- 00:00 checked fixed GRPO first-final rerun 20260903-194915-qwen3_4b_gsm8k_first_final_grpo_jean_zay-d66aeb1; Slurm state FAILED and fetched logs show CUDA OOM during GRPO completion logprob/backward through Qwen3 lm_head, so it produced no usable result records