raw.jsonl contains one FP32 full-logit reference and three plan-cache candidate processes
per model. summary.json compares medians with Experiment 034 and the fixed Experiment 031
PyTorch BF16 reference.
python3 benchmarks/single_gpu/run_bf16_attention_models.py \
--manifest /path/to/local-manifest.json \
--binary build/hip-release/apps/microllm_hf_infer \
--baseline-summary docs/optimization-log/experiments/034-data/summary.json \
--baseline-kind attention \
--pytorch-summary docs/optimization-log/experiments/031-data/summary.json \
--raw-output docs/optimization-log/experiments/036-data/raw.jsonl \
--summary-output docs/optimization-log/experiments/036-data/summary.json \
--runs 3