Six candidate rows cover two models and three independent processes. The retained FP32, BF16 and PyTorch baselines come from Experiment 037.
python3 benchmarks/single_gpu/run_bf16_training_candidate.py \
--manifest /path/to/local-manifest.json \
--binary build/hip-release/apps/microllm_hf_train_step \
--baseline-summary docs/optimization-log/experiments/037-data/summary.json \
--raw-output docs/optimization-log/experiments/039-data/raw.jsonl \
--summary-output docs/optimization-log/experiments/039-data/summary.json \
--runs 3The candidate source was removed after the keep gate failed.