DaoyuanLi2816 · Aug 29, 2026 · Aug 29, 2026
diff --git a/.github/actionlint.yaml b/.github/actionlint.yaml
@@ -2,3 +2,4 @@ self-hosted-runner:
   labels:
     - cuda
     - rtx4080
+    - wsl2
diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
@@ -45,6 +45,9 @@ jobs:
               "PROJECT_STATE.md",
               "environments/known-good-rtx4080-cu130.json",
               "environments/known-good-rtx4080-cu130.txt",
+              "environments/known-good-rtx4080-wsl2-cu130.json",
+              "environments/known-good-rtx4080-wsl2-cu130.txt",
+              "scripts/run_v011_profile_qualification.py",
               "LICENSE",
               "README.md",
           ):

diff --git a/.github/workflows/gpu.yml b/.github/workflows/gpu.yml
@@ -6,7 +6,7 @@ on:
   workflow_dispatch:
     inputs:
       qualification_level:
-        description: "Release smoke, or all three canonical full workloads"
+        description: "Release smoke, or the complete v0.11 WSL2 qualification"
         type: choice
         required: true
         default: smoke
@@ -62,23 +62,21 @@ jobs:
           retention-days: 90
 
   qualify:
-    name: qualify exact candidate on RTX 4080
+    name: qualify exact candidate on WSL2 RTX 4080
     needs: build-candidate
-    runs-on: [self-hosted, cuda, rtx4080]
+    runs-on: [self-hosted, cuda, rtx4080, wsl2]
     timeout-minutes: 360
     env:
       PYTHONUTF8: "1"
       PYTHONPATH: ""
       PYTHONHOME: ""
-    defaults:
-      run:
-        shell: pwsh
     steps:
       - name: Require a fresh workflow dispatch
         run: |
-          if ($env:GITHUB_RUN_ATTEMPT -ne "1") {
-            throw "GPU qualification cannot use a rerun attempt; start a new workflow_dispatch run."
-          }
+          if [ "$GITHUB_RUN_ATTEMPT" != "1" ]; then
+            echo "::error::GPU qualification cannot use a rerun attempt; start a new workflow_dispatch run."
+            exit 1
+          fi
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           fetch-depth: 0
@@ -89,58 +87,72 @@ jobs:
           path: candidate
       - name: Verify the dedicated measured device
         run: |
-          $name = nvidia-smi --query-gpu=name --format=csv,noheader
-          if (($name | Measure-Object).Count -ne 1) { throw "exactly one visible GPU is required" }
-          if ($name.Trim() -ne "NVIDIA GeForce RTX 4080") { throw "unexpected GPU: $name" }
+          mapfile -t names < <(nvidia-smi --query-gpu=name --format=csv,noheader)
+          if [ "${#names[@]}" -ne 1 ]; then echo "exactly one visible GPU is required"; exit 1; fi
+          if [ "${names[0]}" != "NVIDIA GeForce RTX 4080" ]; then
+            echo "unexpected GPU: ${names[0]}"
+            exit 1
+          fi
+          grep -qi microsoft /proc/sys/kernel/osrelease
           nvidia-smi
-          python --version
+          python3 --version
+          test -x "$HOME/.local/bin/uv"
+          "$HOME/.local/bin/uv" --version
       - name: Validate candidate before installation
         run: |
-          python -m venv .gpu-qualification-venv
-          $python = (Resolve-Path '.gpu-qualification-venv/Scripts/python.exe').Path
-          & $python -m pip install --upgrade pip
-          & $python -m pip install pydantic==2.13.4 PyYAML==6.0.3
-          & $python scripts/build_release_candidate.py `
-            --output candidate `
-            --commit $env:GITHUB_SHA `
-            --repository $env:GITHUB_REPOSITORY `
-            --workflow-path .github/workflows/gpu.yml `
-            --run-id $env:GITHUB_RUN_ID `
-            --run-attempt $env:GITHUB_RUN_ATTEMPT `
+          "$HOME/.local/bin/uv" venv --seed --python 3.12.13 .gpu-qualification-venv
+          qualification_python="$PWD/.gpu-qualification-venv/bin/python"
+          "$qualification_python" -m pip install --upgrade pip
+          "$qualification_python" -m pip install pydantic==2.13.4 PyYAML==6.0.3
+          "$qualification_python" scripts/build_release_candidate.py \
+            --output candidate \
+            --commit "$GITHUB_SHA" \
+            --repository "$GITHUB_REPOSITORY" \
+            --workflow-path .github/workflows/gpu.yml \
