DaoyuanLi2816 · DaoyuanLi2816 · Aug 14, 2026 · Aug 14, 2026 · Aug 14, 2026 · Aug 14, 2026
diff --git a/PROJECT_STATE.md b/PROJECT_STATE.md
@@ -4,7 +4,7 @@ Current maintainer handoff for **miniVERL** (`mini-verl` package, `miniverl`
 CLI). `release-state.yaml` is the canonical version source; this page indexes
 current product and evidence state rather than repeating release history.
 
-Last updated: 2026-08-13.
+Last updated: 2026-08-14.
 
 Canonical release state: stable `v0.9.1` (`6c0f3d818c10419e0bfba81f3ad1c5adf24eaf09`), development `0.10.0.dev0`.
 
@@ -31,13 +31,12 @@ Executable compatibility claims are mutation-tested and recorded in
 unknown-size quantized roles require proof instead of receiving an executable
 plan.
 
-Development `0.10.0.dev0` now has a closed typed profile registry and
-torch-free compatibility introspection. New profile-scoped plans, caches,
-checkpoints and exports bind an independent identity; the only registered
-executable profile remains the measured direct-GKD path until PG-k1 conformance
-and runtime work is complete.
+Development `0.10.0.dev0` has a closed typed profile registry and torch-free
+compatibility introspection. Profile-scoped plans, caches, checkpoints and
+exports bind an independent identity. The direct-GKD and sampled-k1 vanilla
+policy-loss profiles both have pinned conformance and measured RTX 4080 paths.
 
-Arbitrary verl YAML, policy-gradient OPD, rewards, PPO/GRPO, Ray, FSDP,
+Arbitrary verl YAML, other policy-gradient modes, rewards, PPO/GRPO, Ray, FSDP,
 Megatron, multi-GPU and distributed execution remain unsupported. The legacy
 environment/PPO artifact bridge is migration-only.
 
@@ -46,7 +45,9 @@ environment/PPO artifact bridge is migration-only.
 | Evidence | Status |
 | --- | --- |
 | Qwen3-0.6B/1.7B developer workload | 32 prompts, 8 current-policy updates, 3.1914 GiB peak reserved on one RTX 4080; matched interruption/resume was byte-identical |
-| SmolLM2-360M/1.7B | one pinned compatibility smoke; not a second measured recipe |
+| Qwen3 sampled-k1 PG | 32 prompts, 8 updates, exact interruption/resume, 3.1914 GiB peak reserved; no quality comparison |
+| SmolLM2-360M/1.7B direct GKD | 32 prompts, 8 updates, 1.4961 GiB peak reserved; exact resume, PEFT reload and materialized export passed |
+| Ubuntu 26.04 WSL2 | plan, bounded probe, rollout, teacher scoring, one update and PEFT reload measured on the same RTX 4080 |
 | external alignment v1 | preregistered early stop: 0 selected checkpoints, teachers, continuation arms or final-test accesses |
 | distributed verl execution | not tested |
 

diff --git a/PYPI.md b/PYPI.md
@@ -163,8 +163,10 @@ and update times were 9.7200, 0.4864 and 2.3260 seconds. A matched 4-update
 interruption resumed to the same byte-identical trajectories, adapter and
 optimizer tensors. See the [data-bound figure and full record](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/verl-opd-reference-workload.md);
 the original one-update [pip smoke](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/opd-quickstart.md) remains preserved.
-A separate pinned SmolLM2-360M/1.7B compatibility smoke completed one full
-rollout/scoring/update cycle; it is not a second measured recipe.
+A separate pinned SmolLM2-360M/1.7B recipe consumed 32 distinct prompts across
+8 updates at 1.4961 GiB peak reserved VRAM. Interruption/resume was
+byte-identical, PEFT reload passed, and the exact-snapshot scale-out bundle
+materialized successfully. See the [full SmolLM2 systems record](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/smollm2-opd-workload.md).
 
