- 01:32 prepared a Qwen3-4B plain EGGROLL GSM8K reproduction using the user-specified setting N=126, rank=1, sigma=1e-3, lr=1, T=1, 1024-token generations, 6 generations per prompt, noise_reuse=4, 100 epochs, and 268x5 validation samples; added microbatched rollouts to avoid a single large generation batch
- 01:52 submitted 20260810-015007-qwen3_4b_gsm8k_plain_eggroll_repro_jean_zay-14b5b16 (job 804069 on jean-zay, commit 14b5b16)
- 01:53 submitted Qwen3-4B plain EGGROLL GSM8K reproduction 20260810-015007-qwen3_4b_gsm8k_plain_eggroll_repro_jean_zay-14b5b16 on Jean Zay as job 804069 using the user-specified large plain-EGGROLL setting
- 05:54 fetched 20260810-015007-qwen3_4b_gsm8k_plain_eggroll_repro_jean_zay-14b5b16
- 05:54 fetched Qwen3-4B plain EGGROLL repro 20260810-015007-qwen3_4b_gsm8k_plain_eggroll_repro_jean_zay-14b5b16 after Jean Zay job 804069 failed; slurm err shows EggRoll.do_updates OOMed during the dense einsum gradient allocation before metrics were written
- 08:03 fixed the Qwen3 H100 EGGROLL update OOM by streaming low-rank ES outer-product accumulation instead of materialising an N-by-d_out-by-d_in einsum intermediate; applied the same update-memory fix to decomposed EGGROLL
- 17:06 submitted 20260810-170449-qwen3_4b_gsm8k_plain_eggroll_repro_jean_zay-818feaf (job 822981 on jean-zay, commit 818feaf)
- 17:07 submitted Qwen3-4B plain EGGROLL GSM8K rerun 20260810-170449-qwen3_4b_gsm8k_plain_eggroll_repro_jean_zay-818feaf on Jean Zay as job 822981 after the H100 update-memory fix; scheduled a two-hour cron check until the run is terminal
- 19:07 checked Qwen3-4B plain EGGROLL rerun 20260810-170449-qwen3_4b_gsm8k_plain_eggroll_repro_jean_zay-818feaf; Jean Zay job 822981 is still PENDING, so the two-hour cron remains enabled
- 21:07 checked Qwen3-4B plain EGGROLL rerun 20260810-170449-qwen3_4b_gsm8k_plain_eggroll_repro_jean_zay-818feaf; Jean Zay job 822981 is still PENDING, so the two-hour cron remains enabled
- 23:07 checked Qwen3-4B plain EGGROLL rerun 20260810-170449-qwen3_4b_gsm8k_plain_eggroll_repro_jean_zay-818feaf; Jean Zay job 822981 is RUNNING, so the two-hour cron remains enabled