+            --run-id "$GITHUB_RUN_ID" \
+            --run-attempt "$GITHUB_RUN_ATTEMPT" \
             --check
-          $manifest = Get-Content -Raw -LiteralPath 'candidate/candidate-manifest.json' | ConvertFrom-Json
-          $wheel = (Resolve-Path (Join-Path 'candidate' $manifest.wheel.filename)).Path
-          "QUALIFICATION_WHEEL=$wheel" | Out-File -FilePath $env:GITHUB_ENV -Append
-          "QUALIFICATION_PYTHON=$python" | Out-File -FilePath $env:GITHUB_ENV -Append
+          "$qualification_python" - <<'PY'
+          import json, os
+          from pathlib import Path
+          manifest = json.loads(Path("candidate/candidate-manifest.json").read_text())
+          wheel = (Path("candidate") / manifest["wheel"]["filename"]).resolve()
+          with Path(os.environ["GITHUB_ENV"]).open("a", encoding="utf-8") as handle:
+              handle.write(f"QUALIFICATION_WHEEL={wheel}\n")
+              handle.write(f"QUALIFICATION_WHEEL_SHA256={manifest['wheel']['sha256']}\n")
+              handle.write(f"QUALIFICATION_PYTHON={Path('.gpu-qualification-venv/bin/python').resolve()}\n")
+          PY
       - name: Install the exact candidate wheel
         run: |
-          & $env:QUALIFICATION_PYTHON -m pip install 'torch==2.13.0+cu130' --index-url https://download.pytorch.org/whl/cu130
-          & $env:QUALIFICATION_PYTHON -m pip install "$env:QUALIFICATION_WHEEL[dev,train,cuda]" --constraint environ
8000
ments/known-good-rtx4080-cu130.txt
-          (Resolve-Path '.gpu-qualification-venv/Scripts').Path | Out-File -FilePath $env:GITHUB_PATH -Append
+          "$QUALIFICATION_PYTHON" -m pip install 'torch==2.13.0+cu130' --index-url https://download.pytorch.org/whl/cu130
+          "$QUALIFICATION_PYTHON" -m pip install \
+            "${QUALIFICATION_WHEEL}[dev,train,cuda,rollout-vllm]" \
+            --constraint environments/known-good-rtx4080-wsl2-cu130.txt
+          echo "$PWD/.gpu-qualification-venv/bin" >> "$GITHUB_PATH"
       - name: Run wheel-installed release smoke
         run: |
-          $offline = @()
-          if ('${{ inputs.offline }}' -eq 'true') { $offline = @('--offline') }
-          New-Item -ItemType Directory -Path qualification | Out-Null
-          Copy-Item -LiteralPath $env:QUALIFICATION_WHEEL -Destination qualification
-          & $env:QUALIFICATION_PYTHON scripts/run_gpu_qualification.py `
-            --commit $env:GITHUB_SHA `
-            --wheel (Join-Path 'qualification' (Split-Path $env:QUALIFICATION_WHEEL -Leaf)) `
-            --candidate-manifest candidate/candidate-manifest.json `
-            --known-good environments/known-good-rtx4080-cu130.json `
-            --output qualification `
-            --work qualification-work `
-            @offline
-          $candidate = Get-Content -Raw candidate/candidate-manifest.json | ConvertFrom-Json
-          $manifestSha = (Get-FileHash -Algorithm SHA256 candidate/candidate-manifest.json).Hash.ToLower()
-          & $env:QUALIFICATION_PYTHON scripts/validate_gpu_qualification.py `
-            qualification/qualification.json `
-            --commit $env:GITHUB_SHA `
-            --wheel-sha256 $candidate.wheel.sha256 `
-            --candidate-manifest-sha256 $manifestSha `
+          offline=()
+          if [ '${{ inputs.offline }}' = 'true' ]; then offline=(--offline); fi
+          mkdir qualification
+          cp -- "$QUALIFICATION_WHEEL" qualification/
+          wheel_copy="qualification/$(basename "$QUALIFICATION_WHEEL")"
+          "$QUALIFICATION_PYTHON" scripts/run_gpu_qualification.py \
+            --commit "$GITHUB_SHA" \
+            --wheel "$wheel_copy" \
+            --candidate-manifest candidate/candidate-manifest.json \
+            --known-good environments/known-good-rtx4080-wsl2-cu130.json \
+            --output qualification \
+            --work qualification-work \
+            "${offline[@]}"
+          manifest_sha=$(sha256sum candidate/candidate-manifest.json | cut -d' ' -f1)
+          "$QUALIFICATION_PYTHON" scripts/validate_gpu_qualification.py \
+            qualification/qualification.json \
+            --commit "$GITHUB_SHA" \
+            --wheel-sha256 "$QUALIFICATION_WHEEL_SHA256" \
