Proposed Version: TBD
Proposed Schedule:
- Code Freeze: TBD
- Release published: TBD
Proposed content:
Project | Lead Contributor | Project Docs | Link to design discussion on dev@ | Notes |
---|---|---|---|---|
Horovod-MXNet Integration | Carl Yang, Lin Yuan, Darren Hu | Horovod-MXNet Integration | Discussion | Upgrade from experimental feature to GA |
Gluon Vision Model Zoo | zhasheng@apache.org | |||
Fix Flaky Tests | zhasheng@apache.org | Github Flaky Tests | Discussion | Resolve remaining flaky tests? |
MKL-DNN | Patric Zhao, Da Zheng | MKL-DNN integration design | Discussion | Already GA in 1.3? |
MKL-DNN based graph optimization and quantization by subgraph | Patric Zhao | MXNet Graph Optimization and Quantization based on subgraph and MKL-DNN | Discussion | experimental feature |
Proper support for Types | Sebastian (Wolfram) | Backend for general purpose numerical packages (support for multipe types, proper testing for all supported types). Derive test priorities from Mathematica internal suite or integrate Mathematica into CI? Feedback: Users are observing instability, incomplete behavior vs. quality delivered by numpy / Pytorch. Haibin: many tests are not well written, most test only cover fp32, don't have good guidelines for community, operator tutorial lacks guidance on test requirements, suggest to develop guidelines (results are deterministic, cover all supported types), need to set quality standards as community grows Sebastian: presentation how it is done in Mathematica based on experience and learnings | ||
Sparse Tensor support for Gluon | Haibin Lin | Gluon Sparse Support | Discussion | Upgrade from experimental feature to GA? |
Android support | Anton | experimental feature or can it be GA? | ||
Topology-aware AllReduce | Topology-aware AllReduce Proposal | Discussion | Upgrade from experimental feature to GA? | |
Control flow operators | Da Zheng, | Optimize dynamic neural networks | Discussion | This is the first step towards optimizing dynamic neural networks by adding symbolic and imperative control flow operators (foreach, while_loop, maybe ifelse). foreach and while_loop have been merged. ifelse has been submitted for review, but is unlikely merged in this release. |
TensorRT Runtime Integration | Discussion | Upgrade from experimental feature for GA? | ||
Java API | MXNet Java Inference API | |||
JVM Memory Management | JVM Memory Management | Discussion | ||
Subgraph API for integrating external accelerators with MXNet | Jun Wu, Da Zheng | Subgraph API | https://github.com/apache/incubator-mxnet/pull/12157 |
Critical PRs to track:
PR | Title | Contributor | Status |
---|---|---|---|
PR candidates suggested v1.3.1 patch release:
From dev@ discussion
- https://github.com/apache/incubator-mxnet/issues/11849 and resolution in https://github.com/apache/incubator-mxnet/pull/12412
- https://github.com/apache/incubator-mxnet/pull/12631
- https://github.com/apache/incubator-mxnet/pull/12603
- https://github.com/apache/incubator-mxnet/pull/12499
- https://github.com/apache/incubator-mxnet/pull/12530
- unit test PR: concat 11371, pool 11608, LRN 11831, Sum 11272, backward 11232, gluon 10921.
- Open issues:
From discussion about gluon toolkits:
- https://github.com/apache/incubator-mxnet/pull/12025
- https://github.com/apache/incubator-mxnet/pull/12554
- https://github.com/apache/incubator-mxnet/pull/12118
Other recommendations:
- https://github.com/apache/incubator-mxnet/pull/12692
- Upgrade dependency libraries - MKL-DNN v0.15
https://github.com/apache/incubator-mxnet/pull/12953 - https://github.com/apache/incubator-mxnet/pull/12804 - critical CUDNN fix