host-batch/: Qwen/DeepSeek × B1/2/4/8 × microLLM/PyTorch on physical visibility 1, explicit host argmax control, one process.batch/: the same matrix and GPU with device row-wise argmax.transfer-control.jsonl: same-window Qwen B8 direct host/device rows with engine transfer calls and bytes.profile-host/andprofile-device/: same Qwen B8 rocprof aggregates.comparison.jsonandprofile-summary.json: machine-checked keep contracts.
Both modes use the same full-sequence model path, tokens, precision, warm-up and measured
iterations. Only batch-argmax-mode changes.