pilot.jsonl: Qwen early keep/stop process.formal/: Qwen/DeepSeek × microLLM/PyTorch × three fresh processes at1×512.fallback128.jsonl: one Qwen process on the unchanged short-row implementation.profile/: retained Qwen Kernel and HIP API aggregates.comparison.json: model throughput, peak and fallback gates.profile-summary.json: forward/backward row-Kernel and whole-process deltas.
All official rows use BF16 Linear compute, FP32 master weights, one warm-up and two measured steps. Raw JSONL/CSV is authoritative over rounded labels.