow operations have been replaced with ML Compute.
- Look for MLCSubgraphOp nodes in this graph. Each of these nodes replaces a TensorFlow subgraph from the original graph, encapsulating all the operations in the subgraph. This, for example, can be used to determine which operations are being optimized by ML Compute.
- Number of subgraphs using ML Compute and how many operations are included in each of these subgraphs.
- Having larger subgraphs that encapsulate big portions of the original graph usually results in better performance from ML Compute. Note that for training, there will usually be at least two MLCSubgraphOp nodes (representing forward and backward/gradient subgraphs).
- TensorFlow subgraphs that correspond to each of the ML Compute graphs.
Eager mode
Unlike graph mode, logging in eager mode is controlled by TF_CPP_MIN_VLOG_LEVEL. The following is the list of information that is logged in eager mode:
- The buffer pointer and shape of input/output tensor.
- The key for associating the tensor’s buffer to built the
MLCTrainingorMLCInferencegraph. This key is used to retrieve the graph and run a backward pass or an optimizer update. - The weight tensor format.
- Caching statistics, such as insertions and deletions.
Tips for debugging
- Larger models being trained on the GPU may use more memory than is available, resulting in paging. If this happens, try decreasing the batch size or the number of layers.
- TensorFlow is multi-threaded, which means that different TensorFlow operations, such as
MLCSubgraphOp, can execute concurrently. As a result, there may be overlapping logging information. To avoid this during the debugging process, set TensorFlow to execute operators sequentially by setting the number of threads to 1 (seetf.config.threading.set_inter_op_parallelism_threads). - In eager mode, you may disable the conversion of any operation to ML Compute by using
TF_DISABLE_MLC_EAGER=“;Op1;Op2;...”. The gradient op may also need to be disabled by modifying the file$PYTHONHOME/site-packages/tensorflow/python/ops/_grad.py(this avoids TensorFlow recompilation). - To initialize allocated memory with a specific value, use
TF_MLC_ALLOCATOR_INIT_VALUE=. - To disable ML Compute acceleration (e.g. for debugging or results verification), set the environment variable
TF_DISABLE_MLC=1.