==============================================================================
target=Qwen/Qwen2.5-1.5B-Instruct  drafter=Qwen/Qwen2.5-0.5B-Instruct  dataset=humaneval  mode=greedy
commit=fdd3be9  n_samples=10
------------------------------------------------------------------------------
[arm_composite-g4]
  workload_tokens_per_second = 80.62466720802476
  tokens_per_second_median = 82.69247134612439
  tokens_per_second_p25 = 74.30528068803494
  tokens_per_second_p75 = 86.66982882303544
  tokens_per_second_mean = 81.78682322631008
  acceptance_rate_mean = 0.5109862676726286
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g6-f2_median = 0.9728776132506674
  speedup_vs_hf_greedy_compiled_median = 1.5838703834487282
  speedup_vs_hf_greedy_median = 3.2569559488245274
  forward_counts.emitted_tokens_per_step = 2.942528735632184
  forward_counts.n_generations = 10
  forward_counts.n_steps = 435
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 445
  forward_counts.target_forwards_per_emitted_token = 0.34765625
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.009712600269589925
  step_profile.bookkeeping_ms_mean = 0.3537306489101772
  step_profile.bookkeeping_ms_total = 153.87283227592707
  step_profile.draft_fraction = 0.4450005557066006
  step_profile.draft_ms_mean = 16.206817017718294
  step_profile.draft_ms_total = 7049.9654027074575
  step_profile.n_steps = 435.0
  step_profile.sample_fraction = 0.010180864843941663
  step_profile.sample_ms_mean = 0.37078473609069296
  step_profile.sample_ms_total = 161.29136019945145
  step_profile.step_ms_mean = 36.419768042725394
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.942528735632184
  step_profile.verify_fraction = 0.5351059791798678
  step_profile.verify_ms_mean = 19.48843564000623
  step_profile.verify_ms_total = 8477.46950340271
  host_syncs.by_kind = {'cuda': 46}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 40
  host_syncs.per_step = 1.15
  host_syncs.prompt_id = HumanEval/23
  host_syncs.total = 46
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
  greedy_token_match.first_divergence = [None, None, 94, None, None, None, None, None, None, None]
  greedy_token_match.n_prompts = 10
  greedy_token_match.prompts_matched = 9
  greedy_token_match.reference = hf_greedy
  greedy_token_match.source = this block
  greedy_token_match_compiled.first_divergence = [None, None, None, None, None, None, None, None, None, None]
  greedy_token_match_compiled.n_prompts = 10
  greedy_token_match_compiled.prompts_matched = 10
  greedy_token_match_compiled.reference = hf_greedy_compiled
  greedy_token_match_compiled.source = this block
[arm_composite-g6-f2]
  workload_tokens_per_second = 82.44928394074864
  tokens_per_second_median = 86.33038153255222
  tokens_per_second_p25 = 73.72680951645049
  tokens_per_second_p75 = 89.79680868040586
  tokens_per_second_mean = 83.98114311711102
  acceptance_rate_mean = 0.38034513434577794
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g4_median = 1.0279037671682698
  speedup_vs_hf_greedy_compiled_median = 1.6450044310517886
  speedup_vs_hf_greedy_median = 3.408510509483143
  forward_counts.emitted_tokens_per_step = 2.4758220502901356
  forward_counts.n_generations = 10
  forward_counts.n_steps = 517
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 527
  forward_counts.target_forwards_per_emitted_token = 0.41171875
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.010763423630084937
  step_profile.bookkeeping_ms_mean = 0.3224823207556625
  step_profile.bookkeeping_ms_total = 166.7233598306775
  step_profile.draft_fraction = 0.32919959185116915
  step_profile.draft_ms_mean = 9.863130173121707
  step_profile.draft_ms_total = 5099.2382995039225
  step_profile.n_steps = 517.0
  step_profile.sample_fraction = 0.01232071711209039
  step_profile.sample_ms_mean = 0.3691403018436764
  step_profile.sample_ms_total = 190.8455360531807
  step_profile.step_ms_mean = 29.9609428968576
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.4758220502901356
  step_profile.verify_fraction = 0.6477162674066556
  step_profile.verify_ms_mean = 19.406190101136552
  step_profile.verify_ms_total = 10033.000282287598
  host_syncs.by_kind = {'cuda': 54}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 49
  host_syncs.per_step = 1.1020408163265305
  host_syncs.prompt_id = HumanEval/23
  host_syncs.total = 54
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
  greedy_token_match.first_divergence = [None, None, None, None, None, None, None, None, None, None]
  greedy_token_match.n_prompts = 10
  greedy_token_match.prompts_matched = 10
  greedy_token_match.reference = hf_greedy
  greedy_token_match.source = this block
  greedy_token_match_compiled.first_divergence = [None, None, 94, None, None, None, None, None, None, None]
  greedy_token_match_compiled.n_prompts = 10
  greedy_token_match_compiled.prompts_matched = 9
  greedy_token_match_compiled.reference = hf_greedy_compiled
  greedy_token_match_compiled.source = this block
[hf_greedy]
  workload_tokens_per_second = 25.331419097732105
  tokens_per_second_median = 25.347841774034904
  tokens_per_second_p25 = 25.232138805855218
  tokens_per_second_p75 = 25.44860211425452
  tokens_per_second_mean = 25.33244326004205
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g4_median = 0.3073587272719037
  speedup_vs_arm_composite-g6-f2_median = 0.2934362449060071
