DaoyuanLi2816 · DaoyuanLi2816 · Aug 12, 2026 · Aug 12, 2026 · Aug 12, 2026 · Aug 12, 2026
diff --git a/PROJECT_STATE.md b/PROJECT_STATE.md
@@ -64,6 +64,26 @@ response or implementation fails closed. The prompt-data integration exercises
 rollout, exact-state scoring, cache reload, one token-mean optimizer update and
 the emitted verl diagnostics.
 
+## v0.8.0 single-GPU verl OPD pivot — PR D
+
+`miniverl plan` now compiles a path or the packaged
+`builtin:qwen3-0.6b-1.7b-opd` profile without importing torch or allocating
+weights. Estimated memory, disk and placement stay distinct from measured and
+unknown values. Auto placement uses parsed model-scale metadata and configured
+headroom, never a GPU-name allowlist; unknown sizes conservatively select swap.
+`miniverl run` validates a generated native `RunConfig`, preserves the source,
+field report and local plan in the run, and executes the supported prompt OPD
+path. `miniverl data sample` makes reward-free verl-style message Parquet from
+an installed wheel.
+
+One RTX 4080 runtime-conformance run at `d441c45` used pinned Qwen3-0.6B /
+Qwen3-1.7B NF4 snapshots, two prompts, 16 response tokens each, top-k 32 and
+one optimizer update. Peak reserved memory was 3.1758 GiB; first rollout,
+teacher target batch and update completed at 10.8684 s, 11.0630 s and 12.0224 s
+from construction start. The standard PEFT adapter load passed. This is a
+systems result only; no alignment endpoint or method comparison ran. The
+machine-readable record is `benchmarks/results/rtx4080-verl-opd-runtime-v1.json`.
+
 ## v0.7.1 Product correction — RELEASE CANDIDATE
 
