PyTorch offers a debugging toolbox that comprises a built-in stats and trace
function. This functionality facilitates the display of the time spent by each
compilation phase, output code, output graph visualization, and IR dump. TORCH_COMPILE_DEBUG invokes this debugging tool that allows
for better problem-solving while troubleshooting the internal issues of TorchDynamo and
TorchInductor. This functionality works for the models optimized using zentorch, and
hence it can be leveraged to debug these models too. To enable this functionality,
either set the environment variable TORCH_COMPILE_DEBUG=1 or specify the environment variable with the
runnable file (for example, test.py) as input.
For example, if the file
test.py contains a
model optimized by torch.compile with zentorch as backend, use:
TORCH_COMPILE_DEBUG=1 python test.py