  speedup_vs_hf_greedy_compiled_median = 0.4859622753017314
  forward_counts.n_generations = 10
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1280
  forward_counts.target_forwards_per_emitted_token = 1.0
  forward_counts.tokens_emitted = 1280
  host_syncs.scope = not probed for HuggingFace baselines in the phase6 runner
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[hf_greedy_compiled]
  workload_tokens_per_second = 52.2244160121773
  tokens_per_second_median = 52.07049803853526
  tokens_per_second_p25 = 51.98004475227953
  tokens_per_second_p75 = 52.506992588813276
  tokens_per_second_mean = 52.22632581218987
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g4_median = 0.63229933383584
  speedup_vs_arm_composite-g6-f2_median = 0.6079290128344301
  speedup_vs_hf_greedy_median = 2.0577729667196065
  forward_counts.n_generations = 10
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1280
  forward_counts.target_forwards_per_emitted_token = 1.0
  forward_counts.tokens_emitted = 1280
  host_syncs.scope = not probed for HuggingFace baselines in the phase6 runner
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
  greedy_token_match.first_divergence = [None, None, 94, None, None, None, None, None, None, None]
  greedy_token_match.n_prompts = 10
  greedy_token_match.prompts_matched = 9
  greedy_token_match.reference = hf_greedy
  greedy_token_match.source = this block

==============================================================================
target=Qwen/Qwen2.5-1.5B-Instruct  drafter=Qwen/Qwen2.5-0.5B-Instruct  dataset=humaneval  mode=sample T=0.7
commit=fdd3be9  n_samples=10
------------------------------------------------------------------------------
[arm_composite-g4]
  workload_tokens_per_second = 68.46233813773001
  tokens_per_second_median = 64.76457268630614
  tokens_per_second_p25 = 62.41346310862339
  tokens_per_second_p75 = 77.35423033265056
  tokens_per_second_mean = 70.01739284875066
  acceptance_rate_mean = 0.4941384849358883
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g6-f2_median = 0.9939617957550109
  speedup_vs_hf_sample_compiled_median = 1.2554473217844095
  forward_counts.emitted_tokens_per_step = 2.800875273522976
  forward_counts.n_generations = 10
  forward_counts.n_steps = 457
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 467
  forward_counts.target_forwards_per_emitted_token = 0.36484375
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.009464413874514398
  step_profile.bookkeeping_ms_mean = 0.3864318598347107
  step_profile.bookkeeping_ms_total = 176.59935994446278
  step_profile.draft_fraction = 0.4990623222587203
  step_profile.draft_ms_mean = 20.376706251528123
  step_profile.draft_ms_total = 9312.154756948352
  step_profile.n_steps = 457.0
  step_profile.sample_fraction = 0.014317692028535358
  step_profile.sample_ms_mean = 0.5845911255028889
  step_profile.sample_ms_total = 267.15814435482025
  step_profile.step_ms_mean = 40.82998323597063
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.800875273522976
  step_profile.verify_fraction = 0.47715557183822993
  step_profile.verify_ms_mean = 19.482253999104906
  step_profile.verify_ms_total = 8903.390077590942
  host_syncs.by_kind = {'cuda': 43}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 38
  host_syncs.per_step = 1.131578947368421
  host_syncs.prompt_id = HumanEval/23
  host_syncs.total = 43
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[arm_composite-g6-f2]
  workload_tokens_per_second = 75.5425340623311
  tokens_per_second_median = 77.0495039950539
  tokens_per_second_p25 = 66.25346521109498
  tokens_per_second_p75 = 89.59897114960685
  tokens_per_second_mean = 77.7675328735106
  acceptance_rate_mean = 0.38243354825314047
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g4_median = 1.0066699115211928
  speedup_vs_hf_sample_compiled_median = 1.4918978008053516
  forward_counts.emitted_tokens_per_step = 2.401500938086304
  forward_counts.n_generations = 10
  forward_counts.n_steps = 533
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 543
  forward_counts.target_forwards_per_emitted_token = 0.42421875
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.010464069978050808
  step_profile.bookkeeping_ms_mean = 0.3319460261354527
  step_profile.bookkeeping_ms_total = 176.92723193019629
  step_profile.draft_fraction = 0.35982274704885836
  step_profile.draft_ms_mean = 11.414462178344463
  step_profile.draft_ms_total = 6083.908341057599
  step_profile.n_steps = 533.0
  step_profile.sample_fraction = 0.018477110649285495
  step_profile.sample_ms_mean = 0.5861393766823674
  step_profile.sample_ms_total = 312.4122877717018
  step_profile.step_ms_mean = 31.722458549277196
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.401500938086304
  step_profile.verify_fraction = 0.6112360723238054
  step_profile.verify_ms_mean = 19.389910968114915
  step_profile.verify_ms_total = 10334.822546005249
  host_syncs.by_kind = {'cuda': 62}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 58
  host_syncs.per_step = 1.0689655172413792
  host_syncs.prompt_id = HumanEval/23
  host_syncs.total = 62
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[hf_sample_compiled]
  workload_tokens_per_second = 51.5730691915891
  tokens_per_second_median = 51.53062472483185
  tokens_per_second_p25 = 51.51815338857841
  tokens_per_second_p75 = 51.632433302414064
  tokens_per_second_mean = 51.57319856270658
