{
 "measured_on": "2026-11-05",
 "host": {
  "name": "a shared 32-core Intel Xeon 2.80 GHz Linux CPU worker, one torch thread per process, load average 14 to 19"
 },
 "versions": {
  "python": "3.10",
  "torch": "2.13.0 (CPU build)",
  "transformers": "5.18.0",
  "torchlens": "2.38.1 (development branch, 2026-11-05)"
 },
 "harness": "standalone timing script (rerun_bench): per variant 3 warmups then up to 15 samples within 20 s (GPT-2: 1 warmup, up to 5 samples within 60 s); median per process; GPT-2 small built from GPT2Config() with random weights",
 "correctness": {
  "staged_fast_rerun_vs_hooked_edit_max_abs_diff": {
   "gpt2-small": 0.0,
   "resnet18": 0.0,
   "lstm-tagger": 1.86e-08
  },
  "unedited_fast_rerun_saved_site_vs_fresh_trace_max_abs_diff": {
   "gpt2-small": 0.0,
   "resnet18": 0.0,
   "lstm-tagger": 0.0
  }
 },
 "engines": {
  "trace.run unedited": "rerun",
  "trace.run staged edit": "guarded_fast",
  "run_transaction fast=True": "guarded_fast"
 },
 "more_saves_second_worker": {
  "lstm-tagger": [
   {
    "plain_forward": 0.74,
    "trace_one_site": 29.99,
    "run_transaction_new_inputs_same_saves": 60.71,
    "run_transaction_new_inputs_more_saves_by_address": 58.84
   },
   {
    "plain_forward": 0.73,
    "trace_one_site": 31.18,
    "run_transaction_new_inputs_same_saves": 60.09,
    "run_transaction_new_inputs_more_saves_by_address": 58.65
   }
  ],
  "resnet18": [
   {
    "plain_forward": 10.61,
    "trace_one_site": 294.02,
    "run_transaction_new_inputs_same_saves": 692.01,
    "run_transaction_new_inputs_more_saves_by_address": 637.45
   },
   {
    "plain_forward": 10.73,
    "trace_one_site": 291.9,
    "run_transaction_new_inputs_same_saves": 719.65,
    "run_transaction_new_inputs_more_saves_by_address": 731.43
   }
  ]
 },
 "rerun": {
  "resnet18": {
   "plain_forward": {
    "process_medians_ms": [
     13.302,
     14.259,
     14.203
    ],
    "job": "rerun"
   },
   "forward_hook_one_site": {
    "process_medians_ms": [
     14.202,
     14.563,
     18.525
    ],
    "job": "hook"
   },
   "trace_one_site": {
    "process_medians_ms": [
     395.311,
     408.932,
     405.281
    ],
    "job": "rerun"
   },
   "trace_more_sites": {
    "process_medians_ms": [
     402.448,
     417.936,
     665.181
    ],
    "job": "rerun"
   },
   "run_new_inputs_same_saves": {
    "process_medians_ms": [
     622.499,
     684.487,
     693.682
    ],
    "job": "rerun"
   },
   "run_transaction_new_inputs_same_saves": {
    "process_medians_ms": [
     849.769,
     898.791,
     855.946
    ],
    "job": "rerun"
   },
   "run_transaction_fast_unstaged": {
    "process_medians_ms": [
     18.186,
     19.183,
     18.85
    ],
    "job": "rerun"
   },
   "trace_with_staged_edit": {
    "process_medians_ms": [
     444.02,
     521.988,
     452.971
    ],
    "job": "rerun"
   },
   "run_staged_edit_new_inputs": {
    "process_medians_ms": [
     48.541,
     56.453,
     49.665
    ],
    "job": "rerun"
   },
   "run_transaction_fast_staged_edit": {
    "process_medians_ms": [
     19.365,
     22.056,
     19.349
    ],
    "job": "rerun"
   },
   "first_trace_in_process_ms": [
    1710,
    1922,
    2380
   ]
  },
  "gpt2-small": {
   "plain_forward": {
    "process_medians_ms": [
     1628.389,
     1645.837
    ],
    "job": "rerun"
   },
   "forward_hook_one_site": {
    "process_medians_ms": [
     1345.331,
     1331.539
    ],
    "job": "hook"
   },
   "trace_one_site": {
    "process_medians_ms": [
     3074.091,
     2913.988
    ],
    "job": "rerun"
   },
   "trace_more_sites": {
    "process_medians_ms": [
     4214.117,
     4336.59
    ],
    "job": "rerun"
   },
   "run_new_inputs_same_saves": {
    "process_medians_ms": [
     4768.862,
     4650.859
    ],
    "job": "rerun"
   },
   "run_transaction_new_inputs_same_saves": {
    "process_medians_ms": [
     7710.558,
     8009.907
    ],
    "job": "rerun"
   },
   "run_transaction_fast_unstaged": {
    "process_medians_ms": [
     1326.296,
     1343.557
    ],
    "job": "rerun"
   },
   "trace_with_staged_edit": {
    "process_medians_ms": [
     3052.274,
     3291.442
    ],
    "job": "rerun"
   },
   "run_staged_edit_new_inputs": {
    "process_medians_ms": [
     1447.357,
     1378.131
    ],
    "job": "rerun"
   },
   "run_transaction_fast_staged_edit": {
    "process_medians_ms": [
     1279.876,
     1365.769
    ],
    "job": "rerun"
   },
   "first_trace_in_process_ms": [
    4886,
    5465
   ]
  },
  "lstm-tagger": {
   "plain_forward": {
    "process_medians_ms": [
     1.026,
     1.464,
     1.078
    ],
    "job": "rerun"
   },
   "forward_hook_one_site": {
    "process_medians_ms": [
     1.11,
     1.081,
     1.129
    ],
    "job": "hook"
   },
   "trace_one_site": {
    "process_medians_ms": [
     43.161,
     42.243,
     44.638
    ],
    "job": "rerun"
   },
   "trace_more_sites": {
    "process_medians_ms": [
     43.949,
     45.283,
     43.035
    ],
    "job": "rerun"
   },
   "run_new_inputs_same_saves": {
    "process_medians_ms": [
     66.791,
     66.084,
     65.837
    ],
    "job": "rerun"
   },
   "run_transaction_new_inputs_same_saves": {
    "process_medians_ms": [
     82.665,
     82.979,
     80.717
    ],
    "job": "rerun"
   },
   "run_transaction_fast_unstaged": {
    "process_medians_ms": [
     3.234,
     3.107,
     3.087
    ],
    "job": "rerun"
   },
   "trace_with_staged_edit": {
    "process_medians_ms": [
     53.411,
     51.34,
     47.439
    ],
    "job": "rerun"
   },
   "run_staged_edit_new_inputs": {
    "process_medians_ms": [
     14.644,
     8.888,
     8.736
    ],
    "job": "rerun"
   },
   "run_transaction_fast_staged_edit": {
    "process_medians_ms": [
     3.746,
     3.301,
     3.323
    ],
    "job": "rerun"
   },
   "first_trace_in_process_ms": [
    2704,
    1397,
    1341
   ]
  }
 },
 "page_rows": [
  {
   "section": "rerun_resnet18",
   "text": "| ResNet-18, CPU, one 64x64 image, n=3 processes | plain forward, nothing attached | 14.20 ms (13.30 ms to 14.26 ms) |"
  },
  {
   "section": "rerun_resnet18",
   "text": "| ResNet-18, CPU, one 64x64 image, n=3 processes | forward hook on the saved module (second job; its plain forward: 14.39 ms) | 14.56 ms (14.20 ms to 18.52 ms) |"
  },
  {
   "section": "rerun_resnet18",
   "text": "| ResNet-18, CPU, one 64x64 image, n=3 processes | fresh `tl.trace(model, x, save=tl.module(site))` | 405.28 ms (395.31 ms to 408.93 ms) |"
  },
  {
   "section": "rerun_resnet18",
   "text": "| ResNet-18, CPU, one 64x64 image, n=3 processes | fresh `tl.trace` saving every listed module | 417.94 ms (402.45 ms to 665.18 ms) |"
  },
  {
   "section": "rerun_resnet18",
   "text": "| ResNet-18, CPU, one 64x64 image, n=3 processes | `trace.run(model, new_x)` on a finished trace (earlier build: recaptures) | 684.49 ms (622.50 ms to 693.68 ms) |"
  },
  {
   "section": "rerun_resnet18",
   "text": "| ResNet-18, CPU, one 64x64 image, n=3 processes | `trace.run_transaction(inputs=new_x)` | 855.95 ms (849.77 ms to 898.79 ms) |"
  },
  {
   "section": "rerun_resnet18",
   "text": "| ResNet-18, CPU, one 64x64 image, n=3 processes | `trace.run_transaction(inputs=new_x, fast=True)`, no edit staged | 18.85 ms (18.19 ms to 19.18 ms) |"
  },
  {
   "section": "rerun_resnet18",
   "text": "| ResNet-18, CPU, one 64x64 image, n=3 processes | fresh `tl.trace(..., intervene=tl.when(...))` | 452.97 ms (444.02 ms to 521.99 ms) |"
  },
  {
   "section": "rerun_resnet18",
   "text": "| ResNet-18, CPU, one 64x64 image, n=3 processes | `trace.run(model, new_x)` on a trace with a staged edit | 49.66 ms (48.54 ms to 56.45 ms) |"
  },
  {
   "section": "rerun_resnet18",
   "text": "| ResNet-18, CPU, one 64x64 image, n=3 processes | `trace.run_transaction(inputs=new_x, fast=True)` with a staged edit | 19.36 ms (19.35 ms to 22.06 ms) |"
  },
  {
   "section": "rerun_gpt2-small",
   "text": "| GPT-2 small (random weights), CPU, one forward of 16 sequences of 32 tokens, n=2 processes | plain forward, nothing attached | 1.64 s (1.63 s to 1.65 s) |"
  },
  {
   "section": "rerun_gpt2-small",
   "text": "| GPT-2 small (random weights), CPU, one forward of 16 sequences of 32 tokens, n=2 processes | forward hook on the saved module (second job; its plain forward: 1.63 s) | 1.34 s (1.33 s to 1.35 s) |"
  },
  {
   "section": "rerun_gpt2-small",
   "text": "| GPT-2 small (random weights), CPU, one forward of 16 sequences of 32 tokens, n=2 processes | fresh `tl.trace(model, x, save=tl.module(site))` | 2.99 s (2.91 s to 3.07 s) |"
  },
  {
   "section": "rerun_gpt2-small",
   "text": "| GPT-2 small (random weights), CPU, one forward of 16 sequences of 32 tokens, n=2 processes | fresh `tl.trace` saving every listed module | 4.28 s (4.21 s to 4.34 s) |"
  },
  {
   "section": "rerun_gpt2-small",
   "text": "| GPT-2 small (random weights), CPU, one forward of 16 sequences of 32 tokens, n=2 processes | `trace.run(model, new_x)` on a finished trace (earlier build: recaptures) | 4.71 s (4.65 s to 4.77 s) |"
  },
  {
   "section": "rerun_gpt2-small",
   "text": "| GPT-2 small (random weights), CPU, one forward of 16 sequences of 32 tokens, n=2 processes | `trace.run_transaction(inputs=new_x)` | 7.86 s (7.71 s to 8.01 s) |"
  },
  {
   "section": "rerun_gpt2-small",
   "text": "| GPT-2 small (random weights), CPU, one forward of 16 sequences of 32 tokens, n=2 processes | `trace.run_transaction(inputs=new_x, fast=True)`, no edit staged | 1.33 s (1.33 s to 1.34 s) |"
  },
  {
   "section": "rerun_gpt2-small",
   "text": "| GPT-2 small (random weights), CPU, one forward of 16 sequences of 32 tokens, n=2 processes | fresh `tl.trace(..., intervene=tl.when(...))` | 3.17 s (3.05 s to 3.29 s) |"
  },
  {
   "section": "rerun_gpt2-small",
   "text": "| GPT-2 small (random weights), CPU, one forward of 16 sequences of 32 tokens, n=2 processes | `trace.run(model, new_x)` on a trace with a staged edit | 1.41 s (1.38 s to 1.45 s) |"
  },
  {
   "section": "rerun_gpt2-small",
   "text": "| GPT-2 small (random weights), CPU, one forward of 16 sequences of 32 tokens, n=2 processes | `trace.run_transaction(inputs=new_x, fast=True)` with a staged edit | 1.32 s (1.28 s to 1.37 s) |"
  },
  {
   "section": "rerun_lstm-tagger",
   "text": "| LSTM tagger, CPU, one sequence of 32 tokens, n=3 processes | plain forward, nothing attached | 1.08 ms (1.03 ms to 1.46 ms) |"
  },
  {
   "section": "rerun_lstm-tagger",
   "text": "| LSTM tagger, CPU, one sequence of 32 tokens, n=3 processes | forward hook on the saved module (second job; its plain forward: 1.07 ms) | 1.11 ms (1.08 ms to 1.13 ms) |"
  },
  {
   "section": "rerun_lstm-tagger",
   "text": "| LSTM tagger, CPU, one sequence of 32 tokens, n=3 processes | fresh `tl.trace(model, x, save=tl.module(site))` | 43.16 ms (42.24 ms to 44.64 ms) |"
  },
  {
   "section": "rerun_lstm-tagger",
   "text": "| LSTM tagger, CPU, one sequence of 32 tokens, n=3 processes | fresh `tl.trace` saving every listed module | 43.95 ms (43.04 ms to 45.28 ms) |"
  },
  {
   "section": "rerun_lstm-tagger",
   "text": "| LSTM tagger, CPU, one sequence of 32 tokens, n=3 processes | `trace.run(model, new_x)` on a finished trace (earlier build: recaptures) | 66.08 ms (65.84 ms to 66.79 ms) |"
  },
  {
   "section": "rerun_lstm-tagger",
   "text": "| LSTM tagger, CPU, one sequence of 32 tokens, n=3 processes | `trace.run_transaction(inputs=new_x)` | 82.67 ms (80.72 ms to 82.98 ms) |"
  },
  {
   "section": "rerun_lstm-tagger",
   "text": "| LSTM tagger, CPU, one sequence of 32 tokens, n=3 processes | `trace.run_transaction(inputs=new_x, fast=True)`, no edit staged | 3.11 ms (3.09 ms to 3.23 ms) |"
  },
  {
   "section": "rerun_lstm-tagger",
   "text": "| LSTM tagger, CPU, one sequence of 32 tokens, n=3 processes | fresh `tl.trace(..., intervene=tl.when(...))` | 51.34 ms (47.44 ms to 53.41 ms) |"
  },
  {
   "section": "rerun_lstm-tagger",
   "text": "| LSTM tagger, CPU, one sequence of 32 tokens, n=3 processes | `trace.run(model, new_x)` on a trace with a staged edit | 8.89 ms (8.74 ms to 14.64 ms) |"
  },
  {
   "section": "rerun_lstm-tagger",
   "text": "| LSTM tagger, CPU, one sequence of 32 tokens, n=3 processes | `trace.run_transaction(inputs=new_x, fast=True)` with a staged edit | 3.32 ms (3.30 ms to 3.75 ms) |"
  },
  {
   "section": "until_resnet18",
   "text": "| ResNet-18, CPU, one 64x64 image, one process | plain forward, nothing attached | 13.65 ms |"
  },
  {
   "section": "until_resnet18",
   "text": "| ResNet-18, CPU, one 64x64 image, one process | `trace.run_transaction(inputs=new_x, fast=True)`, whole forward | 20.30 ms |"
  },
  {
   "section": "until_resnet18",
   "text": "| ResNet-18, CPU, one 64x64 image, one process | the same with `until=\"layer1\"` | 6.16 ms |"
  },
  {
   "section": "until_resnet18",
   "text": "| ResNet-18, CPU, one 64x64 image, one process | the same with `until=\"layer2\"` | 8.43 ms |"
  },
  {
   "section": "until_resnet18",
   "text": "| ResNet-18, CPU, one 64x64 image, one process | `trace.run_transaction(inputs=new_x, until=\"layer1\")`, without `fast=True` | 434.31 ms |"
  },
  {
   "section": "until_gpt2-small",
   "text": "| GPT-2 small (random weights), CPU, one forward of 16 sequences of 32 tokens, one process | plain forward, nothing attached | 1.62 s |"
  },
  {
   "section": "until_gpt2-small",
   "text": "| GPT-2 small (random weights), CPU, one forward of 16 sequences of 32 tokens, one process | `trace.run_transaction(inputs=new_x, fast=True)`, whole forward | 1.51 s |"
  },
  {
   "section": "until_gpt2-small",
   "text": "| GPT-2 small (random weights), CPU, one forward of 16 sequences of 32 tokens, one process | the same with `until=\"transformer.h.1\"` | 161.36 ms |"
  },
  {
   "section": "until_gpt2-small",
   "text": "| GPT-2 small (random weights), CPU, one forward of 16 sequences of 32 tokens, one process | `trace.run_transaction(inputs=new_x, until=\"transformer.h.1\")`, without `fast=True` | 3.40 s |"
  },
  {
   "section": "until_lstm-tagger",
   "text": "| LSTM tagger, CPU, one sequence of 32 tokens, one process | plain forward, nothing attached | 1.01 ms |"
  },
  {
   "section": "until_lstm-tagger",
   "text": "| LSTM tagger, CPU, one sequence of 32 tokens, one process | `trace.run_transaction(inputs=new_x, fast=True)`, whole forward | 3.36 ms |"
  },
  {
   "section": "until_lstm-tagger",
   "text": "| LSTM tagger, CPU, one sequence of 32 tokens, one process | the same with `until=\"embed\"` | 0.77 ms |"
  },
  {
   "section": "until_lstm-tagger",
   "text": "| LSTM tagger, CPU, one sequence of 32 tokens, one process | `trace.run_transaction(inputs=new_x, until=\"embed\")`, without `fast=True` | 51.97 ms |"
  },
  {
   "section": "replay_resnet18",
   "text": "| ResNet-18, CPU, one 64x64 image, one process | plain forward, nothing attached | 20.47 ms |"
  },
  {
   "section": "replay_resnet18",
   "text": "| ResNet-18, CPU, one 64x64 image, one process | fresh `tl.trace(model, x)` | 694.92 ms |"
  },
  {
   "section": "replay_resnet18",
   "text": "| ResNet-18, CPU, one 64x64 image, one process | fresh capture with `CaptureOptions(replayable=True)` | 3.37 s |"
  },
  {
   "section": "replay_resnet18",
   "text": "| ResNet-18, CPU, one 64x64 image, one process | replay of the loaded runnable file on new input, `loaded.run_transaction(inputs=new_x)` | 583.57 ms |"
  },
  {
   "section": "replay_gpt2-small",
   "text": "| GPT-2 small (random weights, no key-value cache), CPU, one forward of 16 sequences of 32 tokens, one process | plain forward, nothing attached | 2.34 s |"
  },
  {
   "section": "replay_gpt2-small",
   "text": "| GPT-2 small (random weights, no key-value cache), CPU, one forward of 16 sequences of 32 tokens, one process | fresh `tl.trace(model, x)` | 5.27 s |"
  },
  {
   "section": "replay_gpt2-small",
   "text": "| GPT-2 small (random weights, no key-value cache), CPU, one forward of 16 sequences of 32 tokens, one process | fresh capture with `CaptureOptions(replayable=True)` | 17.56 s |"
  },
  {
   "section": "replay_gpt2-small",
   "text": "| GPT-2 small (random weights, no key-value cache), CPU, one forward of 16 sequences of 32 tokens, one process | replay of the loaded runnable file on new input, `loaded.run_transaction(inputs=new_x)` | 10.40 s |"
  },
  {
   "section": "replay_lstm-tagger",
   "text": "| LSTM tagger, CPU, one sequence of 32 tokens, one process | plain forward, nothing attached | 1.59 ms |"
  },
  {
   "section": "replay_lstm-tagger",
   "text": "| LSTM tagger, CPU, one sequence of 32 tokens, one process | fresh `tl.trace(model, x)` | 66.93 ms |"
  },
  {
   "section": "replay_lstm-tagger",
   "text": "| LSTM tagger, CPU, one sequence of 32 tokens, one process | fresh capture with `CaptureOptions(replayable=True)` | 290.84 ms |"
  },
  {
   "section": "replay_lstm-tagger",
   "text": "| LSTM tagger, CPU, one sequence of 32 tokens, one process | replay of the loaded runnable file on new input, `loaded.run_transaction(inputs=new_x)` | 58.86 ms |"
  },
  {
   "section": "trace_run_fast_engine",
   "text": "| ResNet-18, CPU, one 64x64 image, n=3 processes | plain forward, nothing attached | 22.68 ms (20.90 ms to 23.02 ms) | 1.00x | 14.42 ms (14.28 ms to 15.26 ms) | 1.00x |"
  },
  {
   "section": "trace_run_fast_engine",
   "text": "| ResNet-18, CPU, one 64x64 image, n=3 processes | fresh `tl.trace(model, x, save=tl.module(site))` | 508.54 ms (490.37 ms to 709.94 ms) | 22.4x | 442.97 ms (423.87 ms to 442.99 ms) | 30.7x |"
  },
  {
   "section": "trace_run_fast_engine",
   "text": "| ResNet-18, CPU, one 64x64 image, n=3 processes | `trace.run(model, new_x)`, no edit staged | 651.95 ms (649.85 ms to 1.08 s) | 28.7x | 47.48 ms (46.99 ms to 48.89 ms) | 3.29x |"
  },
  {
   "section": "trace_run_fast_engine",
   "text": "| ResNet-18, CPU, one 64x64 image, n=3 processes | `trace.run_transaction(inputs=new_x, fast=True)`, no edit staged | 19.70 ms (18.89 ms to 29.97 ms) | 0.87x | 19.18 ms (19.15 ms to 19.80 ms) | 1.33x |"
  },
  {
   "section": "trace_run_fast_engine",
   "text": "| GPT-2 small (random weights), CPU, one forward of 16 sequences of 32 tokens, n=2 processes | plain forward, nothing attached | 1.29 s (1.28 s to 1.30 s) | 1.00x | 1.32 s (1.31 s to 1.34 s) | 1.00x |"
  },
  {
   "section": "trace_run_fast_engine",
   "text": "| GPT-2 small (random weights), CPU, one forward of 16 sequences of 32 tokens, n=2 processes | fresh `tl.trace(model, x, save=tl.module(site))` | 2.91 s (2.86 s to 2.96 s) | 2.26x | 2.95 s (2.87 s to 3.03 s) | 2.23x |"
  },
  {
   "section": "trace_run_fast_engine",
   "text": "| GPT-2 small (random weights), CPU, one forward of 16 sequences of 32 tokens, n=2 processes | `trace.run(model, new_x)`, no edit staged | 4.28 s (4.28 s to 4.29 s) | 3.32x | 1.50 s (1.47 s to 1.54 s) | 1.13x |"
  },
  {
   "section": "trace_run_fast_engine",
   "text": "| GPT-2 small (random weights), CPU, one forward of 16 sequences of 32 tokens, n=2 processes | `trace.run_transaction(inputs=new_x, fast=True)`, no edit staged | 1.45 s (1.45 s to 1.45 s) | 1.12x | 1.47 s (1.47 s to 1.47 s) | 1.11x |"
  },
  {
   "section": "trace_run_fast_engine",
   "text": "| LSTM tagger, CPU, one sequence of 32 tokens, n=3 processes | plain forward, nothing attached | 1.04 ms (1.04 ms to 1.07 ms) | 1.00x | 1.67 ms (1.62 ms to 1.67 ms) | 1.00x |"
  },
  {
   "section": "trace_run_fast_engine",
   "text": "| LSTM tagger, CPU, one sequence of 32 tokens, n=3 processes | fresh `tl.trace(model, x, save=tl.module(site))` | 42.45 ms (42.11 ms to 43.35 ms) | 40.7x | 70.56 ms (67.46 ms to 71.20 ms) | 42.4x |"
  },
  {
   "section": "trace_run_fast_engine",
   "text": "| LSTM tagger, CPU, one sequence of 32 tokens, n=3 processes | `trace.run(model, new_x)`, no edit staged | 64.86 ms (64.10 ms to 90.98 ms) | 62.2x | 11.58 ms (10.93 ms to 12.07 ms) | 6.96x |"
  },
  {
   "section": "trace_run_fast_engine",
   "text": "| LSTM tagger, CPU, one sequence of 32 tokens, n=3 processes | `trace.run_transaction(inputs=new_x, fast=True)`, no edit staged | 3.00 ms (2.97 ms to 3.04 ms) | 2.88x | 4.66 ms (4.45 ms to 4.85 ms) | 2.80x |"
  }
 ],
 "early_stop": {
  "equality": "the stopped site equals a fresh tl.trace of the new input (max abs diff 0) on all three models",
  "engine": "guarded_fast with fast=True"
 },
 "model_free_replay": {
  "host": "a shared 32-core Intel Xeon 2.80 GHz Linux CPU worker, one torch thread per process, load average 14 to 21",
  "route": "tl.trace(model, x, capture=CaptureOptions(replayable=True)); trace.save(path, level='runnable', include_weights=True); tl.load(path).run_transaction(inputs=new_x)",
  "one_time": {
   "resnet18": "save with weights 5.08 s (46.3 MiB), load 3.31 s",
   "gpt2-small": "save with weights 26.85 s (627.8 MiB), load 14.42 s",
   "lstm-tagger": "save with weights 1.15 s (1.7 MiB), load 498 ms"
  },
  "equality_vs_model_new_x": {
   "resnet18": "max abs diff 0",
   "gpt2-small": "max abs diff 0",
   "lstm-tagger": "max abs diff 3e-08"
  },
  "report": "path_faithfulness verified, numeric_attestation not_applicable, state_source embedded_capture_state on all three",
  "warnings": "none from the replay; GPT-2's default trace warned that it kept 1.46 GiB of saved activations"
 },
 "trace_run_fast_engine": {
  "builds": {
   "before": "TorchLens 2.38.1 development build, trace.run on an unedited trace recaptures",
   "after": "TorchLens 2.38.1 development build with the guarded fast engine for unedited trace.run"
  },
  "host": "a shared 32-core Intel Xeon 2.80 GHz Linux CPU worker, one torch thread per process, load average 6 to 15, both jobs at the same time",
  "engines": {
   "before": {
    "trace.run": "rerun",
    "run_transaction_fast": "guarded_fast"
   },
   "after": {
    "trace.run": "guarded_fast",
    "run_transaction_fast": "guarded_fast"
   }
  },
  "correctness": {
   "saved_site_vs_fresh_trace_max_abs_diff": 0.0,
   "output_vs_plain_forward_max_abs_diff": {
    "resnet18": 0.0,
    "gpt2-small": 0.0,
    "lstm-tagger": 2.98e-08
   }
  },
  "raw": {
   "before": {
    "resnet18/plain_forward": {
     "median_ms": 22.679,
     "lowest_ms": 20.903,
     "highest_ms": 23.024,
     "processes": 3
    },
    "resnet18/fresh_trace_one_site": {
     "median_ms": 508.542,
     "lowest_ms": 490.371,
     "highest_ms": 709.936,
     "processes": 3
    },
    "resnet18/trace_run_new_inputs": {
     "median_ms": 651.953,
     "lowest_ms": 649.854,
     "highest_ms": 1075.59,
     "processes": 3
    },
    "resnet18/run_transaction_fast": {
     "median_ms": 19.7,
     "lowest_ms": 18.892,
     "highest_ms": 29.971,
     "processes": 3
    },
    "gpt2-small/plain_forward": {
     "median_ms": 1289.405,
     "lowest_ms": 1282.959,
     "highest_ms": 1295.852,
     "processes": 2
    },
    "gpt2-small/fresh_trace_one_site": {
     "median_ms": 2911.838,
     "lowest_ms": 2860.185,
     "highest_ms": 2963.492,
     "processes": 2
    },
    "gpt2-small/trace_run_new_inputs": {
     "median_ms": 4284.328,
     "lowest_ms": 4281.158,
     "highest_ms": 4287.499,
     "processes": 2
    },
    "gpt2-small/run_transaction_fast": {
     "median_ms": 1448.925,
     "lowest_ms": 1445.931,
     "highest_ms": 1451.919,
     "processes": 2
    },
    "lstm-tagger/plain_forward": {
     "median_ms": 1.043,
     "lowest_ms": 1.04,
     "highest_ms": 1.073,
     "processes": 3
    },
    "lstm-tagger/fresh_trace_one_site": {
     "median_ms": 42.446,
     "lowest_ms": 42.111,
     "highest_ms": 43.348,
     "processes": 3
    },
    "lstm-tagger/trace_run_new_inputs": {
     "median_ms": 64.859,
     "lowest_ms": 64.104,
     "highest_ms": 90.975,
     "processes": 3
    },
    "lstm-tagger/run_transaction_fast": {
     "median_ms": 3.001,
     "lowest_ms": 2.97,
     "highest_ms": 3.04,
     "processes": 3
    }
   },
   "after": {
    "resnet18/plain_forward": {
     "median_ms": 14.422,
     "lowest_ms": 14.278,
     "highest_ms": 15.259,
     "processes": 3
    },
    "resnet18/fresh_trace_one_site": {
     "median_ms": 442.968,
     "lowest_ms": 423.869,
     "highest_ms": 442.994,
     "processes": 3
    },
    "resnet18/trace_run_new_inputs": {
     "median_ms": 47.482,
     "lowest_ms": 46.988,
     "highest_ms": 48.89,
     "processes": 3
    },
    "resnet18/run_transaction_fast": {
     "median_ms": 19.179,
     "lowest_ms": 19.154,
     "highest_ms": 19.801,
     "processes": 3
    },
    "gpt2-small/plain_forward": {
     "median_ms": 1324.271,
     "lowest_ms": 1309.242,
     "highest_ms": 1339.3,
     "processes": 2
    },
    "gpt2-small/fresh_trace_one_site": {
     "median_ms": 2950.042,
     "lowest_ms": 2868.602,
     "highest_ms": 3031.482,
     "processes": 2
    },
    "gpt2-small/trace_run_new_inputs": {
     "median_ms": 1502.593,
     "lowest_ms": 1469.545,
     "highest_ms": 1535.641,
     "processes": 2
    },
    "gpt2-small/run_transaction_fast": {
     "median_ms": 1470.533,
     "lowest_ms": 1470.425,
     "highest_ms": 1470.641,
     "processes": 2
    },
    "lstm-tagger/plain_forward": {
     "median_ms": 1.665,
     "lowest_ms": 1.624,
     "highest_ms": 1.667,
     "processes": 3
    },
    "lstm-tagger/fresh_trace_one_site": {
     "median_ms": 70.559,
     "lowest_ms": 67.465,
     "highest_ms": 71.195,
     "processes": 3
    },
    "lstm-tagger/trace_run_new_inputs": {
     "median_ms": 11.584,
     "lowest_ms": 10.925,
     "highest_ms": 12.074,
     "processes": 3
    },
    "lstm-tagger/run_transaction_fast": {
     "median_ms": 4.659,
     "lowest_ms": 4.451,
     "highest_ms": 4.852,
     "processes": 3
    }
   }
  }
 }
}