DaoyuanLi2816 · DaoyuanLi2816 · Aug 13, 2026 · Aug 13, 2026 · Aug 13, 2026 · Aug 13, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -6,6 +6,22 @@ All notable changes to miniVERL are recorded here. The format follows
 
 ## [Unreleased]
 
+### Measured developer workload
+
+- Published a checksummed RTX 4080 systems workload over 32 distinct consumed
+  prompts, 64-token responses and eight QLoRA updates. The data-bound figure
+  reports steady-state phase time, labelled throughput and 3.1914 GiB peak
+  reserved VRAM without a task-quality or method-comparison claim.
+- Fixed Parquet resume to reconstruct the saved row/epoch cursor. A real
+  Qwen3 interruption after update four now reproduces uninterrupted
+  trajectories, adapter and optimizer tensors byte for byte; all training
+  state fields match apart from the intentionally run-specific resolved-config
+  digest.
+- Qualified the Apache-2.0 SmolLM2-360M/1.7B pair with a pinned one-update
+  compatibility smoke covering tokenizer identity, rollout, teacher scoring,
+  actor update and PEFT reload. This is not a second full recipe or quality
+  benchmark.
+
 ### Immutable plan/run workflow
 
 - Added deterministic `plan --out plan.json` artifacts that bind the source

diff --git a/PROJECT_STATE.md b/PROJECT_STATE.md
@@ -54,6 +54,12 @@ adapter, software and output provenance. `launchable: true` means only that the
 pinned upstream entry point has complete local inputs; distributed execution
 remains `false` and algorithm parity remains unclaimed.
 
+A second-family compatibility smoke used pinned Apache-2.0 SmolLM2-360M and
+SmolLM2-1.7B snapshots. Structural tokenizer identity matched; one strict
+forward-top-k rollout, teacher-score, update and PEFT reload completed on the
+RTX 4080 at 1.416 GiB peak reserved VRAM. This is a tiny compatibility check,
+not a second supported recipe or task-quality result.
+
 ## v0.8.1 product surface
 
 The landing pages now lead with the documented one-GPU verl-style OPD journey,

diff --git a/PYPI.md b/PYPI.md
@@ -144,17 +144,20 @@ arbitrary launch script.
 
 ## Measured RTX 4080 path
 
-The packaged Qwen3-0.6B/1.7B smoke completed two 16-token rollouts and one OPD
-update on one RTX 4080 with **3.1758 GiB peak reserved VRAM**. The first update
-completed in **12.0224 seconds**, the final adapter was exported as standard
-PEFT/safetensors, and a clean reload passed. The recipe records immutable model
-revisions, phase timings, cache identity, checkpoint bytes and adapter hashes.
+The Qwen3-0.6B/1.7B developer workload consumed **32 distinct prompts**, each
+with a 64-token response bound, and completed **8 current-policy updates** at
+**3.1914 GiB peak reserved VRAM**. Median steady-state rollout, teacher-scoring
+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.
 
 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
 or KD. Other NVIDIA GPUs use the same device-name-agnostic CUDA path, but model
-fit depends on VRAM, context length, quantization and installed kernels. Read
-the [exact smoke record and limitations](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/opd-quickstart.md).
+fit depends on VRAM, context length, quantization and installed kernels.
 
 ### Choose a path by hardware, not GPU branding
 

diff --git a/README.md b/README.md
@@ -144,17 +144,20 @@ arbitrary launch script.
 
 ## Measured RTX 4080 path
 
-The packaged Qwen3-0.6B/1.7B smoke completed two 16-token rollouts and one OPD
-update on one RTX 4080 with **3.1758 GiB peak reserved VRAM**. The first update
-completed in **12.0224 seconds**, the final adapter was exported as standard
-PEFT/safetensors, and a clean reload passed. The recipe records immutable model
-revisions, phase timings, cache identity, checkpoint bytes and adapter hashes.
+The Qwen3-0.6B/1.7B developer workload consumed **32 distinct prompts**, each
+with a 64-token response bound, and completed **8 current-policy updates** at
+**3.1914 GiB peak reserved VRAM**. Median steady-state rollout, teacher-scoring
+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.
 
