[syangx39]

Summary

Adds 6 independent reference convergence point (RCP) logs for qwen35_397b_grpo at global_batch_size = 1152 (72 prompts × 16 generations) under llm_post_training/rcp_logs/GBS1152/seed_{1..6}.out, produced using the unmodified MLCommons public reference implementation (NVIDIA-NeMo/RL commit 3fca04c9b313d313302923a5bb6b0c8dc0340ed6).

Related issues / PRs:

• qwen35_397b_grpo Missing RCP error for global_batch_size = 1152 logging#482
• Add qwen35_397b_grpo RCP for BS=1152 logging#483
• https://github.com/mlcommons/submissions_training_v6.1/issues/36

---

Reference Configuration (GBS = 1152)

• Reference Codebase: NVIDIA-NeMo/RL commit 3fca04c9b313d313302923a5bb6b0c8dc0340ed6 (llm_post_training/RL submodule)
• Hardware: 64 NVIDIA GB200 nodes (256 GPUs total: 16 Megatron-Core policy training nodes + 48 vLLM generation nodes)
• Precision: BF16 policy training + BF16 vLLM generation (in_flight_weight_updates: false, max_trajectory_age_steps: 1)
• Batch Shape: global_batch_size = 1152 (num_prompts_per_step = 72, num_generations_per_prompt = 16, micro_batch_size = 1, data_parallel_size = 8, gradient_accumulation_steps = 144)
• Scaled Hyperparameters (per benchmark rules):
  • opt_base_learning_rate = 2.121320343559642e-06 (1.0e-6 * sqrt(1152 / 256))
  • opt_gradient_clip_norm = 0.05892556509887896 (0.125 * sqrt(256 / 1152))
• Validation Schedule:
  • First validation step: Step 6 (ceil(2.5 + 3840 / 1152) = 6, 6912 samples)
  • Validation period: Every step (val_period = 1) on all 251 held-out validation tasks (pass@4 target > 0.69)

---

6-Seed Convergence Results (BS = 1152)

All 6 / 6 seeds converge at the first scheduled validation (Step 6 = 6912 samples):

File	Seed	Converged Step	Samples to Converge	eval_accuracy (pass@4)	Wallclock (run_start to run_stop)	SHA-256
llm_post_training/rcp_logs/GBS1152/seed_1.out	1885626193	Step 6	6912	0.75298804	9023.62 s (150.4 min)	5191a9141deeafe149ea76f6b25012e2cefc2fc3cc2fcb9918bcdd614d1ac4b9
llm_post_training/rcp_logs/GBS1152/seed_2.out	1885626194	Step 6	6912	0.76494026	8777.35 s (146.3 min)	3e31ecadfd2979455735a8f3ec9efcfca237f301d8fe6c2bcaba5b992e06230a
llm_post_training/rcp_logs/GBS1152/seed_3.out	1885626195	Step 6	6912	0.73705178	8989.20 s (149.8 min)	006f7b03954af19103db0d37c92f65f98f4c481568042e90f9d2ed36683b4212
llm_post_training/rcp_logs/GBS1152/seed_4.out	1885626196	Step 6	6912	0.76095617	8634.54 s (143.9 min)	8e15c28685e4356fbe64d23f1fd56543a9943e9cf9b67ef69ebaae1eed35d5f1
llm_post_training/rcp_logs/GBS1152/seed_5.out	1885626197	Step 6	6912	0.78087652	9925.58 s (165.4 min)	247436664926889948b94be1cbb6a9cc698d2deeef563ad110232ba892c41e06
llm_post_training/rcp_logs/GBS1152/seed_6.out	1885626199	Step 6	6912	0.72908366	8702.14 s (145.0 min)	9a2478b75a152cf8891d0947a46bd421927dbf0d706391027405e92c0296afd3

• Epochs (samples) to converge: [6912, 6912, 6912, 6912, 6912, 6912]
• Observed pass@4 range at Step 6: 0.72908366 - 0.78087652 (mean 0.7543)

[github-actions]

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

[uwlei]

nit: do you want to include 1152 here?

[uwlei]

nit: should be the consistent format?

[pavanky]

I merged without the nits being addressed in favor of not blocking for now. The nits can be addressed separately.