rocm sdk builder 6.1.1-2 released

[lamikr]

ROCM SDK Builder release 6.1.1-2 is now available and it is tagged with an release tag v6.1.1-2 both on the master and releases/rocm_sdk_builder_611 git branches. Master branch will transfer shortly for adding support to build the rocm 6.1.2 software stack.

A lot of people has contributed to this release by providing testing, feedpack, improvements and bug fixes both for the source code and documentation. It is released about 3 weeks after the inititial 6.6.1-1 release.

ROCM SDK Builder 6.1.1-2 is based on the source code for AMD's rocm 6.1.1 release but includes many patches in aim of allowing to build the AMD's machine learning stack for consumer level graphic cards on many different Linux distributions. In addition of AMD's ROCm components, this release builds also many other open source machine learning tools like pytorch, pytorch vision, pytorch audio, onnxruntime and DeepSpeed.

Various small example applications are provided under docs/examples folder. and more contributions are welcome.

Full ChangeLog

ROCM SDK Builder 6.1.1-2 (2024-06-14)

Thanks for the following people for contributions, feedbacks, bugreporting and testing since the previous release
- daniandtheweb
- mritunjaymusale
- crizle
- JassonCordones
- hsmalley
- flip111
- Stefan-Oit
- Meso-uca
- eLBart0-DTG
- eitch
- AngryLoki
- Tuleo1

• Support added for a many Linux distributions thanks by the testing by others
  • Fedora 40
  • Ubuntu 24.04
  • Ubuntu 22.04
  • Linux Mint 21
  • Arch Linux
  • Manjaro
• GPU's tested
  • AMD RX 5500
  • AMD RX 6600
  • AMD RX 6800
  • AMD RX 7800 XT
  • AMD RX 7900 XTX
  • AMD 680i Mobile / gfx1035
  • gfx 1036?
• Packages updated to never releases
  • python 3.9.19
  • openmpi 5.0.3
  • pytorch v2.3.1
  • pytorch vision v0.18.1
  • pytorch audio v2.3.1
  • onnxruntime 1.18.0
  • DeepSpeed v0.14.3
  • zstd v1.5.5
• Build System Improvements
  • more heuristic approach for defining the CPU count used for building different apps
    depending of the CPU count and ram available.
    (To avoid running out of memory on some extensive apps like the llvm linking phase)
  • Build now checks and warns more clearly if the git configuration
    is missing or if the read-write permission to amd gpu driver
    /dev/kfd is not configured properly.
  • Lot of improvements to error messages, code comments and bash code syntax used
  • Fixes for the onnxruntime and deepspeed build
  • Aotriton cpu usage fixes (allocated 2x cpus available and put build systems down)
  • Aotriton workaround for issue wheer triton compiler not always producing .hsaco files (issue 28)
  • Upstream bug repors and fixes (rocm_smi_lib, pytorch)
• test apps, documentation and benchmarking
  • Improved various example codes under docs/examples folder to test
    pytorch, opencl, onnxruntime, amdmigraphx, etc.
  • added missing test apps launch scripts
  • pytorch gpu benchmark tests execution improvements in
    https://github.com/lamikr/pytorch-gpu-benchmark

Note that THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.