Skip to content

Add qwen35_397b_grpo reference RCP for BS=1152 #485

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
ShriyaRishab merged 2 commits into from
Oct 27, 2026
Merged

Add qwen35_397b_grpo reference RCP for BS=1152 #485

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
f798681
Add qwen35_397b_grpo reference RCP for BS=1152
uwlei Oct 20, 2026
97f6e28
Merge branch 'master' into add-qwen35-397b-grpo-rcp-1152
ShriyaRishab Oct 27, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
20 changes: 20 additions & 0 deletions mlperf_logging/rcp_checker/training_6.1.0/rcps_qwen35_397b_grpo.json
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -57,5 +57,25 @@
"Epochs to converge": [
7168, 7168, 7168, 7168, 7168, 7168
]
},

"qwen35_397b_grpo_ref_1152":
{
"Benchmark": "qwen35_397b_grpo",
"Creator": "Google",
"When": "Reference RCPs at BS=1152 on 64xGB200 using MLCommons reference code",
"Platform": "64 NVIDIA GB200 nodes",
"Precision": "BF16",
"BS": 1152,
"Hyperparams": {
"opt_base_learning_rate": 2.1213203436e-06,
"opt_gradient_clip_norm": 0.0589255651,
"num_prompts_per_step": 72,
"num_generations_per_prompt": 16,
"gradient_accumulation_steps": 144
},
"Epochs to converge": [
6912, 6912, 6912, 6912, 6912, 6912

@pavanky pavanky Oct 23, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I am not too familiar with the benchmark, but the epochs to converge was generally increasing with batch size, but the epochs to converge here goes down from 1024. Is this expected or just some corner case we are hitting?

@RissyRan RissyRan Oct 27, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for the question @pavanky! For RL evaluation, the rule is to perform offline on checkpoints starting at global_batch_size * CEIL(2.5 + 3840 / global_batch_size) samples and every step after that. With a global_batch_size of 1152, the evaluation began at 6912 samples; our implementation adheres to this rule.

]
}
}
Loading