Release Notes

nGraph is provided as source code, APIs, build scripts, and some binary formats for various Compiler stack configurations and use cases.

For downloads formatted as .zip and tar.gz, see https://github.com/NervanaSystems/ngraph/releases.

This page includes additional documentation updates.

We are pleased to announce the release of version 0.25.

Core updates for 0.25

  • PlaidML support

  • Double-buffering support

  • Constant folding

  • Support for static linking

  • Additional ops

  • Preliminary static linking support

Known issues

  • No PlaidML training support

Latest doc updates 0.25-doc

  • Instructions how to build NGRAPH_PLAIDML backend added to Build and Test.

  • Publishing interim version of doc navigation for updates at ngraph.ai

  • GPU validations: added 5 functional TensorFlow workloads and 4 functional ONNX workloads to Validated workloads.

Important

Pre-releases (-rc-0.*) have newer features, and are less stable.

Changelog on Previous Releases

0.24

  • Fixes reshape sink/swim issue

  • More ONNX ops

  • Elementwise divide defaults to Python semantics

  • GenerateMask seed optional

  • Graph visualization improvements

  • Preserve control dependencies in more places

  • GetOutputElement has single input

0.23

  • More ONNX ops

  • Elementwise divide defaults to Python semantics

  • GenerateMask seed optional

  • Document new debug tool

  • Graph visualization improvements

  • Note deprecation of MXNet’s ngraph-mxnet PyPI

  • Note default change to svg files for graphs and visualization

  • Add more prominent tips for contributors who find the doc-contributor-README

  • Better GSG / Install Guide structure.

  • Added group edits and new illustrations from PR 2994 to introduction.rst.

  • Ensure ngraph-bridge link in README goes to right place.

  • Make project extras their own subdirectory with index to help organize them.

  • Known Issues

    • When using TensorFlow* v1.14.0 with `ngraph-bridge v0.16.0rc0 and CPU backend, we saw notable to severe decreases in throughput in many models.

0.22

  • More ONNX ops

  • Optimizations

  • Don’t reseed RNG on each use

  • Initial doc and API for IntelGPU backend

  • DynamicBackend API

0.21

  • The offset argument in tensor reads and writes has been removed

  • Save/load API

  • More ONNX ops

  • Better tensor creation

  • More shape support

  • Provenance improvements

  • offset arg for tensor creation is deprecated

  • static linking support

  • Initial test of 0.21-doc

  • Updated doc-contributor-README for new community-based contributions.

  • Added instructions on how to test or display the installed nGraph version.

  • Added instructions on building nGraph bridge (ngraph-bridge).

  • Updated Backend Developer Guides and ToC structure.

  • Tested documentation build on Clear Linux OS; it works.

  • Fixed a few links and redirs affected by filename changes.

  • Some coding adjustments for options to render math symbols, so they can be documented more clearly and without excessive JS (see replacements.txt).

  • Consistent filenaming on all BE indexes.

  • Removed deprecated TensorAPI.

0.20

  • Save/load API

  • More ONNX ops

  • Better tensor creation

  • More shape support

  • Provenance improvements

0.19

  • More dynamic shape preparation

  • Distributed interface factored out

  • fp16 and bfloat16 types

  • codegen execution parameterized by context

  • NodeMap, NodeVector, ParameterVector, ResultVector now vectors

    • node_vector.hpp replaced by node.hpp

    • op/parameter_vector.hpp replaced by op/parameter.hpp

    • op/result_vector.hpp replaced by op/result.hpp

  • Additional ONNX ops

  • Add graph visualization tools to doc

  • Update doxygen to be friendlier to frontends

0.18

  • Python formatting issue

  • mkl-dnn work-around

  • Event tracing improvements

  • Gaussian error function

  • Begin tracking framework node names

  • ONNX quantization

  • More fusions

0.17

  • Allow negative padding in more places

  • Add code generation for some quantized ops

  • Preliminary dynamic shape support

  • initial distributed ops

  • Pad op takes CoordinateDiff instead of Shape pad values to allow for negative padding.

0.16

  • NodeInput and NodeOutput classes prepare for simplifications of Node

  • Test improvements

  • Additional quantization ops

  • Performance improvements

  • Fix memory leak

  • Concat optimization

  • Doc updates