[2026-07-15 00:25:12,786] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect) INFO 07-15 00:25:15 __init__.py:183] Automatically detected platform cuda. 2026-07-15 00:25:15,797 | INFO | CUDA preflight passed with training GPU=NVIDIA H100 80GB HBM3 and vLLM GPU=NVIDIA H100 80GB HBM3 2026-07-15 00:25:15,973 | INFO | No checkpoint selected; starting a new training run 2026-07-15 00:25:15,974 | INFO | Starting h100-full profile with model=Qwen/Qwen2.5-3B-Instruct, dataset_size=50000, max_steps=450, max_completion_length=1024 2026-07-15 00:25:15,974 | INFO | Loading dataset Jiayi-Pan/Countdown-Tasks-3to4 2026-07-15 00:25:17,130 | INFO | Selected 50,000 examples 2026-07-15 00:25:17,131 | INFO | Loading tokenizer Qwen/Qwen2.5-3B-Instruct 2026-07-15 00:25:17,622 | INFO | Formatting prompts 2026-07-15 00:25:17,724 | INFO | Prepared 45,000 training examples and 5,000 test examples 2026-07-15 00:25:17,727 | INFO | Writing generated completions in outputs/qwen-countdown-h100-hillclimb/completions.jsonl You are attempting to use Flash Attention 2.0 with a model not initialized on GPU. Make sure to move the model to GPU after initializing it on CPU with `model.to('cuda')`. Loading checkpoint shards: 0%| | 0/2 [00:00