 This is deliberately a runtime and artifact proof. It is not a throughput
 benchmark, an alignment-quality endpoint, or evidence that OPD beats SFT, DPO

diff --git a/README.md b/README.md
@@ -163,8 +163,10 @@ and update times were 9.7200, 0.4864 and 2.3260 seconds. A matched 4-update
 interruption resumed to the same byte-identical trajectories, adapter and
 optimizer tensors. See the [data-bound figure and full record](docs/verl-opd-reference-workload.md);
 the original one-update [pip smoke](docs/opd-quickstart.md) remains preserved.
-A separate pinned SmolLM2-360M/1.7B compatibility smoke completed one full
-rollout/scoring/update cycle; it is not a second measured recipe.
+A separate pinned SmolLM2-360M/1.7B recipe consumed 32 distinct prompts across
+8 updates at 1.4961 GiB peak reserved VRAM. Interruption/resume was
+byte-identical, PEFT reload passed, and the exact-snapshot scale-out bundle
+materialized successfully. See the [full SmolLM2 systems record](docs/smollm2-opd-workload.md).
 
 This is deliberately a runtime and artifact proof. It is not a throughput
 benchmark, an alignment-quality endpoint, or evidence that OPD beats SFT, DPO

diff --git a/README.zh-CN.md b/README.zh-CN.md
@@ -138,8 +138,9 @@ Qwen3-0.6B/1.7B developer workload 实际消费 **32 个不同 prompt**，respon
 匹配的运行在第 4 次 update 后中断并恢复；最终 trajectory、adapter 与 optimizer tensor
 均字节一致。详见[数据绑定图与完整记录](docs/verl-opd-reference-workload.md)；原始一次更新的
 [pip smoke](docs/opd-quickstart.md)仍完整保留。
-另一个固定版本的 SmolLM2-360M/1.7B 兼容性 smoke 完成了一次完整的
-rollout/scoring/update 循环；它不是第二套完整实测 recipe。
+另一套固定版本的 SmolLM2-360M/1.7B recipe 消费 32 个不同 prompt 并完成 8 次
+update，peak reserved VRAM 为 1.4961 GiB；中断恢复保持字节一致，PEFT 重载与精确
+snapshot 的 scale-out 物化均通过。详见[完整 SmolLM2 系统记录](docs/smollm2-opd-workload.md)。
 
 这只证明一个运行时与产物路径，不是吞吐 benchmark、对齐质量 endpoint，也不证明 OPD
 优于 SFT、DPO 或 KD。其他 NVIDIA GPU 使用相同的 device-name-agnostic CUDA 路径，但

