Skip to content

vAccel Tensorflow Bindings

To execute TF applications using vAccel, The first thing we need to do is install tensorflow on the machine we plan to run the plugin. This is a lengthy operation, as we need the C/C++ libraries which are not available from the official tensorflow repo as binaries. To build from source, follow the guide below.

Alternatively you can use a container image we have already built, so feel free to skip to Running in a container.

Get & Build the TF C/C++ API files

Building the TF libraries can be a lengthy process. For instance, on a 4-core machine with 16GB of memory the command below took ~2.5h.

git clone https://github.com/tensorflow/tensorflow.git && \
    cd tensorflow && \
    git checkout v2.11.0 && \
    ./configure && \
    bazel build --jobs=4 \
            --config=v2 \
            --copt=-O3 \
            --copt=-march=native \
            --config=opt \
            --verbose_failures \
            //tensorflow:tensorflow_cc \
            //tensorflow:install_headers \
            //tensorflow:tensorflow \
            //tensorflow:tensorflow_framework \
            //tensorflow/c:c_api \
            //tensorflow/tools/lib_package:libtensorflow

install TF runtime & libs to some dir in the host:

mkdir -p /opt/tensorflow/lib
cp -r bazel-bin/tensorflow/* /opt/tensorflow/lib/
cd /opt/tensorflow/lib && \
     ln -s libtensorflow_cc.so.2.6.0 libtensorflow_cc.so && \
     ln -s libtensorflow_cc.so.2.6.0 libtensorflow_cc.so.2 && \
     ln -s libtensorflow.so.2.6.0 libtensorflow.so && \
     ln -s libtensorflow.so.2.6.0 libtensorflow.so.2
LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH
LD_LIBRARY_PATH=/opt/tensorflow/lib:$LD_LIBRARY_PATH
ldconfig

Running in a container

If you want to avoid the burden (effort & time) to build the TF C/C++ libraries from source, you can use a ready-made container image. Just run:

docker run --rm -it -v$PWD:/data -w /data harbor.nbfc.io/nubificus/tensorflow-base-generic:latest /bin/bash

and perform the steps below inside the container.

Install vAccel on the host

wget https://s3.nbfc.io/nbfc-assets/github/vaccelrt/master/x86_64/Release-deb/vaccel-0.5.0-Linux.deb
sudo dpkg -i vaccel-0.5.0-Linux.deb

Get the Tensorflow plugin & bindings

wget https://s3.nbfc.io/nbfc-assets/github/vaccelrt/plugins/tensorflow/main/x86_64/Release-deb/vaccel-tensorflow-0.1.0-Linux.deb
dpkg -i vaccel-tensorflow-0.1.0-Linux.deb

There should be two .so files installed in /usr/local/lib:

$ find /usr/local/lib -name "libvaccel-tf*.so"
libvaccel-tf-bindings.so
libvaccel-tf.so

Alternatively, get the binary artifacts manually from the binaries table and install them to /usr/local/lib:

wget https://s3.nbfc.io/nbfc-assets/github/vaccelrt/plugins/tensorflow/main/x86_64/Release/libvaccel-tf.so
wget https://s3.nubificus.co.uk/nbfc-assets/github/vaccelrt/plugins/tensorflow/main/x86_64/Release/libvaccel-tf-bindings.so
cp libvaccel-tf-bindings.so /usr/local//lib
cp libvaccel-tf.so /usr/local/lib

Run a simple vAccel test (vAccel TF API)

To test the vAccel TF API we'll need a TF model. You can get a dir with the necessary files from our s3 repo:

wget https://s3.nbfc.io/tensorflow/lstm2.tar.gz
tar -zxvf lstm2.tar.gz --strip-components=1

Then, after setting the parameters below,

export VACCEL_DEBUG_LEVEL=4
export LD_LIBRARY_PATH=/usr/local/lib
export VACCEL_BACKENDS=/usr/local/lib/libvaccel-tf.so

we can run the three examples:

tf_model (just loads a pb model)

