mirror of
https://github.com/ROCm/ROCm.git
synced 2026-02-19 02:34:19 -05:00
84 lines
3.1 KiB
Markdown
84 lines
3.1 KiB
Markdown
# Release Notes
|
|
<!-- Do not edit this file! This file is autogenerated with -->
|
|
<!-- tools/autotag/tag_script.py -->
|
|
|
|
<!-- Disable lints since this is an auto-generated file. -->
|
|
<!-- markdownlint-disable blanks-around-headers -->
|
|
<!-- markdownlint-disable no-duplicate-header -->
|
|
<!-- markdownlint-disable no-blanks-blockquote -->
|
|
<!-- markdownlint-disable ul-indent -->
|
|
<!-- markdownlint-disable no-trailing-spaces -->
|
|
|
|
<!-- spellcheck-disable -->
|
|
|
|
The release notes for the ROCm platform.
|
|
|
|
-------------------
|
|
|
|
## ROCm 5.2.3
|
|
<!-- markdownlint-disable first-line-h1 -->
|
|
<!-- markdownlint-disable no-duplicate-header -->
|
|
### Changes in This Release
|
|
|
|
#### Ubuntu 18.04 End of Life Announcement
|
|
|
|
Support for Ubuntu 18.04 ends in this release. Future releases of ROCm will not provide prebuilt packages for Ubuntu 18.04.
|
|
HIP and Other Runtimes
|
|
|
|
#### HIP Runtime
|
|
|
|
##### Fixes
|
|
|
|
- A bug was discovered in the HIP graph capture implementation in the ROCm v5.2.0 release. If the same kernel is called twice (with different argument values) in a graph capture, the implementation only kept the argument values for the second kernel call.
|
|
|
|
- A bug was introduced in the hiprtc implementation in the ROCm v5.2.0 release. This bug caused the `hiprtcGetLoweredName` call to fail for named expressions with whitespace in it.
|
|
|
|
Example:
|
|
|
|
The named expression `my_sqrt<complex<double>>` passed but `my_sqrt<complex<double >>` failed.
|
|
ROCm Libraries
|
|
|
|
#### RCCL
|
|
|
|
##### Added
|
|
|
|
Compatibility with NCCL 2.12.10
|
|
|
|
- Packages for test and benchmark executables on all supported OSes using CPack
|
|
|
|
- Added custom signal handler - opt-in with RCCL_ENABLE_SIGNALHANDLER=1
|
|
|
|
- Additional details provided if Binary File Descriptor library (BFD) is pre-installed.
|
|
|
|
- Added experimental support for using multiple ranks per device
|
|
|
|
- Requires using a new interface to create communicator (ncclCommInitRankMulti), refer to the interface documentation for details.
|
|
|
|
- To avoid potential deadlocks, user might have to set an environment variables increasing the number of hardware queues. For example,
|
|
|
|
```sh
|
|
export GPU_MAX_HW_QUEUES=16
|
|
```
|
|
|
|
- Added support for reusing ports in NET/IB channels
|
|
|
|
- Opt-in with NCCL_IB_SOCK_CLIENT_PORT_REUSE=1 and NCCL_IB_SOCK_SERVER_PORT_REUSE=1
|
|
|
|
- When "Call to bind failed: Address already in use" error happens in large-scale AlltoAll(for example, >=64 MI200 nodes), users are suggested to opt-in either one or both of the options to resolve the massive port usage issue
|
|
|
|
- Avoid using NCCL_IB_SOCK_SERVER_PORT_REUSE when NCCL_NCHANNELS_PER_NET_PEER is tuned >1
|
|
|
|
##### Removed
|
|
|
|
- Removed experimental clique-based kernels
|
|
|
|
#### Development Tools
|
|
|
|
No notable changes in this release for development tools, including the compiler, profiler, and debugger
|
|
Deployment and Management Tools
|
|
|
|
No notable changes in this release for deployment and management tools.
|
|
Older ROCm Releases
|
|
|
|
For release information for older ROCm releases, refer to <https://github.com/RadeonOpenCompute/ROCm/blob/master/CHANGELOG.md>
|