diff --git a/benchmarks/results/rtx4080-smollm2-opd-developer-v1.json b/benchmarks/results/rtx4080-smollm2-opd-developer-v1.json
@@ -0,0 +1,180 @@
+{
+  "artifacts": {
+    "checkpoint_hashes": {
+      "adapter.safetensors": "7d00e5e99ac19ae8c35055e6dee527b737c3af3504b846feef60a0e60d21856d",
+      "optimizer.safetensors": "1b3473085336cca8b41458c300df368979cb625520462a9152a4d49031b6dd12",
+      "state.json": "a8f459ce233bd59c011ed31f9b9b308d10572865a2227e0f6e77c04d2de56307"
+    },
+    "standard_peft_adapter_sha256": "6132c9aa94b828deed70d03108805beece3392e50409395423a8ae092aac3e4f",
+    "standard_peft_load_verified": true,
+    "total_workload_bytes": 4214630242,
+    "trajectory_sha256": "f8137936991a17d1531352534eb6e5dd4428e2495a0f5de0b9c68fc21d63bb43"
+  },
+  "hardware": {
+    "bitsandbytes": "0.50.0",
+    "cuda_runtime": "13.0",
+    "driver": "596.49",
+    "gpu": "NVIDIA GeForce RTX 4080",
+    "gpu_count": 1,
+    "peft": "0.18.0",
+    "torch": "2.13.0+cu130",
+    "transformers": "5.14.1",
+    "vram_gib": 15.992
+  },
+  "kind": "single_gpu_smollm2_direct_gkd_developer_workload",
+  "measured_at": "2026-08-14T07:46:06Z",
+  "measurements": {
+    "adapter_bytes": 6825692,
+    "batch_downshifts": {
+      "rollout_oom": 0,
+      "update_chunk_oom": 0
+    },
+    "cache_bytes": 81819,
+    "checkpoint_bytes": 9912101,
+    "cold_startup_seconds": 8.4132,
+    "distinct_prompts_consumed": 32,
+    "observed_rollout_physical_batch_sizes": [
+      4
+    ],
+    "peak_allocated_gib": 1.4254,
+    "peak_reserved_gib": 1.4961,
+    "prompts_consumed": 32,
+    "steady_state_median_seconds": {
+      "actor_update": 1.127,
+      "rollout": 16.3067,
+      "teacher_scoring": 0.1941
+    },
+    "steady_state_median_throughput": {
+      "rollout_tokens_per_second": 13.7173,
+      "teacher_scored_positions_per_second": 1190.38,
+      "update_positions_per_second": 186.9733
+    },
+    "time_to_first_rollout_seconds": 21.6392,
+    "time_to_first_teacher_targets_seconds": 21.8723,
+    "time_to_first_update_seconds": 23.1082,
+    "total_run_bytes": 27178611
+  },
+  "miniverl_version": "0.10.0.dev0",
+  "models": {
+    "student": {
+      "adapter": null,
+      "attn_implementation": "sdpa",
+      "dtype": "auto",
+      "gradient_checkpointing": true,
+      "lora": {
+        "alpha": 16,
+        "bias": "none",
+        "dropout": 0.0,
+        "enabled": true,
+        "r": 8,
+        "target_modules": [
+          "q_proj",
+          "v_proj"
+        ]
+      },
+      "model_id": "HuggingFaceTB/SmolLM2-360M-Instruct",
+      "prepare_kbit_training": true,
+      "quantization": "nf4",
+      "revision": "a10cc1512eabd3dde888204e902eca88bddb4951",
+      "tokenizer_id": null,
+      "tokenizer_revision": null,
+      "toy": {
+        "hidden_size": 64,
+        "intermediate_size": 128,
+        "max_position_embeddings": 1024,
+        "num_heads": 4,
+        "num_layers": 2
+      },
+      "trust_remote_code": false
+    },
+    "teacher": {
+      "adapter": null,
+      "attn_implementation": "sdpa",
+      "dtype": "bfloat16",
+      "mode": "standard",
+      "model_id": "HuggingFaceTB/SmolLM2-1.7B-Instruct",
+      "quantization": "nf4",
+      "revision": "31b70e2e869a7173562077fd711b654946d38674",
+      "tokenizer_id": null,
+      "tokenizer_revision": null,
+      "toy": {
+        "hidden_size": 64,
+        "intermediate_size": 128,
+        "max_position_embeddings": 1024,
+        "num_heads": 4,
+        "num_layers": 2
+      },
+      "toy_pretrain_lr": 0.003,
+      "toy_pretrain_steps": 0,
+      "toy_teacher_seed": 99,
+      "trust_remote_code": false
+    }
+  },
+  "profile": "verl-opd-v0.8-single-gpu-v1",
+  "profile_identity": {
+    "digest": "cc30bae03b65fe1bb915990fbd0359bbff436332b1b44c4887bd77ae8e0e4c35",
+    "export_version": "verl-opd-export-v1",
+    "field_rule_digest": "8479344d1d033974b2bbc3a8212c0a01bcc54afbe697ded1987794b7a458ee97",
+    "loss_conformance_version": "forward-kl-topk-verl-v0.8-v1",
+    "native_compiler_version": "direct-gkd-native-v1",
+    "profile_name": "verl-opd-v0.8-single-gpu-v1",
+    "profile_schema_version": 1,
+    "upstream_commit": "7aed6b230776f963fa09509c10d9c3a767d1102c",
+    "upstream_repository": "https://github.com/verl-project/verl",
+    "upstream_tag": "v0.8.0"
+  },
+  "recipe": {
+    "compiled_plan_sha256": "d3ff5ab77f2c57d89a2fb52a04a842cbc06b1c2ca5350a82313aadb6131d6067",