$ /usr/local/bin/tf_model lstm2/saved_model.pb 
2023.01.30-14:49:24.05 - <debug> Initializing vAccel
2023.01.30-14:49:24.05 - <debug> Created top-level rundir: /run/user/0/vaccel.XkutVJ
2023.01.30-14:49:24.10 - <debug> Registered plugin tf
2023.01.30-14:49:24.10 - <debug> Registered function TensorFlow session load from plugin tf
2023.01.30-14:49:24.10 - <debug> Registered function TensorFlow session run from plugin tf
2023.01.30-14:49:24.10 - <debug> Registered function TensorFlow session delete from plugin tf
2023.01.30-14:49:24.10 - <debug> Loaded plugin tf from /usr/local/lib/libvaccel-tf.so
2023.01.30-14:49:24.10 - <debug> session:1 New session
2023.01.30-14:49:24.10 - <debug> New rundir for resource /run/user/0/vaccel.XkutVJ/resource.2
2023.01.30-14:49:24.10 - <debug> Persisting file
2023.01.30-14:49:24.10 - <debug> Destroying resource 1
2023.01.30-14:49:24.10 - <debug> Destroying resource 2
2023.01.30-14:49:24.10 - <debug> Removing file /run/user/0/vaccel.XkutVJ/resource.2/model.pb
2023.01.30-14:49:24.10 - <debug> session:1 Free session
2023.01.30-14:49:24.10 - <debug> Shutting down vAccel
2023.01.30-14:49:24.10 - <debug> Cleaning up plugins
2023.01.30-14:49:24.10 - <debug> Unregistered plugin tf

tf_saved_model (loads a saved model)

$ /usr/local/bin/tf_saved_model lstm2 
2023.01.30-14:50:10.65 - <debug> Initializing vAccel
2023.01.30-14:50:10.65 - <debug> Created top-level rundir: /run/user/0/vaccel.eAMXnP
2023.01.30-14:50:10.71 - <debug> Registered plugin tf
2023.01.30-14:50:10.71 - <debug> Registered function TensorFlow session load from plugin tf
2023.01.30-14:50:10.71 - <debug> Registered function TensorFlow session run from plugin tf
2023.01.30-14:50:10.71 - <debug> Registered function TensorFlow session delete from plugin tf
2023.01.30-14:50:10.71 - <debug> Loaded plugin tf from /usr/local/lib/libvaccel-tf.so
2023.01.30-14:50:10.71 - <info> Testing SavedModel handling from path
2023.01.30-14:50:10.71 - <info> Creating new SavedModel handle
2023.01.30-14:50:10.71 - <info> Setting path of the model
2023.01.30-14:50:10.71 - <debug> Set TensorFlow model path to lstm2
2023.01.30-14:50:10.71 - <info> Registering model resource with vAccel
2023.01.30-14:50:10.71 - <debug> Registering new vAccel TensorFlow model
2023.01.30-14:50:10.71 - <debug> New resource 1
2023.01.30-14:50:10.71 - <info> Registered new resource: 1
2023.01.30-14:50:10.71 - <debug> session:1 New session
2023.01.30-14:50:10.71 - <info> New session: 1
2023.01.30-14:50:10.71 - <info> Registering model 1 with session 1
2023.01.30-14:50:10.71 - <info> Unregistering model 1 from session 1
2023.01.30-14:50:10.71 - <info> Destroying model 1
2023.01.30-14:50:10.71 - <debug> Destroying resource 1
2023.01.30-14:50:10.71 - <info> Destroying session 1
2023.01.30-14:50:10.71 - <debug> session:1 Free session
2023.01.30-14:50:10.71 - <info> Testing SavedModel handling from in memory data
2023.01.30-14:50:10.71 - <info> Creating new SavedModel handle
2023.01.30-14:50:10.71 - <debug> Setting protobuf file for model
2023.01.30-14:50:10.71 - <debug> Setting checkpoint file for model
2023.01.30-14:50:10.71 - <debug> Setting variables index file for model
2023.01.30-14:50:10.71 - <info> Registering model resource with vAccel
2023.01.30-14:50:10.71 - <debug> Registering new vAccel TensorFlow model
2023.01.30-14:50:10.71 - <debug> Persisting file
2023.01.30-14:50:10.71 - <debug> Persisting file
2023.01.30-14:50:10.71 - <debug> Persisting file
2023.01.30-14:50:10.71 - <debug> New resource 1
2023.01.30-14:50:10.71 - <info> Registered new resource: 1
2023.01.30-14:50:10.71 - <debug> session:1 New session
2023.01.30-14:50:10.71 - <info> New session: 1
2023.01.30-14:50:10.71 - <info> Registering model 1 with session 1
2023.01.30-14:50:10.71 - <info> Unregistering model 1 from session 1
2023.01.30-14:50:10.71 - <info> Destroying model 1
2023.01.30-14:50:10.71 - <debug> Destroying resource 1
2023.01.30-14:50:10.71 - <debug> Removing file /run/user/0/vaccel.eAMXnP/resource.1/saved_model.pb
2023.01.30-14:50:10.71 - <debug> Removing file /run/user/0/vaccel.eAMXnP/resource.1/variables/variables.data-00000-of-00001
2023.01.30-14:50:10.71 - <debug> Removing file /run/user/0/vaccel.eAMXnP/resource.1/variables/variables.index
2023.01.30-14:50:10.71 - <info> Destroying session 1
2023.01.30-14:50:10.71 - <debug> session:1 Free session
2023.01.30-14:50:10.71 - <debug> Shutting down vAccel
2023.01.30-14:50:10.71 - <debug> Cleaning up plugins
2023.01.30-14:50:10.71 - <debug> Unregistered plugin tf

