Training only processes one trajectory at a time?

Inizia da src/art/preprocessing/inputs.py alla riga indicata e segui il percorso con cui le traiettorie diventano sequenze di addestramento. Confronta questo percorso con la configurazione di per_device_train_batch_size usando il ciclo di addestramento e i log riportati. Il lavoro è completato quando si determina se elaborare un campione alla volta sia intenzionale e, in caso contrario, si fa funzionare il comportamento di batching configurato oppure se ne documenta la motivazione.

I'm running the training loop with this code, and noticed the trainer only processes one trajectory at a time.

for step in range(TRAINING_STEPS):
    print(f"Step {step+1} rollout")
    train_groups = []
    for scenario in scenarios:
        trajectories = await rollout(model.name, vllm_url, scenario)
        rewards = [traj.reward for traj in trajectories]
        print(f"\n{scenario['id']} rewards: {rewards}")
        train_groups.append(art.TrajectoryGroup(trajectories))

    print(f"Step {step+1} training")
    await model.delete_checkpoints()
    await model.train(train_groups, config=art.TrainConfig(learning_rate=1e-5))

(APIServer pid=1156) Step 19 rollout
(APIServer pid=1156) ....................................................................................................................................................................................
(APIServer pid=1156) s1 rewards: [0.41975308641975306, 0.012345679012345678, 0.1851851851851852, 0.654320987654321, 0.19753086419753085, 0.3950617283950617, 0.1358024691358025, 0.6172839506172839]
(APIServer pid=1156) Step 19 training
(APIServer pid=1156) No "val/reward" metric found in history
(APIServer pid=1156) Deleted checkpoint art/art-rl/models/qwen3-14b-rl/checkpoints/0017
(APIServer pid=1156) Packed 118 trajectories into 118 sequences of length 6144
train: 100%
 118/118 [08:05<00:00,  4.02s/it, loss=0.0146, grad_norm=2.29, policy_loss=0.0146, entropy=0.0266]

This seems to be aligned with the code in inputs.py, which bypasses per_device_train_batch_size config.

Is there a reason to limit to one sample at a time?