NVTX Plugins for Deep Learning

THIS PROJECT IS NOT MAINTAINED ANYMORE

We recommend using Tensorflow-Profiler: https://www.tensorflow.org/guide/profiler

---

NVTX Plugins allows users to add their own NVIDIA Tools Extension (NVTX) events and time ranges to a TensorFlow graph. Applications which integrate NVTX can use NVIDIA Nsight Systems and Nsight Compute to capture and visualize these events and time ranges.

The NVTX ranges are added by wrapping regions of the computation graph with nvtx start and end operations.

NVTX Plugins also provides Keras callbacks and session hooks.

More about:

• NVTX: https://docs.nvidia.com/cuda/profiler-users-guide/index.html#nvtx
• NSight systems: https://developer.nvidia.com/nsight-systems

---

Table of Contents

• Installing or building NVTX Plugins
• Quick start guide
• Documentation
• Acknowledgments
• Disclaimer
• Contributions
• Copyright and License

---

Installing or building NVTX Plugins

Prerequisites

• Linux
• Python 3.4+
• NVIDIA GPU + CUDA toolkit 10.0 or newer
• TensorFlow 1.13 or newer

Installing NVTX-Plugins

The package can be installed from PyPI:

# Stable release
pip install nvtx-plugins

# Pre-release (may present bugs)
pip install nvtx-plugins --pre

The package is also available for download on github: https://github.com/NVIDIA/nvtx-plugins/releases

pip install nvtx-plugins*.tar.gz

Installing from source

You can build and install the package from source:

python setup.py sdist
pip install dist/nvtx-plugins*.tar.gz

For development objectives, you can install the package directly from source with:

python setup.py install

We recommend building the package inside NVIDIA's NGC TensorFlow container: https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow

For more information about how to get started with NGC containers, see the following sections from the NVIDIA GPU Cloud Documentation and the Deep Learning DGX Documentation: Getting Started Using NVIDIA GPU Cloud, Accessing And Pulling From The NGC container registry and Running TensorFlow.

Building the documentation

The documentation is built by running:

cd docs
pip install -r requirements.txt
make html

The documentation files will be generated in docs/build/html

Building the documentation does not require NVTX Plugins to be installed. Nonetheless, due to an issue in Sphinx only Python 3.7 is supported to build the documentation.

---

Quick start guide

Adding markers to the graph

Markers are added by wrapping parts of the computation graph with start and end operations. The operations are identity ops (passing the input to the output without modification) but they have a side effect of generating nvtx markers.

import nvtx.plugins.tf as nvtx_tf

x, nvtx_context = nvtx_tf.ops.start(x, message='Dense 1-3',
    domain_name='Forward', grad_domain_name='Gradient')
x = tf.layers.dense(x, 1000, activation=tf.nn.relu, name='dense_1')
x = tf.layers.dense(x, 1000, activation=tf.nn.relu, name='dense_2')
x = tf.layers.dense(x, 1000, activation=tf.nn.relu, name='dense_3')
x = nvtx_tf.ops.end(x, nvtx_context)
x = tf.layers.dense(x, 1000, activation=tf.nn.relu, name='dense_4')

For convenience, the package also provides a function dectorator:

@nvtx_tf.ops.trace(message='Dense Block', domain_name='Forward',
                   grad_domain_name='Gradient')
def dense_block(x):
    x = tf.layers.dense(x, 1000, activation=tf.nn.relu, name='dense_1')
    x = tf.layers.dense(x, 1000, activation=tf.nn.relu, name='dense_2')
    x = tf.layers.dense(x, 1000, activation=tf.nn.relu, name='dense_3')
    return x

More detailed examples can be found in examples/, also, check the Documentation for more information about other workflows including session hooks, Keras layers and callbacks.

Visualizing the ranges

NVTX requires a logger to register the generated events and ranges, we will use NVIDIA Nsight Systems to capture these event.

Run your code with nsys (pre-installed in NVIDIA's NGC TensorFlow container) to generate a qdrep file:

nsys profile -d 60 \
    -w true \
    --sample=cpu \
    -t 'nvtx,cuda' \
    -o ./generated_timeline \
    python ./network.py

The generated qdrep can be viewed using Nsight Systems.

Nsight Systems and nsys can also be downloaded and from the NVIDIA''s developer website.

More details about nsys and Nsight Systems can be found here.

---

Documentation

More details about NVTX Plugins can be found on here: https://nvtx-plugins.readthedocs.io/en/latest/

---

Acknowledgments

The project structure is heavily influenced by the TensorFlow custom-op example: https://github.com/tensorflow/custom-op

---

Disclaimer

The project is in beta stage, breaking changes are to be expected in the future.

---

Contributions

Contributions to NVTX Plugins are more than welcome. To contribute code, please submit a pull request against the master branch from a local fork.

We appreciate feedback, questions or bug reports. If you need help with the code, create a GitHub issue. Please follow the process outlined in the Stack Overflow https://stackoverflow.com/help/mcve document. Ensure that the posted examples are:

• minimal: Use as little code as possible that still produces the same problem.
• complete: Provide all parts needed to reproduce the problem. Check if you can strip external dependency and still show the problem. The less time we spend on reproducing the problems, the more time we can dedicate to the fixes.
• verifiable: Test the code you are about to provide, to make sure that it reproduces the problem. Remove all other problems that are not related to your request.

---

Copyright and License

This project is released under the Apache License, Version 2.0