tf_inference (performs inference on a saved model)

$ /usr/local/bin/tf_inference lstm2 
2023.01.30-14:50:44.04 - <debug> Initializing vAccel
2023.01.30-14:50:44.04 - <debug> Created top-level rundir: /run/user/0/vaccel.3qjD7J
2023.01.30-14:50:44.09 - <debug> Registered plugin tf
2023.01.30-14:50:44.09 - <debug> Registered function TensorFlow session load from plugin tf
2023.01.30-14:50:44.09 - <debug> Registered function TensorFlow session run from plugin tf
2023.01.30-14:50:44.09 - <debug> Registered function TensorFlow session delete from plugin tf
2023.01.30-14:50:44.09 - <debug> Loaded plugin tf from /usr/local/lib/libvaccel-tf.so
2023.01.30-14:50:44.09 - <debug> Set TensorFlow model path to lstm2
2023.01.30-14:50:44.09 - <debug> Registering new vAccel TensorFlow model
2023.01.30-14:50:44.09 - <debug> New resource 1
Created new model 1
2023.01.30-14:50:44.09 - <debug> session:1 New session
Initialized vAccel session 1
2023.01.30-14:50:44.09 - <debug> TensorFlow: load graph
2023.01.30-14:50:44.09 - <debug> Found implementation in tf plugin
2023.01.30-14:50:44.09 - <debug> [tf-plugin] Loading session from SavedModel
2023-01-30 14:50:44.108279: I tensorflow/cc/saved_model/reader.cc:45] Reading SavedModel from: lstm2
2023-01-30 14:50:44.159135: I tensorflow/cc/saved_model/reader.cc:89] Reading meta graph with tags { serve }
2023-01-30 14:50:44.159176: I tensorflow/cc/saved_model/reader.cc:130] Reading SavedModel debug info (if present) from: lstm2
2023-01-30 14:50:44.263678: I tensorflow/compiler/mlir/mlir_graph_optimization_pass.cc:357] MLIR V1 optimization pass is not enabled
2023-01-30 14:50:44.296034: I tensorflow/cc/saved_model/loader.cc:229] Restoring SavedModel bundle.
2023-01-30 14:50:44.459965: I tensorflow/cc/saved_model/loader.cc:213] Running initialization op on SavedModel bundle at path: lstm2
2023-01-30 14:50:44.605060: I tensorflow/cc/saved_model/loader.cc:305] SavedModel load for tags { serve }; Status: success: OK. Took 496802 microseconds.
2023.01.30-14:50:44.64 - <debug> Model from path loaded correctly
2023.01.30-14:50:44.72 - <debug> TensorFlow: run graph
2023.01.30-14:50:44.72 - <debug> Found implementation in tf plugin
2023.01.30-14:50:44.72 - <debug> [tf-plugin] Running session
2023.01.30-14:50:45.28 - <debug> [tf-plugin] Success
Success!
Output tensor => type:1 nr_dims:3
dim[0]: 1
dim[1]: 30
dim[2]: 61
Result Tensor :
0.016192
0.016538
0.016211
0.016086
0.016314
0.016367
0.016475
0.016484
0.016668
0.016283
2023.01.30-14:50:45.28 - <debug> TensorFlow: delete session
2023.01.30-14:50:45.28 - <debug> Found implementation in tf plugin
2023.01.30-14:50:45.37 - <debug> session:1 Free session
2023.01.30-14:50:45.37 - <debug> Destroying resource 1
2023.01.30-14:50:45.37 - <debug> Shutting down vAccel
2023.01.30-14:50:45.37 - <debug> Cleaning up plugins
2023.01.30-14:50:45.37 - <debug> Unregistered plugin tf

