Added time log for each step and removed Bfloat16cast

[LinlinCui-NOAA]

This PR made two changes to run_graphcast.py:

• Added time log for each step
• Removed conversion to Bfloat16. Bfloat16 should not be used for prediction. The reduced precision greatly increases run-to-run variance.

[aerorahul]

looks good.
One could introduce timing stats in rollout.chunked_predictions as well as converter.save_grib2 to further get info on computation and IO respectively through the model integration.

[RussellManser-NCO]

The additional time logging looks good, but I have concerns about the change to float casting if this is intended for the current production code.

Requested changes were made. Thank you.

[RussellManser-NCO]

I will run a test for this on WCOSS this afternoon.

[RussellManser-NCO]

The print statements are not being written to output while run_graphcast.py is executing. Could you modify the shebang to the following please?

#!/usr/bin/env -S python3 -u

[aerorahul]

The print statements are not being written to output while run_graphcast.py is executing. Could you modify the shebang to the following please?

#!/usr/bin/env -S python3 -u

could also add flush=True to the print statement. However, seems like there is something else at play.

[aerorahul]

If you don't want to make code changes, add export PYTHONUNBUFFERED=1 to the runtime environment.

[LinlinCui-NOAA]

I can switch to use absl.logging, which can write to output while the script is executing, e.g.:

[2025-12-10 20:13:39,365] absl INFO: Elapsed time for loading input: 41.77620339393616 seconds
[2025-12-10 20:13:48,159] absl INFO: Elapsed time for extracting inputs, targets, and forcings: 8.793529987335205 seconds
[2025-12-10 20:13:49,123] absl INFO: Elapsed time for normalization: 0.9637486934661865 seconds

@RussellManser-NCO Please let me know which one you prefer.

[RussellManser-NCO]

I can switch to use absl.logging, which can write to output while the script is executing, e.g.:

[2025-12-10 20:13:39,365] absl INFO: Elapsed time for loading input: 41.77620339393616 seconds
[2025-12-10 20:13:48,159] absl INFO: Elapsed time for extracting inputs, targets, and forcings: 8.793529987335205 seconds
[2025-12-10 20:13:49,123] absl INFO: Elapsed time for normalization: 0.9637486934661865 seconds

@RussellManser-NCO Please let me know which one you prefer.

absl.logging works. It's nice to have the timestamps and logging info.

[RussellManser-NCO]

Logging was not unbuffered, unfortunately, even with a modified shebang. I tried export PYTHONUNBUFFERED=1 as well, which also did not work. The latest push does work.

[LinlinCui-NOAA]

OK. Thanks for testing.