+            --candidate-manifest-sha256 "$manifest_sha" \
             --required-gpu-name 'NVIDIA GeForce RTX 4080'
       - name: Run GPU regression tests
-        run: '& $env:QUALIFICATION_PYTHON -m pytest -q -m gpu -ra'
+        run: '"$QUALIFICATION_PYTHON" -m pytest -q -m gpu -ra'
       - name: Upload exact-commit release smoke
         uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
         with:
@@ -151,19 +163,42 @@ jobs:
       - name: Run full canonical qualification workloads
         if: ${{ inputs.qualification_level == 'full' }}
         run: |
-          $offline = @()
-          if ('${{ inputs.offline }}' -eq 'true') { $offline = @('--offline') }
-          & $env:QUALIFICATION_PYTHON -m pip install 'hydra-core>=1.3,<2'
-          & $env:QUALIFICATION_PYTHON -m pip install --no-deps 'git+https://github.com/verl-project/verl.git@7aed6b230776f963fa09509c10d9c3a767d1102c'
-          New-Item -ItemType Directory -Path qualification-full | Out-Null
-          & $env:QUALIFICATION_PYTHON scripts/run_verl_opd_reference_workload.py --out qualification-full-work/direct --result qualification-full/direct.json @offline
-          & $env:QUALIFICATION_PYTHON scripts/run_verl_pg_k1_workload.py --out qualification-full-work/pg-k1 --result qualification-full/pg-k1.json @offline
-          & $env:QUALIFICATION_PYTHON scripts/run_smollm2_opd_reference_workload.py --out qualification-full-work/smollm2 --result qualification-full/smollm2.json @offline
-          & $env:QUALIFICATION_PYTHON scripts/promote_full_gpu_qualification.py `
-            --qualification qualification/qualification.json `
-            --direct qualification-full/direct.json `
-            --pg-k1 qualification-full/pg-k1.json `
-            --smollm2 qualification-full/smollm2.json
+          offline=()
+          if [ '${{ inputs.offline }}' = 'true' ]; then offline=(--offline); fi
+          "$QUALIFICATION_PYTHON" -m pip install 'hydra-core>=1.3,<2'
+          "$QUALIFICATION_PYTHON" -m pip install --no-deps 'git+https://github.com/verl-project/verl.git@7aed6b230776f963fa09509c10d9c3a767d1102c'
+          mkdir qualification-full
+          "$QUALIFICATION_PYTHON" scripts/run_verl_opd_reference_workload.py --out qualification-full-work/direct --result qualification-full/direct.json "${offline[@]}"
+          "$QUALIFICATION_PYTHON" scripts/run_verl_pg_k1_workload.py --out qualification-full-work/pg-k1 --result qualification-full/pg-k1.json "${offline[@]}"
+          "$QUALIFICATION_PYTHON" scripts/run_smollm2_opd_reference_workload.py --out qualification-full-work/smollm2 --result qualification-full/smollm2.json "${offline[@]}"
+          "$QUALIFICATION_PYTHON" scripts/run_v011_profile_qualification.py \
+            --commit "$GITHUB_SHA" \
+            --wheel "$QUALIFICATION_WHEEL" \
+            --wheel-sha256 "$QUALIFICATION_WHEEL_SHA256" \
+            --output qualification-full/v011-profiles.json \
+            --work qualification-full-work/v011-profiles \
+            "${offline[@]}"
+          "$QUALIFICATION_PYTHON" scripts/benchmark_rollout_runtime_v2_selected.py \
+            --backend hf_cached \
+            --wheel "$QUALIFICATION_WHEEL" \
+            --expected-commit "$GITHUB_SHA" \
+            --output qualification-full/hf-cached-runtime.json \
+            --repetitions 3
+          "$QUALIFICATION_PYTHON" scripts/benchmark_rollout_runtime_v2_selected.py \
+            --backend vllm \
+            --wheel "$QUALIFICATION_WHEEL" \
+            --expected-commit "$GITHUB_SHA" \
+            --output qualification-full/vllm-runtime.json \
+            --repetitions 3
+          "$QUALIFICATION_PYTHON" scripts/promote_full_gpu_qualification.py \
+            --qualification qualification/qualification.json \
+            --direct qualification-full/direct.json \
+            --pg-k1 qualification-full/pg-k1.json \
+            --smollm2 qualification-full/smollm2.json \
+            --v011-profiles qualification-full/v011-profiles.json \
+            --hf-cached-runtime qualification-full/hf-cached-runtime.json \
+            --vllm-runtime qualification-full/vllm-runtime.json \
