Closed RobertTLange closed 2 years ago
I would like to have a set of utilities that automatically generates jobs and schedules them for the different hardware platforms. E.g. CPU/GPU/TPU. This should probably all run on GCP and with docker containers for reproducibility. Have a look pyhpc for benchmarking standards and potentially integrate with mle-toolbox GCP experiment setup.
Generate benchmarks for runtimes of the different environments and compare with plain NumPy/torch versions.