A GPU characterization tool that takes chrome traces and optionally hardware counters for multiple GPUs as input, and generates compute, memory, and architecture visualizations.
To install chopper run:
pip install .
The chopper profiler is invoked using its python module. In this example, GPU and CPU telemetry are enabled.
python -m chopper.profile.collect --gpu-telemetry --cpu-telemetry -- <workload args>
It is up to the workload to enable the PyTorch profiler.
To only execute the workload args and/or hardware profiler on a container while collecting telemetry outside of it, pass --container with the apptainer .sif file path.
Otherwise, chopper can be installed and ran inside a container.
To select and view plots run:
python -m chopper.window
To type check with mypy run:
python -m tests