+            --hf-reference benchmarks/results/rollout-runtime-v2-hf-reference.json
          - name: Upload full qualification records
         uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
         if: ${{ success() && inputs.qualification_level == 'full' }}
@@ -175,11 +210,19 @@ jobs:
       - name: Remove runner-local environments and model work products
         if: always()
         run: |
-          foreach ($relative in @('.gpu-qualification-venv', 'qualification-work', 'qualification-full-work')) {
-            if (Test-Path -LiteralPath $relative) {
-              $target = (Resolve-Path -LiteralPath $relative).Path
-              $workspace = (Resolve-Path -LiteralPath $env:GITHUB_WORKSPACE).Path
-              if (-not $target.StartsWith($workspace, [System.StringComparison]::OrdinalIgnoreCase)) { throw "refusing cleanup outside workspace: $target" }
-              Remove-Item -LiteralPath $target -Recurse -Force
-            }
-          }
+          python3 - <<'PY'
+          import os, shutil
+          from pathlib import Path
+          workspace = Path(os.environ["GITHUB_WORKSPACE"]).resolve()
+          for relative in (
+              ".gpu-qualification-venv",
+              "qualification-work",
+              "qualification-full-work",
+              "qualification-full",
+          ):
+              target = (workspace / relative).resolve()
+              if target.parent != workspace:
+                  raise SystemExit(f"refusing cleanup outside workspace: {target}")
+              if target.exists():
+                  shutil.rmtree(target)
+          PY
diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
@@ -34,7 +34,7 @@ jobs:
             --qualification-artifact-name gpu-full-qualification \
             --required-level full_qualification \
             --required-gpu-name "NVIDIA GeForce RTX 4080" \
-            --known-good environments/known-good-rtx4080-cu130.json \
+            --known-good environments/known-good-rtx4080-wsl2-cu130.json \
             --output verified-qualified-candidate
       - name: Preserve the verified byte chain
         uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1

diff --git a/docs/release-qualification.md b/docs/release-qualification.md
@@ -10,7 +10,7 @@ workflow is manual, not continuous GPU CI and not a pull-request required check.
 | level | cadence | executed scope |
 | --- | --- | --- |
 | release smoke | diagnostic use | install the hosted-runner candidate wheel, verify import/CLI origin, run CLI doctor/plan/dry-run, pinned Qwen actor and teacher, one rollout/score/update, PEFT export/reload and CUDA teardown |
-| full qualification | every formal release | release smoke plus the unchanged direct-GKD, sampled-k1 and SmolLM2 canonical workloads, including interruption/resume and their existing export/materialization checks |
+| full qualification | every formal release | release smoke, the three canonical resume workloads, and the v0.11 direct n=1, grouped PG n=4, rewarded PG n=4, `hf_cached` and vLLM runtime gates |
 
 The smoke budget is intentionally small and is never substituted for a frozen
 full workload or a scientific benchmark. Both levels record runtime
@@ -26,17 +26,20 @@ from the same successful run and first attempt.
 GitHub-hosted `build-candidate` job creates exactly one wheel and one sdist with
 the pinned build toolchain, validates them with Twine, and uploads
 `candidate-distributions` with canonical checksums and a strict manifest. The
-dependent `[self-hosted, cuda, rtx4080]` job downloads that same-run artifact,
+dependent `[self-hosted, cuda, rtx4080, wsl2]` job downloads that same-run artifact,
 validates it before installation, then installs the candidate wheel in a fresh
 virtual environment using the [known-good stack](single-gpu-guide.md). It does
 not build a distribution. `qualification.json` binds the candidate manifest,
 full source SHA, exact wheel hash, workflow run, profile identity, model
 revisions, inputs, measured environment and output hashes. It also proves that
 both the imported package and `miniverl` executable came from the qualification
 environment rather than a checkout or user site.
-For a full run, the workflow additionally validates each canonical result's
-exact resume, resource, PEFT and scale-out fields, then promotes that same
-record to `full_qualification`; three loose result files cannot claim the
+For a v0.11 full run, the workflow additionally performs three real one-update
+profiles, all 24 preregistered cells on each rollout backend, eight policy
+refreshes per backend, conformance and teardown. Promotion verifies the exact
+source, wheel and environment bindings; speed and memory thresholds; grouped
+trajectory identity; reward/advantage completion; and vLLM's
+policy-gradient fail-closed boundary. Loose result files cannot claim the
 higher level.
 
 Every dispatch is single-use. Both jobs reject `GITHUB_RUN_ATTEMPT` values
@@ -54,7 +57,8 @@ SHA-256. Cross-origin artifact redirects retain ordinary API headers but strip
 authorization, proxy authorization and cookies. It publishes the accepted
 wheel and sdist without rebuilding them. Future release runs retain the full
 qualification record, four principal workload JSON files and a deterministic
-subordinate-evidence archive in the canonical layout below. A committed
+subordinate-evidence archive. For v0.11, that archive also contains the three
+profile/backend qualification records. A committed
 JSON file, manual upload, fork run, different workflow or cross-run artifact
 pair cannot satisfy this gate.
 
@@ -80,9 +84,10 @@ qualification-evidence-manifest.json
 ```
 
 The four principal workload records remain directly inspectable. Adapter
-configuration and safetensors, its miniVERL manifest, input prompts and the run
-summary live once in the deterministic archive; its manifest binds every member
-to its semantic role, byte count and SHA-256. `SHA256SUMS` covers only the wheel
+configuration and safetensors, its miniVERL manifest, input prompts, run
+summary and version-specific qualification records live once in the
+deterministic archive; its manifest binds every member to its semantic role,
+byte count and SHA-256. `SHA256SUMS` covers only the wheel
 and sdist. `qualification-SHA256SUMS` covers the nine provenance and evidence
 assets in canonical order. Hashes prove byte integrity, not code signing or
 third-party endorsement.
@@ -127,6 +132,10 @@ portable bounded artifacts, and checks cleanup targets remain under
 `GITHUB_WORKSPACE`. Model caches remain runner-local and are not uploaded.
 Rotate the runner token after suspected exposure.
 
+The v0.11 release runner is Ubuntu under WSL2 with Python 3.12.13 and the
+checked `known-good-rtx4080-wsl2-cu130` stack. The older Windows qualification
+record remains historical and cannot authorize v0.11.
+
 ## Measured release state
 
 The first exact-commit full run completed for v0.10.1 on attempt 1:

diff --git a/docs/single-gpu-guide.md b/docs/single-gpu-guide.md
@@ -37,13 +37,16 @@ not select a CUDA-enabled PyTorch build on its own.
    For the exact stack measured by the maintainer on one RTX 4080, use the
 machine-readable
 [`environments/known-good-rtx4080-cu130.json`](https://github.com/DaoyuanLi2816/mini-verl/blob/main/environments/known-good-rtx4080-cu130.json)
+is the historical Windows v0.10.1 record. The active v0.11 release
+qualification uses
+[`environments/known-good-rtx4080-wsl2-cu130.json`](https://github.com/DaoyuanLi2816/mini-verl/blob/main/environments/known-good-rtx4080-wsl2-cu130.json)
 and its constraints:
 
 ```bash
 python -m pip install "torch==2.13.0+cu130" \
   --index-url https://download.pytorch.org/whl/cu130
-python -m pip install "miniverl[train,cuda]" \
-  --constraint environments/known-good-rtx4080-cu130.txt
+python -m pip install "miniverl[train,cuda,rollout-vllm]" \
+  --constraint environments/known-good-rtx4080-wsl2-cu130.txt
 python scripts/check_known_good_environment.py
 ```
-Original file line number
+Diff line change
@@ Expand Up / @@ -2,3 +2,4 @@ self-hosted-runner: @@
       labels:
         - cuda
         - rtx4080
+        - wsl2