move horovod installation to multi-gpu

tox.ini

@@ -28,6 +28,7 @@ commands =
 ; Runs in: Github Actions
 ; Runs GPU-based tests.
 allowlist_externals =
+    bash
     horovodrun
 deps =
     -rrequirements/test.txt
@@ -41,12 +42,15 @@ commands =
     python -m pip install --upgrade git+https://github.com/NVIDIA-Merlin/core.git@{posargs:main}
     python -m pip install --upgrade git+https://github.com/NVIDIA-Merlin/dataloader.git@{posargs:main}
     python -m pip install --upgrade git+https://github.com/NVIDIA-Merlin/nvtabular.git@{posargs:main}
+    # Reinstall Horovod using `pip install --no-cache-dir` to build with the new version.
+    python -m pip install horovod --no-cache-dir
+    horovodrun --check-build
+    # Install distributed embeddings and check build
+    bash examples/usecases/multi-gpu/install_distributed_embeddings.sh {envtmpdir}
     # Run multi-gpu tests marked with `horovod` marker
     horovodrun -np 2 sh examples/usecases/multi-gpu/hvd_wrapper.sh pytest -m horovod -rxs tests/unit
 
 [testenv:py38-horovod-cpu]
-allowlist_externals =
-    bash
 setenv =
     HOROVOD_WITH_MPI=1
     HOROVOD_WITH_TENSORFLOW=1
@@ -55,11 +59,6 @@ setenv =
 commands =
     conda update --yes --name base --channel defaults conda
     conda env create --prefix {envdir}/env --file requirements/horovod-cpu-environment.yml --force
-    # Install horovod and check build
-    {envdir}/env/bin/python -m pip install horovod --no-cache-dir
-    {envdir}/env/bin/horovodrun --check-build
-    # Install distributed embeddings and check build
-    bash examples/usecases/multi-gpu/install_distributed_embeddings.sh {envtmpdir}
     # Install Merlin packages
     {envdir}/env/bin/python -m pip install --upgrade git+https://github.com/NVIDIA-Merlin/core.git
     {envdir}/env/bin/python -m pip install --upgrade git+https://github.com/NVIDIA-Merlin/dataloader.git