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g4_median = 0.7967838112482855
  speedup_vs_arm_composite-g6-f2_median = 0.6787660619193359
  forward_counts.n_generations = 10
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1280
  forward_counts.target_forwards_per_emitted_token = 1.0
  forward_counts.tokens_emitted = 1280
  host_syncs.scope = not probed for HuggingFace baselines in the phase6 runner
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile

==============================================================================
target=Qwen/Qwen2.5-1.5B-Instruct  drafter=Qwen/Qwen2.5-0.5B-Instruct  dataset=humaneval_edit  mode=greedy
commit=fdd3be9  n_samples=10
------------------------------------------------------------------------------
[arm_composite-g4]
  workload_tokens_per_second = 142.5312928898297
  tokens_per_second_median = 162.1850178199153
  tokens_per_second_p25 = 147.469931562155
  tokens_per_second_p75 = 177.95275944576818
  tokens_per_second_mean = 152.3326317042343
  acceptance_rate_mean = 0.6083112276000873
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g6-f2_median = 0.829452972899918
  speedup_vs_hf_greedy_compiled_median = 3.123346963310837
  speedup_vs_hf_greedy_median = 6.334704848394402
  forward_counts.emitted_tokens_per_step = 3.6260623229461757
  forward_counts.n_generations = 10
  forward_counts.n_steps = 353
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 363
  forward_counts.target_forwards_per_emitted_token = 0.28359375
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.010619624536242612
  step_profile.bookkeeping_ms_mean = 0.26916595997958953
  step_profile.bookkeeping_ms_total = 95.0155838727951
  step_profile.draft_fraction = 0.173378266588746
  step_profile.draft_ms_mean = 4.3944611607209145
  step_profile.draft_ms_total = 1551.2447897344828
  step_profile.n_steps = 353.0
  step_profile.sample_fraction = 0.015259429190859888
  step_profile.sample_ms_mean = 0.3867668666515404
  step_profile.sample_ms_total = 136.52870392799377
  step_profile.step_ms_mean = 25.346090067589586
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 3.6260623229461757
  step_profile.verify_fraction = 0.8007426796841515
  step_profile.verify_ms_mean = 20.295696080237544
  step_profile.verify_ms_total = 7164.3807163238525
  host_syncs.by_kind = {'cuda': 57}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 53
  host_syncs.per_step = 1.0754716981132075
  host_syncs.prompt_id = HumanEval/0:docstring
  host_syncs.total = 57
  nonfinite_logits.prompt_id = HumanEval/0:docstring
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
  greedy_token_match.first_divergence = [None, None, None, None, None, None, None, None, None, None]
  greedy_token_match.n_prompts = 10
  greedy_token_match.prompts_matched = 10
  greedy_token_match.reference = hf_greedy
  greedy_token_match.source = this block
  greedy_token_match_compiled.first_divergence = [None, None, None, None, None, None, None, None, None, None]
  greedy_token_match_compiled.n_prompts = 10
  greedy_token_match_compiled.prompts_matched = 10
  greedy_token_match_compiled.reference = hf_greedy_compiled
  greedy_token_match_compiled.source = this block
[arm_composite-g6-f2]
  workload_tokens_per_second = 165.21993606626214
  tokens_per_second_median = 206.48259325938017
  tokens_per_second_p25 = 172.3910612392009
  tokens_per_second_p75 = 217.98209623208118
  tokens_per_second_mean = 183.52529846144594
  acceptance_rate_mean = 0.5062178287079845
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g4_median = 1.2058009043010172
  speedup_vs_hf_greedy_compiled_median = 3.977571001758399
  speedup_vs_hf_greedy_median = 8.10588051975406
  forward_counts.emitted_tokens_per_step = 4.102564102564102
  forward_counts.n_generations = 10
  forward_counts.n_steps = 312
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 322
  forward_counts.target_forwards_per_emitted_token = 0.2515625
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.011922319771584557
  step_profile.bookkeeping_ms_mean = 0.2949813330689302
  step_profile.bookkeeping_ms_total = 92.03417591750622
  step_profile.draft_fraction = 0.14245810620444557
  step_profile.draft_ms_mean = 3.524690067013487
  step_profile.draft_ms_total = 1099.703300908208
  step_profile.n_steps = 312.0
  step_profile.sample_fraction = 0.015674275448872045
  step_profile.sample_ms_mean = 0.3878119992904174
  step_profile.sample_ms_total = 120.99734377861023
  step_profile.step_ms_mean = 24.74194105848288
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 4.102564102564102
  step_profile.verify_fraction = 0.8299452985750978
  step_profile.verify_ms_mean = 20.534457659110046
  step_profile.verify_ms_total = 6406.750789642334
  host_syncs.by_kind = {'cuda': 59}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 55
  host_syncs.per_step = 1.0727272727272728
  host_syncs.prompt_id = HumanEval/0:docstring
  host_syncs.total = 59
  nonfinite_logits.prompt_id = HumanEval/0:docstring
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
  greedy_token_match.first_divergence = [None, None, None, None, None, None, None, None, None, None]
  greedy_token_match.n_prompts = 10
  greedy_token_match.prompts_matched = 10
  greedy_token_match.reference = hf_greedy
  greedy_token_match.source = this block
  greedy_token_match_compiled.first_divergence = [None, None, None, None, None, None, None, None, None, None]
  greedy_token_match_compiled.n_prompts = 10
  greedy_token_match_compiled.prompts_matched = 10
  greedy_token_match_compiled.reference = hf_greedy_compiled
  greedy_token_match_compiled.source = this block
[hf_greedy]
  workload_tokens_per_second = 25.395624299688595
  tokens_per_second_median = 25.48365280359434
  tokens_per_second_p25 = 25.23306395503338
  tokens_per_second_p75 = 25.564220460024913
  tokens_per_second_mean = 25.397806149177264
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g4_median = 0.1581108475119125
  speedup_vs_arm_composite-g6-f2_median = 0.12355068093682754
  speedup_vs_hf_greedy_compiled_median = 0.49082230037605606
  forward_counts.n_generations = 10
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1280
  forward_counts.target_forwards_per_emitted_token = 1.0
  forward_counts.tokens_emitted = 1280
  host_syncs.scope = not probed for HuggingFace baselines in the phase6 runner
  nonfinite_logits.prompt_id = HumanEval/0:docstring
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[hf_greedy_compiled]
  workload_tokens_per_second = 51.888530111858174
  tokens_per_second_median = 51.86651033594373
  tokens_per_second_p25 = 51.843917106744385
  tokens_per_second_p75 = 51.967543737137376
  tokens_per_second_mean = 51.88866545128307
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g4_median = 0.3207585670229729
  speedup_vs_arm_composite-g6-f2_median = 0.2519566925755855
  speedup_vs_hf_greedy_median = 2.0374219803465103
  forward_counts.n_generations = 10
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1280
  forward_counts.target_forwards_per_emitted_token = 1.0
  forward_counts.tokens_emitted = 1280
  host_syncs.scope = not probed for HuggingFace baselines in the phase6 runner
  nonfinite_logits.prompt_id = HumanEval/0:docstring
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
  greedy_token_match.first_divergence = [None, None, None, None, None, None, None, None, None, None]
  greedy_token_match.n_prompts = 10
  greedy_token_match.prompts_matched = 10
  greedy_token_match.reference = hf_greedy
  greedy_token_match.source = this block

==============================================================================
target=Qwen/Qwen2.5-1.5B-Instruct  drafter=Qwen/Qwen2.5-0.5B-Instruct  dataset=humaneval_edit  mode=sample T=0.7
commit=fdd3be9  n_samples=10
------------------------------------------------------------------------------
[arm_composite-g4]
  workload_tokens_per_second = 144.5027089188522
  tokens_per_second_median = 151.20266463019163
  tokens_per_second_p25 = 141.65342959066368
  tokens_per_second_p75 = 176.54671973149436
  tokens_per_second_mean = 152.86546795785188
  acceptance_rate_mean = 0.6123546770190365
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g6-f2_median = 0.8079334973089571
  speedup_vs_hf_sample_compiled_median = 2.936098808748993
  forward_counts.emitted_tokens_per_step = 3.67816091954023
  forward_counts.n_generations = 10
  forward_counts.n_steps = 348
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 358
  forward_counts.target_forwards_per_emitted_token = 0.2796875
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.010519904312703637
  step_profile.bookkeeping_ms_mean = 0.26682473539278423
  step_profile.bookkeeping_ms_total = 92.85500791668892
  step_profile.draft_fraction = 0.1632987063001443
  step_profile.draft_ms_mean = 4.141875515531363
  step_profile.draft_ms_total = 1441.3726794049144
  step_profile.n_steps = 348.0
  step_profile.sample_fraction = 0.02368513555937458
  step_profile.sample_ms_mean = 0.600745010650021
  step_profile.sample_ms_total = 209.05926370620728
  step_profile.step_ms_mean = 25.363798705903793
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 3.67816091954023
  step_profile.verify_fraction = 0.8024962538277775
  step_profile.verify_ms_mean = 20.354353444329625
  step_profile.verify_ms_total = 7083.314998626709
  host_syncs.by_kind = {'cuda': 59}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 55
  host_syncs.per_step = 1.0727272727272728
  host_syncs.prompt_id = HumanEval/0:docstring
  host_syncs.total = 59
  nonfinite_logits.prompt_id = HumanEval/0:docstring
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[arm_composite-g6-f2]
  workload_tokens_per_second = 186.74422014458275
  tokens_per_second_median = 206.59474760555716
  tokens_per_second_p25 = 176.5079452362672
  tokens_per_second_p75 = 229.18632912677563
  tokens_per_second_mean = 204.02557780425624
  acceptance_rate_mean = 0.5494511212019099
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g4_median = 1.2383625492476038
  speedup_vs_hf_sample_compiled_median = 4.010882244552153
  forward_counts.emitted_tokens_per_step = 4.507042253521127
  forward_counts.n_generations = 10
  forward_counts.n_steps = 284
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 294
  forward_counts.target_forwards_per_emitted_token = 0.2296875
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.011691924764406107
  step_profile.bookkeeping_ms_mean = 0.2810385357850874
  step_profile.bookkeeping_ms_total = 79.81494416296482
  step_profile.draft_fraction = 0.09507043489210346
  step_profile.draft_ms_mean = 2.28520593117975
  step_profile.draft_ms_total = 648.998484455049
  step_profile.n_steps = 284.0
  step_profile.sample_fraction = 0.02520542786670183
  step_profile.sample_ms_mean = 0.605862309605303
  step_profile.sample_ms_total = 172.06489592790604
