miniVERL¶
verl for a single consumer GPU.
Train with PPO, GRPO or on-policy distillation on one NVIDIA GPU. Bring your verl configs, follow the training, resume from a checkpoint and take the resulting models into your next workflow.
Start training Choose a workflow
Start with PPO¶
Install the matching CUDA-enabled build from the PyTorch installer, then:
python -m pip install "miniverl[train,hydra]"
miniverl data sample --reward-profile target-length --rows 8 --out data/rl-prompts.parquet
miniverl run --example hydra-ppo --bind reward.provider=target_length --dry-run
miniverl run --example hydra-ppo --bind reward.provider=target_length --run-id local-ppo
This Qwen3-0.6B example uses a simple length reward you can inspect in
rewards.jsonl. The dry run checks the configuration before downloading models.
Continue with inspect → resume → export.
Three ways to use miniVERL¶
Train with rewards¶
Run PPO or GRPO using your verl config and local data.
Artifact: training recipe and configuration report.
Next: Bring your own config
Distill locally¶
Train a student on its own rollouts with feedback from a teacher.
Artifact: a memory and execution plan for the student/teacher pair.
Prepare a larger run¶
Export your model, data and config for handoff to verl.
Artifact: PEFT, Parquet and config files with a setup checklist.
Next: Review the bundle
Make the most of your GPU¶
miniVERL schedules training roles in phases so actor, critic, reference, reward and teacher can share one GPU. Choose a model and placement strategy with the single-GPU guide and hardware planner. Native recipes also offer SFT, DPO and offline KD.
Explore the results¶
- OPD on an RTX 4080: update timing, memory and exact recovery.
- Rollout Runtime v2: 24 measurements across rollout settings.
- Alignment Lab: a saturated tool-policy case study.
- RecoveryBench: tool recovery outcomes.
- External Alignment Gate: frozen external evaluation results.
About and support¶
miniVERL is an independent project for single-GPU NVIDIA CUDA training. The compatibility matrix lists supported verl versions, profiles and handoff stages. Limitations collects hardware, scientific and security scope. See comparisons for help choosing a training workflow.