For zentorch, enable CPP specific logging by setting the environment variable
TORCH_CPP_LOG_LEVEL. This has four levels: INFO,
WARNING, ERROR and FATAL in
decreasing order of verbosity.
Similarly, enable Python logging by setting the environment variable
ZENTORCH_PY_LOG_LEVEL. This has five levels:
DEBUG, INFO, WARNING,
ERROR, and CRITICAL, again in decreasing order of
verbosity.
Here is an example of how to enable INFO level logs for cpp and DEBUG level for Python
(most
verbose):
export TORCH_CPP_LOG_LEVEL=INFO
export ZENTORCH_PY_LOG_LEVEL=DEBUGWARNING is the default level of logs for both
cpp and Python sources, but it can be overridden.Note: The log levels are the same as those provided by the Python logging
module.
INFO: As all Operators implemented in zentorch are registered with torch using the
TORCH_LIBRARY() and TORCH_LIBRARY_IMPL() macros in
bindings, the PyTorch profiler can be used without any modification to measure the
operator level performance.