Build a tensorflow example application

git clone https://github.com/AmirulOm/tensorflow_capi_sample
cd tensorflow_capi_sample
gcc -I/opt/tensorflow/lib/include/ -L/opt/tensorflow/lib main.c -ltensorflow -o main.out

Make sure the saved model is present:

wget https://s3.nbfc.io/tensorflow/lstm2.tar.gz
tar -zxvf lstm2.tar.gz --strip-components=1

and after we set the LD_LIBRARY_PATH parameter:

export LD_LIBRARY_PATH=/opt/tensorflow/lib:$LD_LIBRARY_PATH

we can run:

$ ./main.out
2022-08-04 18:03:52.615232: I tensorflow/cc/saved_model/reader.cc:38] Reading SavedModel from: lstm2/
2022-08-04 18:03:52.668118: I tensorflow/cc/saved_model/reader.cc:90] Reading meta graph with tags { serve }
2022-08-04 18:03:52.668163: I tensorflow/cc/saved_model/reader.cc:132] Reading SavedModel debug info (if present) from: lstm2/
2022-08-04 18:03:52.884637: I tensorflow/cc/saved_model/loader.cc:211] Restoring SavedModel bundle.
2022-08-04 18:03:53.092219: I tensorflow/cc/saved_model/loader.cc:195] Running initialization op on SavedModel bundle at path: lstm2/
2022-08-04 18:03:53.240309: I tensorflow/cc/saved_model/loader.cc:283] SavedModel load for tags { serve }; Status: success: OK. Took 625082 microseconds.
TF_LoadSessionFromSavedModel OK
TF_GraphOperationByName serving_default_input_1 is OK
TF_GraphOperationByName StatefulPartitionedCall is OK
TF_NewTensor is OK
Session is OK
Result Tensor :
0.016192
0.016538
0.016211
0.016086
0.016314
0.016367
0.016475
0.016484
0.016668
0.016283

This is a generic TF execution example. To run with vAccel, apart from the above env parameters, we'll need to specify the bindings:

$ LD_PRELOAD=/usr/local/lib/libvaccel-tf-bindings.so ./main.out 
2023.01.30-14:54:38.23 - <debug> Initializing vAccel
2023.01.30-14:54:38.23 - <debug> Created top-level rundir: /run/user/0/vaccel.Ldk8YE
2023.01.30-14:54:38.25 - <debug> Registered plugin tf
2023.01.30-14:54:38.25 - <debug> Registered function TensorFlow session load from plugin tf
2023.01.30-14:54:38.25 - <debug> Registered function TensorFlow session run from plugin tf
2023.01.30-14:54:38.25 - <debug> Registered function TensorFlow session delete from plugin tf
2023.01.30-14:54:38.25 - <debug> Loaded plugin tf from /usr/local/lib/libvaccel-tf.so
2023.01.30-14:54:38.27 - <debug> session:1 New session
2023.01.30-14:54:38.27 - <debug> Set TensorFlow model path to lstm2/
2023.01.30-14:54:38.27 - <debug> Registering new vAccel TensorFlow model
2023.01.30-14:54:38.27 - <debug> New resource 1
2023.01.30-14:54:38.27 - <debug> TensorFlow: load graph
2023.01.30-14:54:38.27 - <debug> Found implementation in tf plugin
2023.01.30-14:54:38.27 - <debug> [tf-plugin] Loading session from SavedModel
2023-01-30 14:54:38.273960: I tensorflow/cc/saved_model/reader.cc:45] Reading SavedModel from: lstm2/
2023-01-30 14:54:38.313844: I tensorflow/cc/saved_model/reader.cc:89] Reading meta graph with tags { serve }
2023-01-30 14:54:38.313888: I tensorflow/cc/saved_model/reader.cc:130] Reading SavedModel debug info (if present) from: lstm2/
2023-01-30 14:54:38.409655: I tensorflow/compiler/mlir/mlir_graph_optimization_pass.cc:357] MLIR V1 optimization pass is not enabled
2023-01-30 14:54:38.441268: I tensorflow/cc/saved_model/loader.cc:229] Restoring SavedModel bundle.
2023-01-30 14:54:38.628391: I tensorflow/cc/saved_model/loader.cc:213] Running initialization op on SavedModel bundle at path: lstm2/
2023-01-30 14:54:38.763201: I tensorflow/cc/saved_model/loader.cc:305] SavedModel load for tags { serve }; Status: success: OK. Took 489261 microseconds.
2023.01.30-14:54:38.80 - <debug> Model from path loaded correctly
2023-01-30 14:54:38.870546: I tensorflow/cc/saved_model/reader.cc:45] Reading SavedModel from: lstm2/
2023-01-30 14:54:38.918771: I tensorflow/cc/saved_model/reader.cc:89] Reading meta graph with tags { serve }
2023-01-30 14:54:38.918819: I tensorflow/cc/saved_model/reader.cc:130] Reading SavedModel debug info (if present) from: lstm2/
2023-01-30 14:54:39.050492: I tensorflow/cc/saved_model/loader.cc:229] Restoring SavedModel bundle.
2023-01-30 14:54:39.231115: I tensorflow/cc/saved_model/loader.cc:213] Running initialization op on SavedModel bundle at path: lstm2/
2023-01-30 14:54:39.379673: I tensorflow/cc/saved_model/loader.cc:305] SavedModel load for tags { serve }; Status: success: OK. Took 509137 microseconds.
TF_LoadSessionFromSavedModel OK
TF_GraphOperationByName serving_default_input_1 is OK
TF_GraphOperationByName StatefulPartitionedCall is OK
TF_NewTensor is OK
2023.01.30-14:54:39.41 - <debug> TensorFlow: run graph
2023.01.30-14:54:39.41 - <debug> Found implementation in tf plugin
2023.01.30-14:54:39.41 - <debug> [tf-plugin] Running session
2023.01.30-14:54:40.02 - <debug> [tf-plugin] Success
Session is OK
2023.01.30-14:54:40.02 - <debug> TensorFlow: delete session
2023.01.30-14:54:40.02 - <debug> Found implementation in tf plugin
2023.01.30-14:54:40.12 - <debug> Destroying resource 1
2023.01.30-14:54:40.19 - <debug> session:1 Free session
Result Tensor :
0.016192
0.016538
0.016211
0.016086
0.016314
0.016367
0.016475
0.016484
0.016668
0.016283
2023.01.30-14:54:40.19 - <debug> Shutting down vAccel
2023.01.30-14:54:40.19 - <debug> Cleaning up plugins
2023.01.30-14:54:40.19 - <debug> Unregistered plugin tf