  step_profile.step_ms_mean = 24.03697778150754
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 4.507042253521127
  step_profile.verify_fraction = 0.8680322124767886
  step_profile.verify_ms_mean = 20.8648710049374
  step_profile.verify_ms_total = 5925.623365402222
  host_syncs.by_kind = {'cuda': 61}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 57
  host_syncs.per_step = 1.0701754385964912
  host_syncs.prompt_id = HumanEval/0:docstring
  host_syncs.total = 61
  nonfinite_logits.prompt_id = HumanEval/0:docstring
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[hf_sample_compiled]
  workload_tokens_per_second = 51.49377273142095
  tokens_per_second_median = 51.498787962674484
  tokens_per_second_p25 = 51.46807390563348
  tokens_per_second_p75 = 51.52655645254278
  tokens_per_second_mean = 51.49385726622923
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g4_median = 0.3408132022655452
  speedup_vs_arm_composite-g6-f2_median = 0.24981690566401935
  forward_counts.n_generations = 10
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1280
  forward_counts.target_forwards_per_emitted_token = 1.0
  forward_counts.tokens_emitted = 1280
  host_syncs.scope = not probed for HuggingFace baselines in the phase6 runner
  nonfinite_logits.prompt_id = HumanEval/0:docstring
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile

==============================================================================
target=Qwen/Qwen2.5-1.5B-Instruct  drafter=Qwen/Qwen2.5-0.5B-Instruct  dataset=mixed  mode=greedy
commit=fdd3be9  n_samples=30
------------------------------------------------------------------------------
[arm_composite-g4]
  workload_tokens_per_second = 84.07542690617795
  tokens_per_second_median = 83.49931194723634
  tokens_per_second_p25 = 67.44527421352002
  tokens_per_second_p75 = 131.3565853035022
  tokens_per_second_mean = 99.62666768684197
  acceptance_rate_mean = 0.4929740645944779
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g6-f2_median = 0.9439716766968269
  forward_counts.emitted_tokens_per_step = 2.909090909090909
  forward_counts.n_generations = 30
  forward_counts.n_steps = 1320
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1350
  forward_counts.target_forwards_per_emitted_token = 0.3515625
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 3840
  step_profile.bookkeeping_fraction = 0.010363337248721759
  step_profile.bookkeeping_ms_mean = 0.35778707826905176
  step_profile.bookkeeping_ms_total = 472.27894331514835
  step_profile.draft_fraction = 0.4105690878881661
  step_profile.draft_ms_mean = 14.174614881052438
  step_profile.draft_ms_total = 18710.49164298922
  step_profile.n_steps = 1320.0
  step_profile.sample_fraction = 0.010992874854821803
  step_profile.sample_ms_mean = 0.3795214303740949
  step_profile.sample_ms_total = 500.9682880938053
  step_profile.step_ms_mean = 34.52431100929213
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 2.909090909090909
  step_profile.verify_fraction = 0.5680747000082903
  step_profile.verify_ms_mean = 19.61238761959654
  step_profile.verify_ms_total = 25888.35165786743
  host_syncs.by_kind = {'cuda': 50}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 46
  host_syncs.per_step = 1.0869565217391304
  host_syncs.prompt_id = mt_bench/121
  host_syncs.total = 50
  nonfinite_logits.prompt_id = mt_bench/121
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
  greedy_token_match.first_divergence = [None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, 94, None]
  greedy_token_match.n_prompts = 30
  greedy_token_match.prompts_matched = 29
  greedy_token_match.reference = hf_greedy
  greedy_token_match.source = per-set blocks, by prompt id
  greedy_token_match_compiled.first_divergence = [None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None]
  greedy_token_match_compiled.n_prompts = 30
  greedy_token_match_compiled.prompts_matched = 30
  greedy_token_match_compiled.reference = hf_greedy_compiled
  greedy_token_match_compiled.source = per-set blocks, by prompt id
[arm_composite-g6-f2]
  workload_tokens_per_second = 90.56017198477832
  tokens_per_second_median = 87.16112205108863
  tokens_per_second_p25 = 69.55390026280371
  tokens_per_second_p75 = 155.34229369114695
  tokens_per_second_mean = 112.03591132388652
  acceptance_rate_mean = 0.38572982267068295
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g4_median = 1.0593685677483764
  forward_counts.emitted_tokens_per_step = 2.6265389876880985
  forward_counts.n_generations = 30
  forward_counts.n_steps = 1462
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1492
  forward_counts.target_forwards_per_emitted_token = 0.3885416666666667
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 3840
  step_profile.bookkeeping_fraction = 0.011509742653466375
  step_profile.bookkeeping_ms_mean = 0.3330234309350091
  step_profile.bookkeeping_ms_total = 486.88025602698326
  step_profile.draft_fraction = 0.2987126958957002
  step_profile.draft_ms_mean = 8.642967079813234
  step_profile.draft_ms_total = 12636.017870686948
  step_profile.n_steps = 1462.0
  step_profile.sample_fraction = 0.013087435163448399
  step_profile.sample_ms_mean = 0.37867245962779594
  step_profile.sample_ms_total = 553.6191359758377
  step_profile.step_ms_mean = 28.934046656092075
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 2.6265389876880985
  step_profile.verify_fraction = 0.676690126287385
  step_profile.verify_ms_mean = 19.579383685716035
  step_profile.verify_ms_total = 28625.058948516846
  host_syncs.by_kind = {'cuda': 59}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 55
  host_syncs.per_step = 1.0727272727272728
  host_syncs.prompt_id = mt_bench/121
  host_syncs.total = 59
  nonfinite_logits.prompt_id = mt_bench/121
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
  greedy_token_match.first_divergence = [None, None, None, None, None, 74, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, 44, None, None, None, None]
  greedy_token_match.n_prompts = 30
  greedy_token_match.prompts_matched = 28
  greedy_token_match.reference = hf_greedy
  greedy_token_match.source = per-set blocks, by prompt id
  greedy_token_match_compiled.first_divergence = [None, None, None, None, None, 74, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, 44, None, None, 94, None]
  greedy_token_match_compiled.n_prompts = 30
  greedy_token_match_compiled.prompts_matched = 27
  greedy_token_match_compiled.reference = hf_greedy_compiled
  greedy_token_match_compiled.source = per-set blocks, by prompt id

==============================================================================
target=Qwen/Qwen2.5-1.5B-Instruct  drafter=Qwen/Qwen2.5-0.5B-Instruct  dataset=mixed  mode=sample T=0.7
commit=fdd3be9  n_samples=30
------------------------------------------------------------------------------
[arm_composite-g4]
  workload_tokens_per_second = 78.28316554888809
  tokens_per_second_median = 68.59000370625498
  tokens_per_second_p25 = 62.87793205360281
  tokens_per_second_p75 = 138.64051690441855
  tokens_per_second_mean = 95.1234411055792
  acceptance_rate_mean = 0.4550193832379414
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g6-f2_median = 0.9267900386524093
  forward_counts.emitted_tokens_per_step = 2.8635346756152127
  forward_counts.n_generations = 30
  forward_counts.n_steps = 1341
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1371
  forward_counts.target_forwards_per_emitted_token = 0.35703125
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 3840
  step_profile.bookkeeping_fraction = 0.010279615172995669
  step_profile.bookkeeping_ms_mean = 0.3752665530715213
  step_profile.bookkeeping_ms_total = 503.23244766891
  step_profile.draft_fraction = 0.43637254189412633
  step_profile.draft_ms_mean = 15.930170234567766
  step_profile.draft_ms_total = 21362.358284555376
  step_profile.n_steps = 1341.0
  step_profile.sample_fraction = 0.016231410322885456
  step_profile.sample_ms_mean = 0.5925421624108982
  step_profile.sample_ms_total = 794.5990397930145
  step_profile.step_ms_mean = 36.50589509005537
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 2.8635346756152127
  step_profile.verify_fraction = 0.5371164326099925
  step_profile.verify_ms_mean = 19.60791614000518
  step_profile.verify_ms_total = 26294.21554374695
  host_syncs.by_kind = {'cuda': 52}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 48
  host_syncs.per_step = 1.0833333333333333
  host_syncs.prompt_id = mt_bench/121
  host_syncs.total = 52
  nonfinite_logits.prompt_id = mt_bench/121
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[arm_composite-g6-f2]
  workload_tokens_per_second = 89.04019967993284
  tokens_per_second_median = 86.99574546523223
  tokens_per_second_p25 = 65.03467254068062
  tokens_per_second_p75 = 175.27476018804788
  tokens_per_second_mean = 117.26473158950604
  acceptance_rate_mean = 0.4145248426591108
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g4_median = 1.0790599315679197
  forward_counts.emitted_tokens_per_step = 2.6759581881533103
  forward_counts.n_generations = 30
  forward_counts.n_steps = 1435
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1465
  forward_counts.target_forwards_per_emitted_token = 0.3815104166666667
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 3840
  step_profile.bookkeeping_fraction = 0.011130691149286583
  step_profile.bookkeeping_ms_mean = 0.3337405096466948
  step_profile.bookkeeping_ms_total = 478.9176313430071
  step_profile.draft_fraction = 0.31367717180101645
  step_profile.draft_ms_mean = 9.405236186803632
  step_profile.draft_ms_total = 13496.513928063214
  step_profile.n_steps = 1435.0
  step_profile.sample_fraction = 0.019745387812537454
  step_profile.sample_ms_mean = 0.5920419229447634
  step_profile.sample_ms_total = 849.5801594257355
  step_profile.step_ms_mean = 29.983808298201303
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 2.6759581881533103
  step_profile.verify_fraction = 0.6554467492371595
  step_profile.verify_ms_mean = 19.65278967880621
  step_profile.verify_ms_total = 28201.753189086914
  host_syncs.by_kind = {'cuda': 63}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 59
  host_syncs.per_step = 1.0677966101694916
  host_syncs.prompt_id = mt_bench/121
  host_syncs.total = 63
  nonfinite_logits.prompt_id = mt_bench/121
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile

==============================================================================
target=Qwen/Qwen2.5-1.5B-Instruct  drafter=Qwen/Qwen2.5-0.5B-Instruct  dataset=mt_bench  mode=greedy
commit=fdd3be9  n_samples=10
------------------------------------------------------------------------------
[arm_composite-g4]
  workload_tokens_per_second = 60.28842019692385
  tokens_per_second_median = 58.05117973943692