 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
 or KD. Other NVIDIA GPUs use the same device-name-agnostic CUDA path, but model
-fit depends on VRAM, context length, quantization and installed kernels. Read
-the [exact smoke record and limitations](docs/opd-quickstart.md).
+fit depends on VRAM, context length, quantization and installed kernels.
 
 ### Choose a path by hardware, not GPU branding
 

diff --git a/README.zh-CN.md b/README.zh-CN.md
@@ -120,14 +120,18 @@ Megatron、多 GPU 与多节点均不支持。已知不支持值会得到机器
 
 ## RTX 4080 实测路径
 
-打包的 Qwen3-0.6B/1.7B smoke 在一张 RTX 4080 上完成两条 16-token rollout 与一次
-OPD update，**peak reserved VRAM 为 3.1758 GiB**，首次 update 在 **12.0224 秒**完成；
-标准 PEFT/safetensors adapter 导出与干净重载均通过。recipe 同时记录不可变模型 revision、
-阶段计时、cache 身份、checkpoint 字节数与 adapter 哈希。
+Qwen3-0.6B/1.7B developer workload 实际消费 **32 个不同 prompt**，response 上限为
+64 token，完成 **8 次 current-policy update**，**peak reserved VRAM 为 3.1914 GiB**。
+稳态 rollout、teacher scoring 与 update 的中位耗时分别为 9.7200、0.4864 与 2.3260 秒。
+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。
 
 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
-能否装下仍取决于显存、上下文、量化与 kernel。见[精确实测记录](docs/opd-quickstart.md)。
+能否装下仍取决于显存、上下文、量化与 kernel。
 
