[0z5a]

Incremental review: Files changed.

Direct dependency and ancestor review: #3 → #1. Baseline: cleaned #3 head 94129cedf3a91edd92d16bbd3499258b93590cff. This fork PR shows only this layer.

Adds a synchronous global-budget Ouro GRPO recipe using VIME's existing Megatron optimizer, loss, gradient reduction and checkpoint path. Each update uses one deterministic K on every DP rank; prefill, decode and training execute the same budget. Group identities and checkpoint schedule/data contracts are preserved.

Historical fixed-budget rollout companion: 0z5a/vllm-rlt#1. The public-contract follow-up is #5, consolidated upstream in #459.

Variant	Mean step (s)	Ratio vs K=4 / latency reduction	Job (s) / GPU-h	Training reward	Truncated	Mean response tokens
fixed4	44.526	1.000× / 0.00%	633.63 / 0.3520	1/96	73/96	436.1
fixed2	22.786	1.954× / 48.83%	499.86 / 0.2777	0/96	61/96	378.0
fixed3	33.568	1.326× / 24.61%	556.56 / 0.3092	1/96	70/96	428.5
mixed	33.415	1.333× / 24.96%	558.63 / 0.3104	1/96	78/96	450.6

2×H20, full Ouro-1.4B, six updates per arm, one seed, 16 completions/update and a 512-token cap. Step ratios include generation/scoring/training/publication, not evaluation or checkpoint saving. They do not establish equal-quality acceleration. Fixed4 finishes at 1/4 held-out success at K=4; every final mixed evaluation is 0/4. The mixed pilot has not established the quality/compute tradeoff needed to justify adding a bucket scheduler.

14 budget/normalization tests, provider/HF numerical checks, real two-rank schedule checks, tiny mixed-depth save/resume and full-model K=4 save/resume pass. All four full-model arms finish with complete checkpoints and normal process exit. Completed-task model files have since been cleaned up.

Test scripts, validation reports and logs were removed. Mixed-depth bucketing, learned stopping and time-to-target remain unestablished.

[0z5a]

Superseded by vllm-project#459, now merged into 0z5a-branch. This historical Ouro layer is included in the consolidated upstream integration; closing the duplicate PR.