| nsys profile \ | |
| -o /pscratch/sd/j/joshuaho/full_stats_profile_1_gpu_batch_size_1028 \ | |
| --capture-range=cudaProfilerApi \ | |
| --duration=100 \ | |
| --force-overwrite true \ | |
| --trace=nvtx \ | |
| --cudabacktrace=all \ | |
| python scripts/training_script.py --config configs/stats_all/ttH_CP_even_vs_odd.yaml --preshuffle --nocompile --lazy --restart --profile | |
| nsys profile \ | |
| -o /pscratch/sd/j/joshuaho/full_stats_profile_1_gpu_batch_size_2048 \ | |
| --capture-range=cudaProfilerApi \ | |
| --duration=100 \ | |
| --force-overwrite true \ | |
| --trace=nvtx \ | |
| --cudabacktrace=all \ | |
| python scripts/training_script.py --config configs/stats_all/ttH_CP_even_vs_odd_batch_size_2048.yaml --preshuffle --nocompile --lazy --restart --profile | |
| nsys profile \ | |
| -o /pscratch/sd/j/joshuaho/full_stats_profile_1_gpu_batch_size_4096 \ | |
| --capture-range=cudaProfilerApi \ | |
| --duration=100 \ | |
| --force-overwrite=true \ | |
| --trace=nvtx \ | |
| --cudabacktrace=all \ | |
| python scripts/training_script.py --config configs/stats_all/ttH_CP_even_vs_odd_batch_size_4096.yaml --preshuffle --nocompile --lazy --restart --profile | |
| nsys profile \ | |
| -o /pscratch/sd/j/joshuaho/full_stats_profile_1_gpu_batch_size_8192 \ | |
| --capture-range=cudaProfilerApi \ | |
| --duration=100 \ | |
| --force-overwrite true \ | |
| --trace=nvtx \ | |
| --cudabacktrace=all \ | |
| python scripts/training_script.py --config configs/stats_all/ttH_CP_even_vs_odd_batch_size_8192.yaml --preshuffle --nocompile --lazy --restart --profile | |