Plan a workload for your GPU

Start by inspecting the schedule, then calibrate OPD memory with a short probe. Model size, sequence lengths and the roles resident at each phase determine how much VRAM a run needs. The single-GPU guide explains the settings you can adjust.

Inspect the PPO/GRPO schedule

The maintained qualification device is one RTX 4080. For the new PPO/GRPO workflow, train local.yaml --dry-run --json reports logical prompts, trajectories, actor updates and critic updates before model loading. The output counts planned work; actual peak VRAM comes from executing the recipe.

Hardware coverage

VRAM class	PPO/GRPO planning status	Useful evidence
8 GB	unknown until measured for the exact recipe	doctor, token bounds, actual peak/retries
12 GB	unknown until measured for the exact recipe	same recipe and stack identity
16 GB	maintainer RTX 4080 qualification for named workloads	exact-wheel release record and inspect
24 GB	unknown on other devices until measured	community submissions stay unreviewed

Measurements include the model/revision, algorithm, role strategy, token bounds, package versions and GPU so you can compare them to your setup. Check kernel and dtype support as well as VRAM capacity. Placement rules keep quantized roles resident and PPO's trainable critic separate from a frozen teacher.

OPD calibration

Normal planning is CPU-safe and weight-free. It labels memory as estimated and time as unknown:

miniverl plan --config verl-opd.yaml --out plan.json --offline

After reviewing the config, an explicit probe can calibrate the exact pinned models on one visible CUDA GPU:

miniverl plan --config verl-opd.yaml \
  --accept-local-reinterpretations --out probed-plan.json \
  --probe --offline

The bounded probe loads roles sequentially. It measures actor static memory, greedy 2-token rollout candidates, one selected-position backward, teacher static memory, and one top-k teacher score. It creates no optimizer, performs zero parameter updates, and publishes no checkpoint. Role objects and temporary tensors are destroyed between phases; failure to return allocation near the starting CUDA baseline invalidates the probe.

Results separate measured phase values, recommendations and failed OOM candidates. The cache key binds GPU UUID/name/capability/memory, driver, CUDA runtime, Torch and miniVERL versions, plan digest, model/tokenizer revisions, quantization, LoRA, token bounds and top-k. A mismatched or modified cache is never reused. Use --force-probe to remeasure deliberately.

Use the recommended batches as starting points and retain the configured headroom: the probe's tiny inputs calibrate phases, while a complete run measures the full workload. An OOM stops execution with the model, teacher, context, top-k and loss settings preserved. See limitations for probe scope.