artifacts/stage4_cpu_benchmark.json

{
  "stage": 4,
  "model_path": "models/stage4-bert-tiny-ner-improved",
  "device": "cpu",
  "torch_threads": 8,
  "parameter_count": 4371601,
  "parameter_memory_mb": 17.486404,
  "artifact_size_mb": 18.209406,
  "rss_before_load_mb": 442.400768,
  "rss_after_load_mb": 474.218496,
  "model_load_rss_delta_mb": 31.817728,
  "rss_after_inference_mb": 483.885056,
  "inference_rss_delta_mb": 41.484288,
  "benchmark_examples": 2000,
  "batch_size": 32,
  "elapsed_seconds": 0.1722599580007227,
  "throughput_examples_per_second": 11610.359268702534,
  "mean_latency_ms_per_example_at_batch_size": 0.08612997900036135,
  "scope": "end-to-end fast-tokenizer plus PyTorch CPU forward pass on the 40 wild probes"
}