Rollout Runtime v2: baseline and selected-backend evidence

The pre-v0.11 Hugging Face path is now frozen as the compatibility and performance baseline for Rollout Runtime v2. All 24 preregistered cells completed on one WSL2 RTX 4080 without an OOM or discarded cell.

This result measures actor rollout only. It is not a task-quality experiment, does not include teacher scoring or an optimizer update, and does not yet exercise typed grouped-sample semantics.

Frozen setup

Item	Exact value
Source	0d6e0070ae73ef35f718aec3624ee5263ac96e3a
Candidate wheel	0256bd9e63ca6ed52999a5073a3577a008581a8d4418062314750d86f21cd5fe
Actor	Qwen/Qwen3-0.6B at c1899de289a04d12100db370d81485cdf75e47ca
Weight payload digest	b29cd98b83f9bddc7ec8943be5f142243e956f448f13456847849e5b8615b413
Runtime	Python 3.12.13, Torch 2.13.0+cu130, Transformers 5.14.1, NF4/BF16
GPU / driver	NVIDIA GeForce RTX 4080 16 GiB / 596.49
Workload	4 logical prompts; prompt 128/512; response 64/256/512; n=1/4; greedy and seeded stochastic

Measured rollout throughput

Every row generated the full configured response bound. Rates use actual generated tokens and median wall time over three measured repetitions after one warmup.

Prompt	Response	Samples/prompt	Greedy tokens/s	Seeded stochastic tokens/s	Peak reserved GiB
128	64	1	95.42	29.93	1.053
128	64	4	97.07	29.95	1.053
128	256	1	95.87	29.95	1.084
128	256	4	92.35	30.07	1.084
128	512	1	93.16	30.23	1.086
128	512	4	92.91	30.30	1.086
512	64	1	92.68	30.28	1.086
512	64	4	83.81	30.21	1.086
512	256	1	80.10	30.31	1.086
512	256	4	77.24	30.06	1.086
512	512	1	66.28	29.92	1.086
512	512	4	66.26	30.31	1.086

The two sampling paths have different pre-v0.11 implementations. Greedy uses one padded batch but recomputes the full prefix at each step; seeded stochastic preserves per-request RNG by running the established sequential KV-cache path. These measurements are therefore separate baselines, not an algorithm ranking.

What remains unmeasured

The legacy public backend does not expose clean prefill, decode or time-to-first token boundaries, so those fields are not_measured, not zero. Policy sync is not_applicable for an actor already loaded in the same process. Teacher scoring, actor update and full-cycle time are also not_measured in this rollout-only baseline and remain required before v0.11 release.

n=4 here is four independently repeated requests per prompt. Transactional group identity, grouped resume and reward composition do not exist in this baseline and cannot be inferred from it.

Artifact integrity

Artifact	SHA-256
Preregistration	8cc3ba738c69b59ed19c22c1de874fd00249404198a3e05983477dc8899bb7e5
Raw measurement	2e303eabb559b843d25377a7c72e0aeb0219eec7eb8bc109c0419839b3251170
Final identity-bound result	b25daee7ee726b7a7be18d7dbe26590fb61325bf212cfd9e7110b69f5fa8889c
Frozen calculator benchmark	53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc

See the method for timing and invalidation rules. The next backend must use this workload unchanged and pass the preregistered correctness, memory and 256/512-token speedup gates.

First development candidate (preserved negative result)

The exact 0.11.0.dev0 wheel at source 690db9b079bfecfec14dc3bedc3aa0308cbacf60 ran the same 24 cells with hf_cached and managed vLLM 0.28.0. Both runs used wheel SHA-256 266fcd59bb3e85b02974a92b26c9a4e59c7c9b9205853d17baadcfffdb898e27 and the baseline software stack: Python 3.12.13, Torch 2.13.0+cu130, Transformers 5.14.1, PEFT 0.20.0 and bitsandbytes 0.50.2.

Response	Sampling	hf_cached / reference	vLLM / hf_cached	vLLM tokens/s
64	greedy	0.65–0.76×	1.73–1.80×	112.7–114.9
64	seeded stochastic	1.83–1.87×	1.94–2.06×	109.3–114.2
256	greedy	0.71–0.85×	1.71–1.77×	111.8–117.4
256	seeded stochastic	0.84–1.85×	2.04–4.55×	113.1–114.5
512	greedy	0.67–1.05×	1.65–1.87×	114.8–117.7
512	seeded stochastic	1.84–1.90×	1.90–2.05×	108.5–114.4

Ranges contain both prompt lengths and both n values. One hf_cached p512/r256/n4 stochastic cell had measured runs of 118.807, 267.930 and 162.641 seconds; the median and resulting 0.84× speedup remain in the result.

Gate outcome

• hf_cached >= 2× hf_reference at response 256 and 512: failed. The conservative preregistration interpretation requires every paired cell to reach the threshold.
• vLLM >= 1.2× hf_cached at response 256 and 512: passed in every cell; the observed range was 1.65–4.55×.
• Peak total GPU memory below 14.5 GiB: passed. hf_cached peaked at 4,598 MiB and vLLM at 11,820 MiB.
• Eight vLLM refreshes: passed. Every sync and identity was unique, with no monotonic memory growth; the localhost port closed at teardown.
• vLLM PG-k1 log-probability conformance: failed. Greedy token agreement was 32/32, while maximum absolute log-probability difference was 0.0194 against the 0.01 NF4 threshold. Direct GKD remains the supported scope.

This candidate failed the local-backend speed gate and triggered the profiling work that produced the qualified candidate below. Its bytes and failed outcome remain part of the record.

Candidate artifact integrity

Artifact	SHA-256
hf_cached raw	a9f8f0b0275d940f30497a0d88d76da0e112ffbe2bc1b37a42c6ed313b852242
vLLM raw	abf14d73a7289f9619515943b4c8d7dafe7cd187fefe928655739ac7ed1ab16c
Aggregate result	32be86856263ccdf787986c4ec54570d323af8c238dc1664e00a9ce41ca393c4
Frozen calculator	53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc

The aggregate JSON contains every cell, exact raw hashes, gate calculation and backend-selection state. scripts/publish_rollout_runtime_v2_evidence.py --check reproduces it.

Qualified v0.11 candidate

The exact 0.11.0.dev0 wheel at source 3cfa09b0b7aa2ee63f5702a5766b73d9a32fa8b9 reran the same 24 cells. Both backends used wheel SHA-256 33c30834dcd01001a5c2ae619a1c982ceb2ad5fcc457998baa9f960a87267eba. The local backend keeps the NF4 actor as the training source of truth, builds a BF16 inference mirror, and synchronizes the exact live LoRA tensors before each policy version. The managed vLLM backend uses CUDA Graph execution.

Response	hf_cached / reference	vLLM / hf_cached	hf_cached tokens/s	vLLM tokens/s
64	2.39–4.30×	2.99–5.71×	124.2–242.5	626.6–806.0
256	2.38–4.30×	3.19–5.97×	125.9–240.8	685.6–836.4
512	2.54–4.23×	3.08–5.88×	124.3–248.7	693.4–821.2

Ranges include both prompt lengths, both n values and both sampling modes. Each cell used one warmup and three measured repetitions. Compilation and engine warmup happen before measured repetitions and remain visible in the raw artifact.

Gate outcome

• hf_cached >= 2× hf_reference at response 256 and 512: passed in every cell, with conservative minima of 2.375× and 2.541×.
• vLLM >= 1.2× hf_cached at response 256 and 512: passed in every cell; the full observed range was 3.076–5.966×.
• Peak total GPU memory below 14.5 GiB: passed. hf_cached peaked at 6,504 MiB and vLLM at 11,931 MiB.
• Exact policy refresh and teardown: passed for both backends across eight unique policy identities, without monotonic memory growth.
• Local NF4/mirror conformance: passed with exact greedy tokens and maximum sampled-token log-probability difference 0.002483 against the 0.01 limit.
• vLLM PG-k1 log-probability conformance: failed closed. Greedy tokens were exact, but the maximum difference was 0.019646. vLLM is therefore selected for direct GKD only; PG-k1 stays on hf_cached.

The runtime performance gate is complete. Release publication still requires the exact-wheel qualification and release workflow.

Qualified-candidate artifact integrity

Artifact	SHA-256
hf_cached raw	dd29402642ab68f7854579bd68accb4b449ca0a2ec4cb8b90d834ee2e6f03757
vLLM CUDA Graph raw	00c55a25deb9ecf7786d7f54ccba4c5c5b0bcce4bf1e46fe8ac6b48d0596194c
Aggregate result	0fb2066a8b567adbb011113141b2b4e3a4fa281cf3f3d13f6eecd1d7b0499a5d
Frozen calculator	53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc

The qualified aggregate JSON binds every cell to the exact wheel, raw files and preregistration. scripts/publish_rollout_runtime_v2_candidate.py --check reproduces it.