 ### 按硬件条件选择路径，而不是按显卡名称
 

diff --git a/benchmarks/results/rtx4080-smollm2-opd-family-smoke-v1.json b/benchmarks/results/rtx4080-smollm2-opd-family-smoke-v1.json
@@ -0,0 +1,83 @@
+{
+  "schema_version": 1,
+  "kind": "single_gpu_opd_model_family_compatibility_smoke",
+  "status": "measured",
+  "measured_at": "2026-08-13T07:48:14Z",
+  "source_commit": "2819f7290ffb0c14a0ff5d13509a0fec29c8100d",
+  "miniverl_version": "0.9.0.dev0",
+  "profile": "verl-opd-v0.8-single-gpu-v1",
+  "models": {
+    "student": {
+      "id": "HuggingFaceTB/SmolLM2-360M-Instruct",
+      "revision": "a10cc1512eabd3dde888204e902eca88bddb4951",
+      "license": "apache-2.0",
+      "quantization": "nf4",
+      "adapter": "lora-r8-alpha16"
+    },
+    "teacher": {
+      "id": "HuggingFaceTB/SmolLM2-1.7B-Instruct",
+      "revision": "31b70e2e869a7173562077fd711b654946d38674",
+      "license": "apache-2.0",
+      "quantization": "nf4"
+    }
+  },
+  "tokenizer_identity": {
+    "tokenizer_class": "GPT2Tokenizer",
+    "length": 49152,
+    "structural_digest_v2": "4bf67fa8fdbf333c2f79c00bdc43909ed2ffb4bd7ed1f44a11d65f171ea9d1c5",
+    "behavioral_fingerprint_v1": "5772ce854a388a7613d1284a34fa0360f9c603256796cb1f72e25713ab82f7b5",
+    "student_teacher_match": true
+  },
+  "workload": {
+    "distinct_prompts": 1,
+    "response_limit": 16,
+    "logical_batch": 1,
+    "optimizer_updates": 1,
+    "top_k": 32,
+    "loss": "forward_kl_topk",
+    "runtime_strategy": "dual_model_resident"
+  },
+  "measurements": {
+    "construction_seconds": 8.5539,
+    "time_to_first_rollout_seconds": 10.084,
+    "time_to_first_teacher_targets_seconds": 10.1989,
+    "time_to_first_update_seconds": 10.7449,
+    "rollout_tokens_per_second": 10.46,
+    "teacher_scored_positions_per_second": 139.24,
+    "update_positions_per_second": 35.5365,
+    "selected_positions": 16,
+    "peak_allocated_gib": 1.384,
+    "peak_reserved_gib": 1.416,
+    "oom_retries": 0,
+    "checkpoint_bytes": 19798324,
+    "run_disk_bytes": 26507663
+  },
+  "artifacts": {
+    "adapter_sha256": "1eda290c7e19b55d9d497ec014711fa451c67595525ad5aa304167fd1be189a6",
+    "peft_adapter_model_sha256": "90bf81a33c6c23acb87c93a1df53caeb5e9b91aa72e5678edfe9832b60716aec",
+    "peft_load_verified": true
+  },
+  "runtime": {
+    "gpu": "NVIDIA GeForce RTX 4080",
+    "gpu_count": 1,
+    "vram_gib": 15.992,
+    "torch": "2.13.0+cu130",
+    "cuda_runtime": "13.0",
+    "transformers": "5.14.1",
+    "peft": "0.18.0",
+    "bitsandbytes": "0.50.0",
+    "distributed_execution_tested": false
+  },
+  "scope": {
+    "config_compiled": true,
+    "tokenizer_identity_verified": true,
+    "student_loaded": true,
+    "teacher_loaded": true,
+    "rollout_completed": true,
+    "teacher_scoring_completed": true,
+    "optimizer_update_completed": true,
+    "task_quality_evaluated": false,
+    "alignment_quality_evaluated": false,
+    "full_recipe_supported": false
+  }
+}
diff --git a/benchmarks/results/rtx4080-verl-opd-developer-v1.json b/benchmarks/results/rtx4080-verl-opd-developer-v1.json
@@ -0,0 +1,122 @@
+{
+  "artifacts": {
+    "checkpoint_hashes": {
+      "adapter.safetensors": "60e2eced54a3a13d5d66c765b1a9f45d26b2f16bf982fa0011f9316b0cc0433f",
+      "optimizer.safetensors": "0bb7dde497204151b952d1e5866059822334f71150bfe4a5ce46e253cc84f494",
+      "state.json": "e92abdf28f18f07c2ebb7c49a712b806c9462ad422bfeb99fb4cce037a248962"
+    },
+    "standard_peft_adapter_sha256": "085b9c1241acaf0a2ab2c6c9081484b77eafbd329d92c7a6181ddfa3709bfe0c",
+    "standard_peft_load_verified": true,
+    "trajectory_sha256": "f0d7e256e977508d698c4db506f1c37f42188b91b27c372bce1dad8bcf6d54d0",
+    "workload_output_bytes": 130233537
+  },
+  "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_opd_developer_workload",
+  "measured_at": "2026-08-13T07:31:03Z",
+  "measurements": {
+    "adapter_bytes": 9213148,