+    "input_parquet_sha256": "baa99ba8c59d4c30b2e3081346860e7e8b2665dd70fdbcd7952935545a0cd770",
+    "interrupt_after_updates": 4,
+    "logical_batch": 4,
+    "optimizer_updates": 8,
+    "prompt_limit": 128,
+    "prompts_available": 64,
+    "prompts_consumed": 32,
+    "response_limit": 64,
+    "teacher_target": "top_k_ids_log_probs_and_mass",
+    "top_k": 32
+  },
+  "resource_contract": {
+    "gpu_hour_limit": 3.0,
+    "peak_reserved_limit_gib": 14.5,
+    "peak_reserved_within_limit": true,
+    "total_gpu_workload_seconds": 301.27
+  },
+  "resume": {
+    "adapter_and_optimizer_byte_identical": true,
+    "excluded_run_identity_field": "resolved_config_digest",
+    "global_optimizer_steps": 8,
+    "interrupt_checkpoint_bytes": 9912101,
+    "resume_load_seconds": 3.1039,
+    "status": "exact_match",
+    "task_cursor": 32,
+    "training_state_fields_identical": true,
+    "trajectories_byte_identical": true
+  },
+  "scaleout": {
+    "artifact_bundle_complete": true,
+    "bundle_bytes": 4160177169,
+    "distributed_execution_tested": false,
+    "launchable": true,
+    "model_data_load_smoke_passed": true,
+    "upstream_config_parse_passed": true
+  },
+  "schema_version": 1,
+  "scientific_scope": {
+    "algorithm_comparison": false,
+    "alignment_quality_evaluated": false,
+    "runtime_correctness_only": true,
+    "task_quality_evaluated": false
+  },
+  "source_commit": "6cf11cfacb455c872559dc1ca48d928594556b15",
+  "status": "maintainer_measured",
+  "verl": {
+    "commit": "7aed6b230776f963fa09509c10d9c3a767d1102c",
+    "distributed_execution_tested": false,
+    "repository": "https://github.com/verl-project/verl",
+    "tag": "v0.8.0"
+  },
+  "workload_driver_sha256": "f249766c774e1b44e09ff323d22f548aa5ca9bfa782f54bfbf9ca8782ce59fcd"
+}
diff --git a/docs/evidence/wsl2-rtx4080-smollm2-opd-v1.json b/docs/evidence/wsl2-rtx4080-smollm2-opd-v1.json
@@ -0,0 +1,71 @@
+{
+  "schema_version": 1,
+  "kind": "wsl2_single_gpu_opd_smoke",
+  "status": "maintainer_measured",
+  "measured_at": "2026-08-14T07:49:21Z",
+  "source_commit": "6cf11cfacb455c872559dc1ca48d928594556b15",
+  "miniverl_version": "0.10.0.dev0",
+  "profile": "verl-opd-v0.8-single-gpu-v1",
+  "profile_identity_digest": "cc30bae03b65fe1bb915990fbd0359bbff436332b1b44c4887bd77ae8e0e4c35",
+  "platform": {
+    "distribution": "Ubuntu 26.04 LTS",
+    "kernel": "6.18.33.2-microsoft-standard-WSL2",
+    "python": "3.12.13",
+    "gpu": "NVIDIA GeForce RTX 4080",
+    "vram_gib": 15.992,
+    "driver": "596.49",
+    "torch": "2.13.0+cu130",
+    "cuda_runtime": "13.0",
+    "transformers": "5.15.0",
+    "bitsandbytes": "0.50.1",
+    "peft": "0.20.0"
+  },
+  "models": {
+    "student": {
+      "id": "HuggingFaceTB/SmolLM2-360M-Instruct",
+      "revision": "a10cc1512eabd3dde888204e902eca88bddb4951",
+      "quantization": "nf4"
+    },
+    "teacher": {
+      "id": "HuggingFaceTB/SmolLM2-1.7B-Instruct",
+      "revision": "31b70e2e869a7173562077fd711b654946d38674",
+      "quantization": "nf4"
+    }
+  },
+  "probe": {
+    "status": "measured",
+    "parameter_updates": 0,
+    "duration_seconds": 38.648,
+    "peak_reserved_gib": 1.0449
+  },
+  "execution": {
+    "plan_digest": "a7e941448a73562235a5b05ad8294dcfff5c508c1424adfbddf908523bb9075e",
+    "prompts": 1,
+    "response_limit": 64,
+    "generated_tokens": 42,
+    "optimizer_updates": 1,
+    "time_to_first_update_seconds": 9.5106,
+    "rollout_tokens_per_second": 12.84,
+    "teacher_scored_positions_per_second": 313.15,
+    "update_positions_per_second": 75.1401,
+    "peak_allocated_gib": 1.3808,
+    "peak_reserved_gib": 1.4043,
+    "peft_export_and_reload": "passed"
+  },
+  "artifacts": {
+    "checkpoint_adapter_sha256": "0ad396b576ab89bba894407331bedff529272dbc80a10d5b73935f42f0d444de",
+    "checkpoint_optimizer_sha256": "b41f8da8946c3d6691bbf9266866a95a7fbd0e76a0d927a56f802749c4d4cd0e",
+    "trajectory_sha256": "0adbe928a45d5aba8c2c85234fcbba246b4c1bbd1c48397c4f6ea584d701d3e3",
+    "peft_adapter_sha256": "9429bc79db18ac773a65c7f6a776eef45f9275f7a7f5c783ccd10286d5127c55"
+  },
+  "scope": {
+    "plan_completed": true,
+    "probe_completed": true,
+    "rollout_completed": true,
+    "teacher_scoring_completed": true,
+    "optimizer_update_completed": true,
+    "peft_reload_completed": true,
+    "task_quality_evaluated": false,
+    "distributed_execution_tested": false
+  }
+}
diff --git a/docs/for-verl-users.md b/docs/for-verl-users.md
@@ -1,6 +1,6 @@
 # For verl users
 
-miniVERL is a local runtime for one documented subset of verl v0.8 OPD. It
+miniVERL is a local runtime for two documented subsets of verl v0.8 OPD. It
 keeps familiar field names and Parquet data, then compiles distributed resource
 intent into sequential phases on one CUDA GPU. It is an independent project;
 the mapping is explicit and does not imply endorsement or full compatibility.
@@ -19,16 +19,18 @@ by plans, caches, checkpoints and exports.
 
 - A resolved YAML using the `verl-opd-v0.8-single-gpu-v1` field subset.
 - Reward-free verl-style Parquet prompts with structured chat messages.
-- One actor, one teacher, `n=1`, forward top-k GKD and token-mean aggregation.
+- One actor, one teacher, `n=1`, token-mean aggregation, and either direct
+  forward-top-k GKD or the sampled-k1 vanilla policy-loss profile.
 - Immutable Hugging Face revisions, PEFT adapters and tokenizer snapshots.
 - Familiar fields such as `actor_rollout_ref.model.path`,
   `distillation.teacher_models.teacher_model.model_path`, response bounds,
   learning rate and LoRA configuration.
 
 What is not reusable: arbitrary Hydra composition inside miniVERL, shell launch scripts,
 resource pools, Ray actors, FSDP/Megatron checkpoints, PPO/GRPO, critics,
-policy-gradient OPD, task-reward mixtures, multiple teachers and multimodal
-workers. Unsupported semantics fail closed instead of falling back silently.
+policy-gradient modes beyond the closed sampled-k1 profile, task-reward
+mixtures, multiple teachers and multimodal workers. Unsupported semantics fail
+closed instead of falling back silently.
 
 ## Command mapping
 

diff --git a/docs/release-checklist.md b/docs/release-checklist.md
@@ -7,11 +7,11 @@ after the exact release commit and its remote checks are green.
 ## v0.10.0 development
 
 - [x] Begin from the verified v0.9.1 release and `0.10.0.dev0` canonical state.
-- [ ] Add the versioned internal compatibility-profile registry and profile
+- [x] Add the versioned internal compatibility-profile registry and profile
       identity binding without an unrestricted plugin loader.
-- [ ] Add the pinned verl v0.8 policy-gradient k1 profile with scalar, gradient,
+- [x] Add the pinned verl v0.8 policy-gradient k1 profile with scalar, gradient,
       freshness, resume and export conformance.
-- [ ] Publish bounded RTX 4080 systems evidence for PG k1 and a full SmolLM2
+- [x] Publish bounded RTX 4080 systems evidence for PG k1 and a full SmolLM2
       direct-GKD developer recipe; do not add a task-quality benchmark.
 - [ ] Add validated community hardware records and document the measured or
       explicitly unmeasured Linux/WSL state.