After recently upgrading to rocm 5.7, we see that tensorflow-rocm uses "CPU" and never "GPU". We confirmed this with rocm-smi. Tests with py-torch and HIP-EXAMPLES show gpu usage. This happens with latest tensorflow-rocm and anaconda, and with the container ( rocm/tensorflow:latest )
( we run with apptainer/singularity ).
Our tensorflow in the system anaconda setup is : tensorflow-rocm 2.12.0.560 .
Anybody have a working rocm5.7 with tensorflow on the MI100 ? Is there a trick or environment setting needed to get GPU tensorflow working? The examples were working in previous rocm 5.2 .
