artifacts/stage4_cpu_benchmark.json

{
  "stage": 4,
  "model_path": "models/stage4-bert-mini-final",
  "device": "cpu",
  "torch_threads": 8,
  "parameter_count": 11109137,
  "parameter_memory_mb": 44.436548,
  "artifact_size_mb": 45.163283,
  "rss_before_load_mb": 442.351616,
  "rss_after_load_mb": 474.64448,
  "model_load_rss_delta_mb": 32.292864,
  "rss_after_inference_mb": 500.875264,
  "inference_rss_delta_mb": 58.523648,
  "benchmark_examples": 2000,
  "batch_size": 32,
  "elapsed_seconds": 0.48290212500069174,
  "throughput_examples_per_second": 4141.626007541663,
  "mean_latency_ms_per_example_at_batch_size": 0.24145106250034587,
  "scope": "end-to-end fast-tokenizer plus PyTorch CPU forward pass on the 40 wild probes"
}