 Branch `v0.7.1-product-correction` starts from synchronized main

diff --git a/benchmarks/results/rtx4080-verl-opd-runtime-v1.json b/benchmarks/results/rtx4080-verl-opd-runtime-v1.json
@@ -0,0 +1,70 @@
+{
+  "schema_version": 1,
+  "kind": "single_gpu_runtime_conformance",
+  "status": "measured",
+  "measured_at": "2026-08-12T07:34:15Z",
+  "miniverl_version": "0.8.0.dev0",
+  "source_commit": "d441c458ab5e30a939352e64391b6f875b489c99",
+  "profile": "verl-opd-v0.8-single-gpu-v1",
+  "verl": {
+    "tag": "v0.8.0",
+    "commit": "7aed6b230776f963fa09509c10d9c3a767d1102c",
+    "distributed_execution_tested": false
+  },
+  "hardware": {
+    "gpu": "NVIDIA GeForce RTX 4080",
+    "gpu_count": 1,
+    "vram_gib": 15.992,
+    "driver": "596.49",
+    "torch": "2.13.0+cu130"
+  },
+  "models": {
+    "student": {
+      "id": "Qwen/Qwen3-0.6B",
+      "revision": "c1899de289a04d12100db370d81485cdf75e47ca",
+      "quantization": "nf4",
+      "adapter": "lora-r8-alpha16"
+    },
+    "teacher": {
+      "id": "Qwen/Qwen3-1.7B",
+      "revision": "70d244cc86ccca08cf5af4e1e306ecf908b1ad5e",
+      "quantization": "nf4"
+    }
+  },
+  "recipe": {
+    "runtime_strategy": "dual_model_resident",
+    "prompt_limit": 128,
+    "response_limit": 16,
+    "logical_batch": 2,
+    "rollout_physical_batch": 2,
+    "update_physical_batch": 1,
+    "top_k": 32,
+    "optimizer_updates": 1,
+    "compiled_plan_sha256": "a4e7a377e313022766c80098cadd67577743905c9c2a96862696b5afeeabfa8c",
+    "input_parquet_sha256": "a686aaa602a5a6cd3411de137190779e692806ddedd03f4d728e808539800c54"
+  },
+  "measurements": {
+    "peak_allocated_gib": 2.1713,
+    "peak_reserved_gib": 3.1758,
+    "time_to_first_rollout_seconds": 10.8684,
+    "time_to_first_teacher_target_batch_seconds": 11.063,
+    "time_to_first_optimizer_update_seconds": 12.0224,
+    "rollout_tokens_per_second": 23.39,
+    "teacher_scored_positions_per_second": 164.44,
+    "update_selected_positions_per_second": 41.2463,
+    "checkpoint_bytes": 27649321,
+    "run_disk_bytes": 27757228,
+    "resume_load_seconds": 9.4933
+  },
+  "artifacts": {
+    "checkpoint_adapter_sha256": "8c20e2136593c7a37bd084df6324bdd8e937561e63af6c8eec05e5b543589667",
+    "standard_peft_adapter_sha256": "595046de169790870bc87d6a04f043ebab0ed94bee2ab7bc37db8cf30665ef58",
+    "standard_peft_manifest_sha256": "e6d0e7b0ad8b39221775ad0655c3fbb355baa0d659ca3c89927e46030890353c",
+    "standard_peft_load_verified": true
+  },
+  "scientific_scope": {
+    "runtime_correctness_only": true,
+    "alignment_quality_evaluated": false,
+    "opd_beats_sft_dpo_or_kd_claimed": false
+  }
+}
diff --git a/docs/opd-quickstart.md b/docs/opd-quickstart.md
@@ -0,0 +1,47 @@
+# Run verl-style OPD locally
+
+miniVERL v0.8 supports one pinned, fail-closed subset of verl v0.8: one actor,
+one teacher, one generation per prompt, reward-free direct GKD with
+`forward_kl_topk`, token-mean aggregation and a LoRA/QLoRA student. It is local
+single-GPU execution, not Ray/FSDP or distributed verl execution.
+
+Install a CUDA build of PyTorch that matches your machine first, then:
+
+```bash
+python -m pip install "miniverl[train,cuda,bridge]"
+miniverl data sample --format verl-parquet --out data/opd-smoke.parquet
+miniverl plan --profile verl-opd-v0.8-single-gpu-v1 \
+  --config builtin:qwen3-0.6b-1.7b-opd
+miniverl run --profile verl-opd-v0.8-single-gpu-v1 \
+  --config builtin:qwen3-0.6b-1.7b-opd
+```
+
+The built-in recipe downloads pinned Qwen3-0.6B and Qwen3-1.7B snapshots when
+they are not cached. Allow roughly 6 GiB of download/cache space and 0.25 GiB
+for run artifacts. The plan command itself is CPU-only and weight-free; use
+`--offline` for a zero-network compiler smoke.
+
+## Measured reference
+
+| 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 |
+| 12 GiB CUDA GPU | same built-in recipe | planner-selected | same | — | — | not measured |
+| 24 GiB CUDA GPU | same built-in recipe | planner-selected | same | — | — | not measured |
+
+The measured run completed one current-policy rollout/scoring/update cycle,
+exported a loadable PEFT adapter, and used one RTX 4080. It demonstrates
+runtime and artifact correctness only; it did not evaluate alignment quality.
+The checksummed record is
+[`rtx4080-verl-opd-runtime-v1.json`](https://github.com/DaoyuanLi2816/mini-verl/blob/main/benchmarks/results/rtx4080-verl-opd-runtime-v1.json).
+
+## What `plan` means
+
+`plan` reports estimates separately from measurements. Auto placement uses
+model metadata plus the configured VRAM headroom; it never branches on a GPU
+product name. Unknown model sizes conservatively select swap. `--probe` is
+reserved but fails closed in v0.8.0 rather than loading weights unexpectedly.
+
+Unsupported settings—including policy-gradient OPD, task rewards, reference
+KL, multiple teachers, multiple generations, multimodal inputs and every
+distributed dimension—are rejected rather than reinterpreted.
diff --git a/examples/verl-opd-v0.8-single-gpu.yaml b/examples/verl-opd-v0.8-single-gpu.yaml
@@ -1,5 +1,6 @@
 # Resolved fixture for the pinned verl-opd-v0.8-single-gpu-v1 compiler.
-# This file compiles offline. Runtime execution is added in later v0.8 PRs.
+# This file compiles offline and can be executed through `miniverl run` once
+# the referenced Parquet files exist.
 data:
   train_files: [data/train.parquet]
   val_files: [data/val.parquet]
@@ -76,6 +77,8 @@ trainer:
   n_gpus_per_node: 1
   nnodes: 1
 miniverl:
+  student_revision: c1899de289a04d12100db370d81485cdf75e47ca
+  teacher_revision: 70d244cc86ccca08cf5af4e1e306ecf908b1ad5e
   runtime:
     mode: auto
   memory:

diff --git a/mkdocs.yml b/mkdocs.yml
@@ -39,6 +39,7 @@ plugins:
   - search
 nav:
   - Home: index.md
+  - Run verl-style OPD: opd-quickstart.md
   - Start on one GPU: single-gpu-guide.md
   - Align:
       - "External Alignment Gate v1: early stop": alignment-external/alignment-external-v1.md

diff --git a/pyproject.toml b/pyproject.toml
@@ -127,6 +127,7 @@ core-metadata-version = "2.4"
 "benchmarks/schema/alignment-external-result.schema.json" = "miniverl/evidence/data/alignment-external-v1/result.schema.json"
 "benchmarks/preregistration/alignment-external-v1.yaml" = "miniverl/evidence/data/alignment-external-v1/preregistration.yaml"
 "benchmarks/evidence/alignment-external-v1/jsonnav-selection-records.jsonl" = "miniverl/evidence/data/alignment-external-v1/task-evidence.jsonl"
+"src/miniverl/resources/qwen3_0_6b_1_7b_opd.yaml" = "miniverl/resources/qwen3_0_6b_1_7b_opd.yaml"
 
 [tool.hatch.build.targets.sdist]
 core-metadata-version = "2.4"