Fix disabling progress bar on non-zero ranks using Horovod backend (#1709)

* Fix Horovod backend to disable progress bar on all ranks except 0

* Add join barriers

* Added changelog

* Make protected and add verbosity

* Refactor to disable progress bar callback in train

* Removed vebose setting

* Add cache check for Horovod

* Test run again

* Updated comment

* Always skip cache for Horovod

* Only reinstall when necessary

* Added separate step

* Fixed spacing

* Skip Python 3.8
This commit is contained in:
Travis Addair
2020-05-04 13:02:57 -04:00
committed by GitHub
parent 1a9f1c80a1
commit f90afa29b8
5 changed files with 39 additions and 2 deletions
+11
View File
@@ -86,6 +86,17 @@ jobs:
pip list
shell: bash
- name: Reinstall Horovod if necessary
if: runner.os != 'windows' && matrix.python-version != '3.8'
run: |
HOROVOD_BUILT=$(python -c "import horovod.torch; horovod.torch.nccl_built(); print('SUCCESS')")
if [[ $HOROVOD_BUILT != "SUCCESS" ]]; then
pip uninstall -y horovod
HOROVOD_BUILD_ARCH_FLAGS="-mfma" pip install --no-cache-dir $(grep "horovod" requirements-extra.txt)
fi
horovodrun --check-build
shell: bash
- name: Cache datasets
uses: actions/cache@v1
with: