DaoyuanLi2816 · Aug 29, 2026 · Aug 29, 2026 · Aug 29, 2026 · Aug 29, 2026 · Aug 29, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -71,12 +71,16 @@ All notable changes to miniVERL are recorded here. The format follows
 - Published the exact-wheel 24-cell RTX 4080 comparison for `hf_reference`,
   `hf_cached` and managed vLLM across prompt lengths 128/512, response bounds
   64/256/512, `n=1/4`, greedy and seeded sampling.
-- Selected vLLM 0.28.0 for the direct-GKD development path after a measured
-  1.65–4.55× advantage over `hf_cached` at 256/512 tokens, eight confirmed
-  policy refreshes, 11,820 MiB peak total GPU memory and clean teardown.
-- Recorded the blocking negative result: `hf_cached` did not reach the
-  preregistered 2× speedup over `hf_reference`. v0.11.0 remains unpublished
-  until the local backend is optimized and the frozen workload is rerun.
+- Added an owned BF16 rollout mirror for NF4 training actors, exact live-LoRA
+  synchronization and StaticCache compilation. The final exact-wheel candidate
+  passed the `hf_cached >= 2× hf_reference` gate in every required cell.
+- Selected CUDA Graph vLLM 0.28.0 for the direct-GKD development path after a
+  measured 3.08–5.97× advantage over `hf_cached` at 256/512 tokens, eight
+  confirmed policy refreshes, 11,931 MiB peak total GPU memory and clean
+  teardown.
+- Preserved the first failed candidate as negative evidence and published the
+  qualified rerun as a separate identity-bound aggregate. PG-k1 remains on
+  `hf_cached` because vLLM did not pass log-probability conformance.
 
 ### Documentation narrative
 

diff --git a/PYPI.md b/PYPI.md
@@ -140,20 +140,20 @@ rewards and explicit group advantage composition; it has no task-quality claim.
 
 ## Measured systems evidence
 
-Rollout Runtime v2 now has a complete 24-cell RTX 4080 measurement across
-64/256/512-token responses, `n=1/4`, greedy and seeded sampling. Managed vLLM
-0.28.0 delivered 108.5–117.7 output tokens/s and was 1.65–4.55× faster than
-`hf_cached` in the 256/512-token cells, with 11.54 GiB peak total GPU memory,
-eight confirmed policy refreshes and clean teardown. It is the measured
-direct-GKD engine; PG-k1 stays on `hf_cached` because the engine log-probability
-probe exceeded its numerical threshold.
-
-The same evidence found the next optimization target: `hf_cached` reached
-about 1.8–1.9× on most seeded 256/512-token cells but did not meet the
-preregistered 2× gate, while the greedy path was usually slower than the legacy
-oracle. The development line therefore remains `0.11.0.dev0` rather than a
-published v0.11.0 release. See the [full runtime result](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/benchmarks/rollout-runtime-v2.md)
-for every cell and artifact hash.
+Rollout Runtime v2 has a complete 24-cell RTX 4080 measurement across
+64/256/512-token responses, `n=1/4`, greedy and seeded sampling. The compiled
+`hf_cached` path reached 124.2–248.7 output tokens/s and passed the
+preregistered 2× gate in every required cell. It keeps the NF4 actor for
+training and synchronizes each policy version into an owned BF16 rollout
+mirror.
+
+Managed vLLM 0.28.0 with CUDA Graph execution reached 626.6–836.4 output
+tokens/s and was 3.08–5.97× faster than `hf_cached` in the 256/512-token cells.
+Both backends passed memory, eight-refresh and teardown gates. vLLM is the
+measured direct-GKD engine; PG-k1 stays on `hf_cached` because its engine
+log-probability probe exceeded the numerical threshold. See the
+[full runtime result](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/benchmarks/rollout-runtime-v2.md) for every cell,
+the preserved first candidate and artifact hashes.
 
 The Qwen3-0.6B/1.7B developer workload consumed **32 distinct prompts**, used a
 64-token response bound, and completed **8 current-policy updates** at

diff --git a/README.md b/README.md
@@ -140,20 +140,20 @@ rewards and explicit group advantage composition; it has no task-quality claim.
 
 ## Measured systems evidence
 
-Rollout Runtime v2 now has a complete 24-cell RTX 4080 measurement across
-64/256/512-token responses, `n=1/4`, greedy and seeded sampling. Managed vLLM
-0.28.0 delivered 108.5–117.7 output tokens/s and was 1.65–4.55× faster than
-`hf_cached` in the 256/512-token cells, with 11.54 GiB peak total GPU memory,
-eight confirmed policy refreshes and clean teardown. It is the measured
-direct-GKD engine; PG-k1 stays on `hf_cached` because the engine log-probability
-probe exceeded its numerical threshold.
-
-The same evidence found the next optimization target: `hf_cached` reached
-about 1.8–1.9× on most seeded 256/512-token cells but did not meet the
-preregistered 2× gate, while the greedy path was usually slower than the legacy
-oracle. The development line therefore remains `0.11.0.dev0` rather than a
-published v0.11.0 release. See the [full runtime result](docs/benchmarks/rollout-runtime-v2.md)
-for every cell and artifact hash.
+Rollout Runtime v2 has a complete 24-cell RTX 4080 measurement across
+64/256/512-token responses, `n=1/4`, greedy and seeded sampling. The compiled
+`hf_cached` path reached 124.2–248.7 output tokens/s and passed the
+preregistered 2× gate in every required cell. It keeps the NF4 actor for
+training and synchronizes each policy version into an owned BF16 rollout
+mirror.
+
+Managed vLLM 0.28.0 with CUDA Graph execution reached 626.6–836.4 output
+tokens/s and was 3.08–5.97× faster than `hf_cached` in the 256/512-token cells.
+Both backends passed memory, eight-refresh and teardown gates. vLLM is the
+measured direct-GKD engine; PG-k1 stays on `hf_cached` because its engine
+log-probability probe exceeded the numerical threshold. See the
+[full runtime result](docs/benchmarks/rollout-runtime-v2.md) for every cell,
+the preserved first candidate and artifact hashes.
 
 The Qwen3-0.6B/1.7B developer workload consumed **32 distinct prompts**, used a
 64-token response bound, and completed **8 current-policy updates** at

diff --git a/README.zh-CN.md b/README.zh-CN.md
@@ -130,15 +130,15 @@ group advantage 组合；目前不对任务质量作结论。
 ## 实测系统证据
 
 Rollout Runtime v2 已在 RTX 4080 上完成 24-cell 测量，覆盖 64/256/512-token
-response、`n=1/4`、greedy 与 seeded sampling。Managed vLLM 0.28.0 达到
-108.5–117.7 output tokens/s；在 256/512-token cell 中比 `hf_cached` 快
-1.65–4.55×，peak total GPU memory 为 11.54 GiB，并通过 8 次 policy refresh 与
-完整 teardown。它是实测的 direct-GKD engine；engine log-probability probe 超出数值
-门槛，因此 PG-k1 继续使用 `hf_cached`。
-
-同一份证据也给出了下一步优化目标：`hf_cached` 在多数 seeded 256/512-token cell 中
-约为旧路径的 1.8–1.9×，没有达到预注册的 2× gate；greedy 路径通常比旧 oracle 更慢。
-因此开发线保持 `0.11.0.dev0`，尚未发布 v0.11.0。每个 cell 与 artifact hash 见
+response、`n=1/4`、greedy 与 seeded sampling。编译后的 `hf_cached` 达到
+124.2–248.7 output tokens/s，并在所有必需 cell 中通过预注册的 2× gate。训练仍以
+NF4 actor 为准，每个 policy version 都会精确同步到独立的 BF16 rollout mirror。
+
+启用 CUDA Graph 的 managed vLLM 0.28.0 达到 626.6–836.4 output tokens/s；在
+256/512-token cell 中比 `hf_cached` 快 3.08–5.97×。两个 backend 都通过显存、8 次
+policy refresh 与 teardown gate。vLLM 是实测的 direct-GKD engine；其 engine
+log-probability probe 超出数值门槛，因此 PG-k1 继续使用 `hf_cached`。每个 cell、保留的
+第一版失败候选与 artifact hash 见
 [完整 runtime 结果](docs/benchmarks/rollout-runtime-v2.md)。
 
 Qwen3-0.6B/1.7B developer workload 消费 **32 个不同 prompt**，response 上限为 64 token，

diff --git a/benchmarks/README.md b/benchmarks/README.md
@@ -80,6 +80,7 @@ Listed so you can check that this repository does not.
 
 | File | Hardware | What it shows |
 | --- | --- | --- |
+| `results/rollout-runtime-v2-v0.11.0-candidate-rtx4080.json` | WSL2 RTX 4080 16 GB | Exact-wheel qualified candidate: `hf_cached` passed the 2× local gate in every required cell; CUDA Graph vLLM passed the direct-GKD value gate. Both passed refresh, memory and teardown checks. [Result, method and hashes.](../docs/benchmarks/rollout-runtime-v2.md) |
 | `results/rollout-runtime-v2-rtx4080.json` | WSL2 RTX 4080 16 GB | Exact-wheel 24-cell `hf_reference` / `hf_cached` / vLLM comparison. vLLM passed the direct-GKD engine gates; `hf_cached` missed the preregistered 2× speed gate, so v0.11.0 remains blocked. [Result, method and hashes.](../docs/benchmarks/rollout-runtime-v2.md) |
 | `results/rollout-runtime-v2-hf-reference.json` | WSL2 RTX 4080 16 GB | Pre-v0.11 actor-rollout systems baseline: all 24 preregistered prompt/response/n/sampling cells completed. It freezes throughput and memory for the future `hf_cached` comparison; teacher scoring, updates, full-cycle timing and typed grouped semantics are not measured here. [Method, table and boundaries.](../docs/benchmarks/rollout-runtime-v2.md) |
 | `results/alignment-external-v1.json` | RTX 4080 16 GB | **Preregistered early stop, not a method comparison.** Two declared lineages and eight candidates all measured 0/64 retained JSONNav utility against an unchanged 20% floor. No checkpoint, teacher, continuation arm or final-test result exists. [Study, correction provenance and evaluator boundary.](../docs/alignment-external/alignment-external-v1.md) |