  tokens_per_second_p25 = 56.81607268654288
  tokens_per_second_p75 = 66.21050684461966
  tokens_per_second_mean = 61.36854243985274
  acceptance_rate_mean = 0.36448637869169703
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g6-f2_median = 0.9512799363535468
  speedup_vs_hf_greedy_compiled_median = 1.1161790895142627
  speedup_vs_hf_greedy_median = 2.316411596178774
  forward_counts.emitted_tokens_per_step = 2.406015037593985
  forward_counts.n_generations = 10
  forward_counts.n_steps = 532
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 542
  forward_counts.target_forwards_per_emitted_token = 0.4234375
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.010737726048164928
  step_profile.bookkeeping_ms_mean = 0.42775410485166804
  step_profile.bookkeeping_ms_total = 227.5651837810874
  step_profile.draft_fraction = 0.4970118408679133
  step_profile.draft_ms_mean = 19.79924372604635
  step_profile.draft_ms_total = 10533.197662256658
  step_profile.n_steps = 532.0
  step_profile.sample_fraction = 0.009421669063600797
  step_profile.sample_ms_mean = 0.37532691730368406
  step_profile.sample_ms_total = 199.67392000555992
  step_profile.step_ms_mean = 39.83656343372357
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.406015037593985
  step_profile.verify_fraction = 0.482828764020321
  step_profile.verify_ms_mean = 19.234238685521863
  step_profile.verify_ms_total = 10232.614980697632
  host_syncs.by_kind = {'cuda': 54}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 50
  host_syncs.per_step = 1.08
  host_syncs.prompt_id = 81
  host_syncs.total = 54
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
  greedy_token_match.first_divergence = [None, None, None, None, None, None, None, None, None, None]
  greedy_token_match.n_prompts = 10
  greedy_token_match.prompts_matched = 10
  greedy_token_match.reference = hf_greedy
  greedy_token_match.source = this block
  greedy_token_match_compiled.first_divergence = [None, None, None, None, None, None, None, None, None, None]
  greedy_token_match_compiled.n_prompts = 10
  greedy_token_match_compiled.prompts_matched = 10
  greedy_token_match_compiled.reference = hf_greedy_compiled
  greedy_token_match_compiled.source = this block
[arm_composite-g6-f2]
  workload_tokens_per_second = 65.64071641266041
  tokens_per_second_median = 65.40865810966457
  tokens_per_second_p25 = 59.51450710814835
  tokens_per_second_p75 = 72.03927134184724
  tokens_per_second_mean = 66.42539972831128
  acceptance_rate_mean = 0.281242119473989
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g4_median = 1.0513822288245316
  speedup_vs_hf_greedy_compiled_median = 1.2567906433133345
  speedup_vs_hf_greedy_median = 2.6175018048782546
  forward_counts.emitted_tokens_per_step = 2.022116903633491
  forward_counts.n_generations = 10
  forward_counts.n_steps = 633
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 643
  forward_counts.target_forwards_per_emitted_token = 0.50234375
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.011726929475416118
  step_profile.bookkeeping_ms_mean = 0.3605518866842392
  step_profile.bookkeeping_ms_total = 228.2293442711234
  step_profile.draft_fraction = 0.3511018307287136
  step_profile.draft_ms_mean = 10.794848536687063
  step_profile.draft_ms_total = 6833.139123722911
  step_profile.n_steps = 633.0
  step_profile.sample_fraction = 0.012215678737736347
  step_profile.sample_ms_mean = 0.3755787928334156
  step_profile.sample_ms_total = 237.7413758635521
  step_profile.step_ms_mean = 30.74563443398259
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.022116903633491
  step_profile.verify_fraction = 0.624955561058134
  step_profile.verify_ms_mean = 19.214655217777874
  step_profile.verify_ms_total = 12162.876752853394
  host_syncs.by_kind = {'cuda': 71}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 67
  host_syncs.per_step = 1.0597014925373134
  host_syncs.prompt_id = 81
  host_syncs.total = 71
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
  greedy_token_match.first_divergence = [74, 44, None, None, None, None, None, None, None, None]
  greedy_token_match.n_prompts = 10
  greedy_token_match.prompts_matched = 8
  greedy_token_match.reference = hf_greedy
  greedy_token_match.source = this block
  greedy_token_match_compiled.first_divergence = [74, 44, None, None, None, None, None, None, None, None]
  greedy_token_match_compiled.n_prompts = 10
  greedy_token_match_compiled.prompts_matched = 8
  greedy_token_match_compiled.reference = hf_greedy_compiled
  greedy_token_match_compiled.source = this block
[hf_greedy]
  workload_tokens_per_second = 25.270848472299207
  tokens_per_second_median = 25.2248646317596
  tokens_per_second_p25 = 25.10785844955806
  tokens_per_second_p75 = 25.51860935304506
  tokens_per_second_mean = 25.274245665039576
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g4_median = 0.4319078365698047
  speedup_vs_arm_composite-g6-f2_median = 0.3820552364402474
  speedup_vs_hf_greedy_compiled_median = 0.48610484754403316
  forward_counts.n_generations = 10
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1280
  forward_counts.target_forwards_per_emitted_token = 1.0
  forward_counts.tokens_emitted = 1280
  host_syncs.scope = not probed for HuggingFace baselines in the phase6 runner
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[hf_greedy_compiled]
  workload_tokens_per_second = 51.99135226846701
