Chore/sync upstream 2026 05 12

.claude/skills

@@ -0,0 +1 @@
+../skills

.cursor/BUGBOT.md

@@ -2,10 +2,16 @@
 
 ## Changelog Enforcement
 
-Any PR that modifies configuration structures or usage patterns must update `CHANGELOG.md`. This includes changes to config fields (added, removed, renamed, moved, or default value changes) in:
+Any PR that introduces **breaking** configuration changes must update `CHANGELOG.md`. Breaking changes are those that require users to update existing configs:
 
-- `src/prime_rl/*/config.py`
-- `src/prime_rl/rl.py`
-- `src/prime_rl/utils/config.py`
+- **Renamed** config fields (old name no longer accepted)
+- **Removed** config fields (field deleted or moved to a different path)
+- **Moved** config fields (field relocated in the config hierarchy)
 
-If such changes are detected without a corresponding `CHANGELOG.md` update, request that the author add an entry.
+Additive changes (new fields with defaults, new optional features) and default value changes do **not** require a changelog entry.
+
+Config files live in:
+
+- `src/prime_rl/configs/`
+
+If breaking changes are detected without a corresponding `CHANGELOG.md` update, request that the author add an entry.

.github/workflows/cpu_tests.yaml

@@ -39,34 +39,58 @@ jobs:
       - name: Cleanup output_dir
         run: rm -rf /tmp/outputs
 
-  integration-tests:
-    name: Integration tests
+  slim-configs-install:
+    name: Slim install (prime-rl-configs only)
     runs-on: ubuntu-latest
     steps:
       - name: Checkout repository
         uses: actions/checkout@v4
-        with:
-          submodules: true
       - name: Install uv
         uses: astral-sh/setup-uv@v5
         with:
           enable-cache: true
           cache-dependency-glob: "uv.lock"
-      - name: Install dependencies
-        run: uv sync --all-extras --locked
-      - name: Run tests
-        env:
-          USERNAME_CI: CI_RUNNER
-          WANDB_API_KEY: ${{ secrets.WANDB_API_KEY }}
-          # Set WANDB_MODE to online only if WANDB_API_KEY is available, otherwise set to offline
-          # This is to allow running tests on forks without WANDB_API_KEY
-          WANDB_MODE: ${{ secrets.WANDB_API_KEY && 'online' || 'offline' }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
-          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
-          PRIME_API_KEY: ${{ secrets.PRIME_API_KEY }}
-          GITHUB_REF_NAME: ${{ github.ref_name }}
-          GITHUB_HEAD_REF: ${{ github.head_ref }}
-          PYTEST_OUTPUT_DIR: /tmp/outputs
-        run: PYTEST_OUTPUT_DIR=/tmp/outputs uv run pytest tests/integration -m "not gpu"
-      - name: Cleanup output_dir
-        run: rm -rf /tmp/outputs
+      - name: Build the slim wheel
+        run: uv build --wheel ./packages/prime-rl-configs --out-dir /tmp/slim-wheel
+      - name: Install slim wheel from a directory with no workspace context
+        # Build the wheel and install it from /tmp so uv can't discover the
+        # repo's root pyproject.toml [tool.uv.sources] and silently apply
+        # them. This mirrors what an external pip-only user gets.
+        run: |
+          uv venv --python 3.12 /tmp/slim
+          cd /tmp && uv pip install --python /tmp/slim/bin/python /tmp/slim-wheel/*.whl
+      - name: Verify configs import without heavy deps
+        run: |
+          /tmp/slim/bin/python - <<'PY'
+          import sys
+
+          from prime_rl.configs.rl import RLConfig
+          from prime_rl.configs.sft import SFTConfig
+          from prime_rl.configs.inference import InferenceConfig
+          from prime_rl.configs.orchestrator import OrchestratorConfig
+          from prime_rl.configs.trainer import TrainerConfig
+          from prime_rl.configs.env_server import EnvServerConfig
+          from prime_rl.utils.config import BaseConfig, find_package_resource, rgetattr, rsetattr
+          from prime_rl.utils.validation import validate_shared_ckpt_config
+
+          forbidden = ["torch", "transformers", "vllm", "wandb", "ring_flash_attn",
+                       "verifiers", "prime", "datasets", "liger_kernel", "loguru"]
+          leaked = [m for m in forbidden if m in sys.modules]
+          if leaked:
+              raise SystemExit(f"slim install leaked heavy deps into sys.modules: {leaked}")
+
+          # Slim wheel does not ship the templates/ data dir; resolver returns None.
+          assert find_package_resource("templates") is None, "templates should not be in slim wheel"
+          print("slim install OK — configs import with no heavy deps loaded")
+          PY
+      - name: Verify a real config (with [slurm]) parses on slim install
+        run: |
+          /tmp/slim/bin/python - <<'PY'
+          from pydantic_config import cli
+          from prime_rl.configs.rl import RLConfig
+          rl = cli(RLConfig, args=["@", "examples/Intellect-3.1/rl.toml"])
+          assert rl.slurm is not None, "expected [slurm] in example config"
+          # On slim, template_path can't be auto-resolved (heavy wheel absent) — must stay None.
+          assert rl.slurm.template_path is None, f"unexpected template_path on slim: {rl.slurm.template_path}"
+          print("slim parse OK")
+          PY

