Why are we using olympic scores stdev for RCP interpolation?

[ShriyaPalsamudram]

The general idea of stdev is to get the variance of a population, but Olympic scoring drops the extreme values. So we should consider using stdev of the entire population for interpolation.

Need to analyze and fix for v4.1

[pgmpablo157321]

[pgmpablo157321]

[hiwotadese]

Can we see how this affect the v4.0 scores?

[pgmpablo157321]

@ShriyaPalsamudram I am getting the same results when with and without the olympic stdev. It seems to be, because it the RCP Stdev is only being used to compute the min_epochs

logging/mlperf_logging/rcp_checker/rcp_checker.py

Lines 272 to 275 in 369260b

	min_epochs = self._find_min_acceptable_mean(
	record_contents['RCP Mean'],
	record_contents['RCP Stdev'],
	len(epoch_list)-samples_rejected*2)

And since we are no longer pruning based on min_epochs, it doesn't seem to have an effect on the results. The min_epochs later affects the Max Speedup, but this only later used in a condition to check if the RCP passed.

logging/mlperf_logging/rcp_checker/rcp_checker.py

Line 276 in 369260b

record_contents['Max Speedup'] = record_contents['RCP Mean'] / min_epochs

logging/mlperf_logging/rcp_checker/rcp_checker.py

Line 438 in 369260b

if mean_subm_epochs >= (rcp_record["RCP Mean"] / rcp_record["Max Speedup"]):

@ShriyaPalsamudram What changes were expected when changing the Stdev?

[ShriyaPalsamudram]

Since this impacts max speedup, can we compare max speedup before and after the change for all RCP points?

[pgmpablo157321]

[pgmpablo157321]

Additionally, an example of the max_speed_up values for last training results:
HPE-Cray-XD670-Gen11-H100-SXM5-80GB_n1_mxnet_24.04
With olympic score:

[1.018748075108887, 1.0601459916687128]

Without olympic score:

[1.0262860923600325, 1.0699036853548622]