outputs/space_training.log

Starting pilot training at Thu May 14 08:57:54 2026
Command: /usr/local/bin/python3.10 training/train_grpo.py --model-name Qwen/Qwen3-1.7B --max-steps 100 --num-generations 8 --gradient-accumulation-steps 8 --max-completion-length 512 --learning-rate 1e-05 --logging-steps 5 --output-dir /data/outputs/releaseops-grpo-pilot --metrics-json /data/outputs/grpo_env_metrics_pilot.json --best-loss-dir outputs/best_by_loss --bf16

Warning: installed TRL supports environment_factory but GRPOConfig does not expose env_kwargs_keys; dataset scenario columns may not be passed into reset(**kwargs).

Generating train split: 0 examples [00:00, ? examples/s]
Generating train split: 120 examples [00:00, 6533.69 examples/s]
Skipping GRPOConfig args unsupported by installed TRL: max_prompt_length
Traceback (most recent call last):
  File "/app/training/train_grpo.py", line 877, in <module>
    main()
  File "/app/training/train_grpo.py", line 822, in main
    training_args = build_config(args)
  File "/app/training/train_grpo.py", line 558, in build_config
    return GRPOConfig(**config_kwargs)
  File "<string>", line 179, in __init__
  File "/usr/local/lib/python3.10/site-packages/trl/trainer/grpo_config.py", line 879, in __post_init__
    super().__post_init__()
  File "/usr/local/lib/python3.10/site-packages/trl/trainer/base_config.py", line 107, in __post_init__
    super().__post_init__()
  File "/usr/local/lib/python3.10/site-packages/transformers/training_args.py", line 1576, in __post_init__
    self._validate_args()
  File "/usr/local/lib/python3.10/site-packages/transformers/training_args.py", line 1738, in _validate_args
    raise ValueError(error_message)
ValueError: Your setup doesn't support bf16/gpu. You need to assign use_cpu if you want to train the model on CPU.