  tokens_per_second_median = 51.978255671224986
  tokens_per_second_p25 = 51.93331754532797
  tokens_per_second_p75 = 52.0684287540002
  tokens_per_second_mean = 51.991523912544814
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g4_median = 0.8960673077860575
  speedup_vs_arm_composite-g6-f2_median = 0.7956776433424948
  speedup_vs_hf_greedy_median = 2.057171012403326
  forward_counts.n_generations = 10
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1280
  forward_counts.target_forwards_per_emitted_token = 1.0
  forward_counts.tokens_emitted = 1280
  host_syncs.scope = not probed for HuggingFace baselines in the phase6 runner
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
  greedy_token_match.first_divergence = [None, None, None, None, None, None, None, None, None, None]
  greedy_token_match.n_prompts = 10
  greedy_token_match.prompts_matched = 10
  greedy_token_match.reference = hf_greedy
  greedy_token_match.source = this block

==============================================================================
target=Qwen/Qwen2.5-1.5B-Instruct  drafter=Qwen/Qwen2.5-0.5B-Instruct  dataset=mt_bench  mode=sample T=0.7
commit=fdd3be9  n_samples=10
------------------------------------------------------------------------------
[arm_composite-g4]
  workload_tokens_per_second = 58.27963130033713
  tokens_per_second_median = 58.119863994220665
  tokens_per_second_p25 = 53.88868549663778
  tokens_per_second_p75 = 64.63719254841506
  tokens_per_second_mean = 59.23523668262179
  acceptance_rate_mean = 0.353721257472155
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g6-f2_median = 0.9512209420988531
  speedup_vs_hf_sample_compiled_median = 1.1274404156580013
  forward_counts.emitted_tokens_per_step = 2.388059701492537
  forward_counts.n_generations = 10
  forward_counts.n_steps = 536
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 546
  forward_counts.target_forwards_per_emitted_token = 0.4265625
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.010673672804746518
  step_profile.bookkeeping_ms_mean = 0.4366273432136027
  step_profile.bookkeeping_ms_total = 234.03225596249104
  step_profile.draft_fraction = 0.5028254411057087
  step_profile.draft_ms_mean = 20.569052515134437
  step_profile.draft_ms_total = 11025.012148112059
  step_profile.n_steps = 536.0
  step_profile.sample_fraction = 0.014371110187456715
  step_profile.sample_ms_mean = 0.5878782097750636
  step_profile.sample_ms_total = 315.10272043943405
  step_profile.step_ms_mean = 40.906944704116945
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.388059701492537
  step_profile.verify_fraction = 0.4721297759020881
  step_profile.verify_ms_mean = 19.313386635993844
  step_profile.verify_ms_total = 10351.9752368927
  host_syncs.by_kind = {'cuda': 56}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 52
  host_syncs.per_step = 1.0769230769230769
  host_syncs.prompt_id = 81
  host_syncs.total = 56
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[arm_composite-g6-f2]
  workload_tokens_per_second = 64.97537113354078
  tokens_per_second_median = 63.66822145393372
  tokens_per_second_p25 = 59.70671338712481
  tokens_per_second_p75 = 69.59687770256593
  tokens_per_second_mean = 66.66711096014413
  acceptance_rate_mean = 0.29846675268184264
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g4_median = 1.051322093574076
  speedup_vs_hf_sample_compiled_median = 1.2354066654768245
  forward_counts.emitted_tokens_per_step = 2.071197411003236
  forward_counts.n_generations = 10
  forward_counts.n_steps = 618
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 628
  forward_counts.target_forwards_per_emitted_token = 0.490625
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.011465275136311341
  step_profile.bookkeeping_ms_mean = 0.3647674820762622
  step_profile.bookkeeping_ms_total = 225.42630392313004
  step_profile.draft_fraction = 0.36293137491800825
  step_profile.draft_ms_mean = 11.546653893725
  step_profile.draft_ms_total = 7135.83210632205
  step_profile.n_steps = 618.0
  step_profile.sample_fraction = 0.018489742815790456
  step_profile.sample_ms_mean = 0.5882507703450891
  step_profile.sample_ms_total = 363.5389760732651
  step_profile.step_ms_mean = 31.814978510286043
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.071197411003236
  step_profile.verify_fraction = 0.6071136071298899
  step_profile.verify_ms_mean = 19.315306364139694
  step_profile.verify_ms_total = 11936.85933303833
  host_syncs.by_kind = {'cuda': 72}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 68
  host_syncs.per_step = 1.0588235294117647
  host_syncs.prompt_id = 81
  host_syncs.total = 72
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[hf_sample_compiled]
  workload_tokens_per_second = 51.53563279267184
  tokens_per_second_median = 51.5501509943132
  tokens_per_second_p25 = 51.52132209604921
  tokens_per_second_p75 = 51.66465190122399
  tokens_per_second_mean = 51.536223295766014
  gpu_memory_gb = 4.282809257507324
  speedup_vs_arm_composite-g4_median = 0.8882761298597817
  speedup_vs_arm_composite-g6-f2_median = 0.8095303966596755
  forward_counts.n_generations = 10
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1280
  forward_counts.target_forwards_per_emitted_token = 1.0
  forward_counts.tokens_emitted = 1280
  host_syncs.scope = not probed for HuggingFace baselines in the phase6 runner
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile

