dlrmv4: replace the GBS 8192 RCP with a GB300 rerun of the reference

[mmarcinkiewicz]

Follow up to mlcommons/training#908

GBS16k and GBS32k look alright. The proposed RCPs represent the true distribution very well

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅

[ShriyaRishab]

WG meeting notes (10/1/26):

1. AMD concern is that a set of submission runs like the previous RCPs fails the current RCPs?
2. Changing very close to the submission is a concern from AMD
3. Keeping new RCPs saves us from cherry-picking and unlucky/unfairness

WG decided to require N=12 runs (drop 2 slowest and 2 fastest) instead of N=10 to work around this. Apply this for all GBS. Need to apply logging and rules. @matthew-frank will make these PRs.

[mrmhodak]

@mmarcinkiewicz and @ShriyaRishab:

We have had an extensive internal discussion on this:

1. Changing RCPs this close is disruptive, our optimizations have assumed the RCP convergence
2. MI355X GPUs is the reference HW for this model and thus the RCP runs should come from MI355X and not from GB300.

Based on that, we are against changing the RCPs on a short notice.

However, the longer runs are a concern. Our offer to improve this is to do a 1/3 removal out of 12: Remove 1 shortest run and 3 longest.

[mmarcinkiewicz]

Miro,

Changing RCPs this close is disruptive, our optimizations have assumed the RCP convergence

this is actually you shooting yourself in the foot - the submission is supposed to be run on random seeds, so using fixed, you would miss (as we have till now) that there are bad seeds till you actually see them in your submission - what would you do then? Just rerun with a different seed? This is something we'd like to avoid for everyone.

However, what we can do one time, and on a non-precedent setting basis, to allow you (and everyone else) to run with the same seeds the submission is using, so everyone converges more or less the same. What do you think about that?