+    "batch_downshifts": {
+      "rollout_oom": 0,
+      "update_chunk_oom": 0
+    },
+    "cache_bytes": 90727,
+    "checkpoint_bytes": 27649609,
+    "cold_startup_seconds": 9.0802,
+    "distinct_prompts_consumed": 32,
+    "observed_rollout_physical_batch_sizes": [
+      4
+    ],
+    "peak_allocated_gib": 2.3033,
+    "peak_reserved_gib": 3.1914,
+    "prompts_consumed": 32,
+    "steady_state_median_seconds": {
+      "actor_update": 2.326,
+      "rollout": 9.72,
+      "teacher_scoring": 0.4864
+    },
+    "steady_state_median_throughput": {
+      "rollout_tokens_per_second": 26.3374,
+      "teacher_scored_positions_per_second": 526.32,
+      "update_positions_per_second": 110.0602
+    },
+    "time_to_first_rollout_seconds": 19.0015,
+    "time_to_first_teacher_targets_seconds": 19.5584,
+    "time_to_first_update_seconds": 22.0412,
+    "total_run_bytes": 65065194
+  },
+  "miniverl_version": "0.9.0.dev0",
+  "models": {
+    "student": {
+      "adapter": "lora-r8-alpha16",
+      "id": "Qwen/Qwen3-0.6B",
+      "quantization": "nf4",
+      "revision": "c1899de289a04d12100db370d81485cdf75e47ca"
+    },
+    "teacher": {
+      "id": "Qwen/Qwen3-1.7B",
+      "quantization": "nf4",
+      "revision": "70d244cc86ccca08cf5af4e1e306ecf908b1ad5e"
+    }
+  },
+  "profile": "verl-opd-v0.8-single-gpu-v1",
+  "recipe": {
+    "compiled_plan_sha256": "9b2b0b522585e5a88603f83f3ce714080a1560d92940d9d5dd661fe586dc1a5b",
+    "dataset_rows": 64,
+    "distinct_dataset_prompts": 64,
+    "distinct_prompts_consumed": 32,
+    "input_parquet_sha256": "baa99ba8c59d4c30b2e3081346860e7e8b2665dd70fdbcd7952935545a0cd770",
+    "logical_batch": 4,
+    "optimizer_updates": 8,
+    "prompt_limit": 128,
+    "prompts_consumed": 32,
+    "response_limit": 64,
+    "rollout_physical_batch": 4,
+    "runtime_strategy": "dual_model_resident",
+    "teacher_score_batch": 4,
+    "top_k": 32,
+    "update_physical_batch": 1
+  },
+  "resource_contract": {
+    "gpu_hour_limit": 4.0,
+    "peak_reserved_limit_gib": 14.5,
+    "peak_reserved_within_limit": true,
+    "total_gpu_workload_seconds": 235.49
+  },
+  "resume": {
+    "adapter_and_optimizer_byte_identical": true,
+    "excluded_run_identity_field": "resolved_config_digest",
+    "global_optimizer_steps": 8,
+    "interrupt_after_optimizer_updates": 4,
+    "interrupt_checkpoint_bytes": 27649609,
+    "resume_load_seconds": 8.2339,
+    "status": "exact_match",
+    "task_cursor": 32,
+    "training_state_fields_identical": true,
+    "trajectories_byte_identical": true
+  },
+  "schema_version": 1,
+  "scientific_scope": {
+    "alignment_quality_evaluated": false,
+    "opd_beats_sft_dpo_or_kd_claimed": false,
+    "runtime_correctness_only": true,
+    "task_quality_evaluated": false
+  },
+  "source_commit": "2819f7290ffb0c14a0ff5d13509a0fec29c8100d",
+  "status": "measured",
+  "verl": {
+    "commit": "7aed6b230776f963fa09509c10d9c3a767d1102c",
+    "distributed_execution_tested": false,
+    "tag": "v0.8.0"
+  },
+  "workload_driver_sha256": "00356532e8d1ea6a488340bffc42acf59300e9c130d483866645233e5cc823af"
+}
diff --git a/docs/opd-quickstart.md b/docs/opd-quickstart.md
@@ -23,6 +23,10 @@ for run artifacts. The plan command itself is CPU-only and weight-free; use
 
 ## Measured reference
 
+The v0.9 developer workload extends this install smoke to 32 distinct consumed
+prompts, 64 response tokens and eight optimizer updates. See its
+[phase-time, throughput, VRAM and resume evidence](verl-opd-reference-workload.md).
+
 | GPU | student / teacher | strategy | limits / top-k | peak reserved | first update | status |
 | --- | --- | --- | --- | ---: | ---: | --- |
 | RTX 4080 16 GiB | Qwen3-0.6B / Qwen3-1.7B, both NF4 | dual resident | 128 + 16 tokens / 32 | 3.176 GiB | 12.02 s | measured |

diff --git a/docs/verl-opd-reference-workload-mobile.svg b/docs/verl-opd-reference-workload-mobile.svg