- Preserve upstream R3 Tree-RL of BAGEL-7B-MoT: group size 16, policy group size 4, two rounds, seeds 4396 / 42, learning rate 5e-6, warm-up 10, AdamW beta2 0.95 and eps 1e-8, clipping 1.0, KL weights 0.001 and frozen ViT. Do not change model, prompts/order, rewards, losses, optimizer, sampling or RNG consumption for speed. No FP8, INT8 or FP16.
- The current result is fusion_b (
1367ad7) on eight H100 80 GB GPUs, with ZeRO-3 and the same loading adapter on baseline and candidate. Runtime Python files under modeling/ and train/ must match docs/validation/h100.json unless a new candidate is being evaluated. - Headline 2.0831x includes the reward server configuration change. Training-code-only speedup under the faster reward configuration is 1.8803x. Each is the minimum of three paired window ratios from fresh baseline A -> candidate -> baseline B launches (16 steps, 10 warm-up, three two-step windows). Never multiply independent speedups or mix the old four-GPU 1.4709x result with this one.
- Numerical evidence covers deterministic steps 0-13 on all eight ranks, including final parameter/Adam digests. Rank-0 wandb lists were empty; evaluation-only values were not compared. The owner accepts the completed bit-exact checks as the numerical acceptance criterion for this release; do not require or schedule additional long-training consistency/quality runs. Keep the 14-step scope explicit and do not present it as a measurement of full-training time or quality. Production timing values were not numerically compared.
- New numerical changes must preserve arithmetic, reduction order and RNG draws unless a new protocol is explicitly authorized. Fused reductions depend on PyTorch 2.9.1 / Triton 3.5.1; rerun the GPU diagnostics after upgrades. Defaults of all fusion switches are the measured configuration.
- Graph captures must not consume extra RNG or run extra forwards. Graphs and pools must not outlive each call because FSDP reshards and weights change. Preserve the static graph diagnostic output.
- Memory is tight: report allocated and reserved peaks for residency/lifetime changes. Eight-GPU baseline and candidate use the same ZeRO-3 adapter; the old four-GPU memory figures do not apply.
- Respect shared-host GPU locks and never stop another user's job. Rerecord baseline twice after changing host. Keep the complete judge external; do not claim that the portable launcher has been benchmarked unless it has.
- Use public datasets/checkpoints only. Never commit checkpoints, generated images, reward traces, profiler traces, credentials, private paths, host names or per-file data listings. Keep aggregate public benchmark evidence and source provenance in docs/validation/.