.github/workflows/devx_tag.yaml

@@ -0,0 +1,94 @@
+name: DevX Tag
+
+on:
+  push:
+    branches: [main]
+  workflow_dispatch:
+
+permissions:
+  contents: write
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: false
+
+jobs:
+  create-dev-tag:
+    name: Create .dev tag
+    if: github.ref == 'refs/heads/main'
+    runs-on: ubuntu-latest
+    steps:
+      - name: Create next .dev tag from latest release
+        uses: actions/github-script@v7
+        with:
+          script: |
+            const { owner, repo } = context.repo;
+            const sha = context.sha;
+
+            let latestTag;
+            try {
+              const latestRelease = await github.rest.repos.getLatestRelease({ owner, repo });
+              latestTag = latestRelease.data.tag_name;
+            } catch (error) {
+              if (error.status === 404) {
+                core.notice("No releases found. Skipping .dev tag creation.");
+                return;
+              }
+              throw error;
+            }
+
+            const semverMatch = /^(v?)(\d+)\.(\d+)(?:\.(\d+))?$/.exec(latestTag);
+            if (!semverMatch) {
+              core.setFailed(`Latest release tag '${latestTag}' is not semver-like.`);
+              return;
+            }
+
+            const [, versionPrefix, majorRaw, minorRaw, patchRaw] = semverMatch;
+            const major = Number.parseInt(majorRaw, 10);
+            const minor = Number.parseInt(minorRaw, 10);
+            const patch = Number.parseInt(patchRaw ?? "0", 10) + 1;
+            const baseTag = `${versionPrefix}${major}.${minor}.${patch}`;
+
+            const prefix = `${baseTag}.dev`;
+            const escapeRegex = (value) => value.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+            const pattern = new RegExp(`^refs/tags/${escapeRegex(baseTag)}\\.dev(\\d+)$`);
+
+            for (let attempt = 0; attempt < 5; attempt += 1) {
+              const refs = await github.paginate(github.rest.git.listMatchingRefs, {
+                owner,
+                repo,
+                ref: `tags/${prefix}`,
+              });
+
+              let maxVersion = 0;
+              for (const ref of refs) {
+                const match = pattern.exec(ref.ref);
+                if (match) {
+                  const version = Number.parseInt(match[1], 10);
+                  if (version > maxVersion) {
+                    maxVersion = version;
+                  }
+                }
+              }
+
+              const tag = `${prefix}${maxVersion + 1}`;
+
+              try {
+                await github.rest.git.createRef({
+                  owner,
+                  repo,
+                  ref: `refs/tags/${tag}`,
+                  sha,
+                });
+                core.notice(`Created tag ${tag} at ${sha}`);
+                return;
+              } catch (error) {
+                if (error.status === 422) {
+                  core.warning(`Tag ${tag} already exists, retrying...`);
+                  continue;
+                }
+                throw error;
+              }
+            }
+
+            core.setFailed("Could not create a unique .dev tag after retries.");

.github/workflows/gpu_tests.yaml

@@ -15,7 +15,7 @@ jobs:
     name: Unit tests
     runs-on: vm
     container:
-      image: pytorch/pytorch:2.5.1-cuda12.4-cudnn9-devel
+      image: pytorch/pytorch:2.10.0-cuda13.0-cudnn9-devel
       options: --gpus all
     timeout-minutes: 45
     if: github.event_name == 'push' || github.event.pull_request.draft == false
@@ -53,9 +53,9 @@ jobs:
 
   integration-tests:
     name: Integration tests (${{ matrix.test }})
-    runs-on: vm
+    runs-on: ${{ matrix.runner }}
     container:
-      image: pytorch/pytorch:2.5.1-cuda12.4-cudnn9-devel
+      image: pytorch/pytorch:2.10.0-cuda13.0-cudnn9-devel
       options: --gpus all
     timeout-minutes: 60
     if: github.event_name == 'push' || github.event.pull_request.draft == false
@@ -64,11 +64,20 @@ jobs:
       matrix:
         include:
           - test: base
-            pytest_args: "tests/integration -m gpu --ignore=tests/integration/test_rl_lora.py --ignore=tests/integration/test_rl_multi_run_lora.py"
+            runner: vm
+            pytest_args: "tests/integration -m gpu --ignore=tests/integration/test_rl_lora.py --ignore=tests/integration/test_rl_multi_run_lora.py --ignore=tests/integration/test_benchmark_regression.py --ignore=tests/integration/test_alphabet_sort.py"
           - test: rl_lora
+            runner: vm
             pytest_args: "tests/integration/test_rl_lora.py -m gpu"
+          - test: alphabet_sort
+            runner: vm
+            pytest_args: "tests/integration/test_alphabet_sort.py -m gpu"
           - test: rl_multi_run
+            runner: 4xa6000
             pytest_args: "tests/integration/test_rl_multi_run_lora.py -m gpu"
+          - test: benchmark_regression
+            runner: 4xa6000
+            pytest_args: "tests/integration/test_benchmark_regression.py -m gpu"
     steps:
       - name: Install Git
         run: |
@@ -96,7 +105,10 @@ jobs:
           HF_TOKEN: ${{ secrets.HF_TOKEN }}
           OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
           PRIME_API_KEY: ${{ secrets.PRIME_API_KEY }}
+          PRIME_BASE_URL: ${{ secrets.PRIME_BASE_URL }}
+          PRIME_TEAM_ID: ${{ secrets.PRIME_TEAM_ID }}
           GITHUB_REF_NAME: ${{ github.ref_name }}
           GITHUB_HEAD_REF: ${{ github.head_ref }}
           PYTEST_OUTPUT_DIR: /tmp/outputs
+          VLLM_WORKER_MULTIPROC_METHOD: spawn
         run: uv run pytest ${{ matrix.pytest_args }}

.github/workflows/nightly_tests.yaml

@@ -56,8 +56,11 @@ jobs:
         needs: discover-experiments
         runs-on: research-cluster
         container:
-            image: pytorch/pytorch:2.5.1-cuda12.4-cudnn9-devel
-            options: --gpus all
+            image: pytorch/pytorch:2.10.0-cuda13.0-cudnn9-devel
+            # --shm-size: vLLM V1's DP engine/API-server IPC uses POSIX shm-backed
+            # MessageQueue (~160MB per queue); Docker's default 64MB /dev/shm triggers
+            # SIGBUS in DP engine cores on VLM runs (mm data overflows tmpfs).
+            options: --gpus all --shm-size=16g
         timeout-minutes: 1440 # 24 hours
         strategy:
             fail-fast: false

.github/workflows/style.yaml

@@ -20,3 +20,11 @@ jobs:
           submodules: true
       - name: Run ruff
         uses: astral-sh/ruff-action@v3
+        with:
+          version: "0.13.0"
+          args: "check --config=pyproject.toml"
+      - name: Run ruff format (check)
+        uses: astral-sh/ruff-action@v3
+        with:
+          version: "0.13.0"
+          args: "format --check --config=pyproject.toml"

.github/workflows/sync-docs.yml

@@ -12,6 +12,39 @@ jobs:
   trigger-docs-sync:
     runs-on: ubuntu-latest
     steps:
+      - uses: actions/checkout@v4
+
+      - name: Validate mint.json pages exist
+        run: |
+          missing=()
+          pages=$(python3 -c "
+          import json, sys
+          data = json.load(open('docs/mint.json'))
+          pages = []
+          def collect(obj):
+              if isinstance(obj, list):
+                  for item in obj: collect(item)
+              elif isinstance(obj, dict):
+                  if 'pages' in obj: collect(obj['pages'])
+                  if 'groups' in obj: collect(obj['groups'])
+              elif isinstance(obj, str):
+                  pages.append(obj)
+          collect(data.get('navigation', []))
+          print('\n'.join(pages))
+          ")
+          while IFS= read -r page; do
+            [ -z "$page" ] && continue
+            if [ ! -f "docs/${page}.md" ] && [ ! -f "docs/${page}.mdx" ]; then
+              missing+=("${page}")
+            fi
+          done <<< "$pages"
+          if [ ${#missing[@]} -gt 0 ]; then
+            echo "ERROR: The following pages listed in mint.json do not exist in docs/:"
+            printf '  - %s\n' "${missing[@]}"
+            exit 1
+          fi
+          echo "All mint.json pages verified."
+
       - name: Trigger public-docs workflow
         uses: peter-evans/repository-dispatch@v3
         with:

.gitignore

@@ -5,6 +5,9 @@ configs/endpoints.py
 
 # Artifacts
 **/slurm
+!src/prime_rl/slurm
+!examples/slurm
+!docs/slurm.md
 **/torchrun
 **/weights
 **/checkpoints
@@ -196,3 +199,5 @@ debug_I2_zero_band
 .idea
 
 outputs/
+
+third_party/

.pre-commit-config.yaml

@@ -1,10 +1,10 @@
 repos:
 - repo: https://github.com/astral-sh/ruff-pre-commit
   # Ruff version.
-  rev: v0.9.3
+  rev: v0.13.0
   hooks:
     # Run the linter.
-    - id: ruff
+    - id: ruff-check
       args: [ --fix, --config=pyproject.toml ]
     # Run the formatter.
     - id: ruff-format