miniverl 0.6.2

PyPI · Stable docs · Development docs · 中文

miniVERL is a local, inspectable runtime for a documented subset of single-GPU LLM alignment and distillation. It keeps rollout provenance, assistant-only loss masks, teacher targets, update budgets and run artifacts explicit, then exports portable artifacts through a fail-closed bridge to one pinned upstream verl profile.

PyPI v0.6.1 is stable; main is development. The CUDA path has no GPU-name allowlist, but fit depends on the model pair, context budget, kernels and VRAM. miniVERL is independent from verl and does not claim distributed execution or full algorithmic compatibility.

Install and run the 60-second demo

python -m pip install "miniverl[train]"
miniverl doctor
miniverl demo --output runs/demo
miniverl inspect runs/demo

The demo is deterministic, needs no network or GPU, and performs a real toy optimization in about 50 seconds on the measured laptop CPU. For inspection, schemas and reports without the ML stack, use pip install miniverl. For CUDA training, install the matching CUDA-enabled PyTorch wheel first, then install miniverl[train,cuda]; the extra does not select a CUDA PyTorch build. See the single-GPU guide.

Three paths

The bridge is a verified artifact bridge: a pinned config/data/model parse-load smoke at miniVERL-defined compatibility Level 3. It has never run a distributed verl job, and no OPD-to-PPO semantic parity is claimed.

The import is deliberately not generic YAML conversion. If dataset or environment, teacher identity, objective, or schedule semantics are missing, import-verl writes <stem>.import-report.json and a non-executable <stem>.template.yaml with status: needs_user_input. It never silently substitutes calculator tasks or an unqualified same-base teacher. An unresolved ${...} value can never reach an accepted recipe, outputs are stem-specific, and an existing output family is replaced only with an explicit --overwrite.

One measured alignment result

Alignment Lab v1 is a saturated tool-policy case study, not a broad safety benchmark. The shared SFT checkpoint already achieved 100% policy compliance and 100% retained tool utility in all three seeds. No continuation method improved it; continued SFT and both OPD variants retained measured regressions.

The two sandbox safety checks tied at zero while utility still regressed. IFEval, XSTest, HarmBench and RewardBench were not executed. “Preference win rate” is a deterministic Minipolicy paired outcome, not human preference. Read the study, seed-level values and limitations.

One measured systems result

On one RTX 4080 with Qwen3-0.6B and eight fixed SQLite trajectories, physical batch 4 improved update throughput from 2.369 to 3.866 trajectories/s in the dual-model runtime. The shared-backbone batch-4 cell used 2.227 GiB peak reserved memory versus 3.035 GiB for dual model, while running 10.1% slower. All 12 preregistered equivalence comparisons passed. These are one-workload, one-machine measurements, not promises for other GPUs.

Consumer Runtime v1 methods and caveats

Compatibility boundary

The bridge targets official verl v0.8.0 at commit 7aed6b23 and uses the term miniVERL-defined compatibility Level 3. That means a checksummed standard-artifact bundle plus pinned upstream config-parse/model-data-load smoke—not arbitrary verl YAML or a completed distributed job.

Current exported bundles are intentionally launchable: false: the base snapshot is absent, the reward implementation fails closed, and required user mappings remain placeholders. The generated entry point is therefore launch.template.sh. Readiness is reported as separate facts for artifact completeness, parse/load smoke, reward completeness, launchability, distributed execution and algorithm-semantic parity. The target is a PPO/reward scaffold, not an executable continuation of miniVERL OPD semantics.

Detailed studies and preserved negative evidence

• RecoveryBench v1: frozen-student KD outperformed much slower fresh-state OPD on the preregistered primary view; the verifier gate remained insufficient_evidence.
• Alignment Lab v1: the starting SFT checkpoint was at the ceiling, so no positive OPD result is claimed.
• Calculator benchmark: both negative controls completed normally and measured 0% strict success. They were not configuration failures. Because they used the historical ambiguous protocol-v1 prompt, their failure cannot be attributed solely to intrinsic teacher behavior.
• Consumer Runtime v1: padded update batches and shared adapters preserve the measured one-update objective within declared tolerances; rollout generation remains sequential.
• Limitations, math, reproducibility and compatibility policy.

New runs establish tokenizer compatibility through structural identity. The legacy behavioral fingerprint—token IDs for one fixed probe plus metadata—is only a migration fallback for older artifacts and is not an identity proof.

Scope

miniVERL supports one local CUDA process. It does not implement or wrap Ray, FSDP, Megatron, PPO, GRPO or a distributed launcher. The public studies cover small Qwen3 models, deterministic tool environments and one RTX 4080; they do not establish cross-model, cross-task, cross-GPU or broad safety generality.

git clone https://github.com/DaoyuanLi2816/mini-verl.git
cd mini-verl
python -m pip install -e ".[dev]"
pytest -q -m "not gpu and not network"

Apache-2.0 licensed. See CONTRIBUTING.md and SECURITY.md. Project records: default GPU recipe, frozen calculator JSON, changelog, citation and license.