==============================================================================
target=Qwen/Qwen2.5-1.5B-Instruct  drafter=Qwen/Qwen2.5-0.5B-Instruct  dataset=humaneval  mode=greedy
commit=36134c0  n_samples=10
------------------------------------------------------------------------------
[flashspec_ucb]
  tokens_per_second_median = 26.23852723249483
  tokens_per_second_p25 = 24.63765887975667
  tokens_per_second_p75 = 27.162960607917213
  tokens_per_second_mean = 25.728897698873812
  acceptance_rate_mean = 0.7748324561339393
  gpu_memory_gb = 3.831118583679199
  speedup_vs_hf_assisted_mean = 1.0727548392153887
  speedup_vs_hf_assisted_median = 1.0601084433981502
  speedup_vs_hf_assisted_std = 0.07760942137070022
  speedup_vs_hf_greedy_mean = 0.9973799925317689
  speedup_vs_hf_greedy_median = 1.0229054923869696
  speedup_vs_hf_greedy_std = 0.07468506981515628
  forward_counts.drafter_forwards = 1230
  forward_counts.drafter_forwards_per_step = 4.0
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 4.19672131147541
  forward_counts.n_generations = 10
  forward_counts.n_steps = 305
  forward_counts.target_forwards = 315
  forward_counts.target_forwards_per_emitted_token = 0.24609375
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.0017677451622977649
  step_profile.bookkeeping_ms_mean = 0.2899338489184614
  step_profile.bookkeeping_ms_total = 88.42982392013073
  step_profile.draft_fraction = 0.7706923687375018
  step_profile.draft_ms_mean = 126.4038559210105
  step_profile.draft_ms_total = 38553.1760559082
  step_profile.n_steps = 305.0
  step_profile.sample_fraction = 0.0007640269379620193
  step_profile.sample_ms_mean = 0.1253106361285585
  step_profile.sample_ms_total = 38.21974401921034
  step_profile.step_ms_mean = 164.01337427030333
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 4.19672131147541
  step_profile.verify_fraction = 0.2267758591622384
  step_profile.verify_ms_mean = 37.19427386424581
  step_profile.verify_ms_total = 11344.25352859497
  host_syncs.by_kind = {'cuda': 38}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 36
  host_syncs.per_step = 1.0555555555555556
  host_syncs.per_target_forward = 1.0555555555555556
  host_syncs.prompt_id = HumanEval/23
  host_syncs.target_forwards_after_prefill = 36
  host_syncs.total = 38
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 145
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = the untimed probe generation on the first prompt
  nonfinite_logits.target_entries = 0
  nonfinite_logits.target_forwards_checked = 37
  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
[hf_assisted]
  tokens_per_second_median = 24.67250490527404
  tokens_per_second_p25 = 22.646022509962553
  tokens_per_second_p75 = 25.612035488652108
  tokens_per_second_mean = 24.10582075018854
  gpu_memory_gb = 3.9205241203308105
  forward_counts.drafter_forwards = 1385
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 7.231638418079096
  forward_counts.n_generations = 10
  forward_counts.target_forwards = 187
  forward_counts.target_forwards_per_emitted_token = 0.14609375
  forward_counts.tokens_emitted = 1280
  host_syncs.by_kind = {'cuda': 617}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.per_target_forward = 24.68
  host_syncs.prompt_id = HumanEval/23
  host_syncs.target_forwards_after_prefill = 25
  host_syncs.total = 617
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 153
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = the untimed probe generation on the first prompt
  nonfinite_logits.target_entries = 0
  nonfinite_logits.target_forwards_checked = 26
  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
[hf_greedy]
  tokens_per_second_median = 25.747536265737693
  tokens_per_second_p25 = 25.667860736052027
  tokens_per_second_p75 = 25.94495327957157
  tokens_per_second_mean = 25.79951422204662
  gpu_memory_gb = 3.8375372886657715
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 1.0078740157480315
  forward_counts.n_generations = 10
  forward_counts.target_forwards = 1280
  forward_counts.target_forwards_per_emitted_token = 1.0
  forward_counts.tokens_emitted = 1280
  host_syncs.by_kind = {'cuda': 130}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.per_target_forward = 1.0236220472440944
  host_syncs.prompt_id = HumanEval/23
  host_syncs.target_forwards_after_prefill = 127
  host_syncs.total = 130
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 0
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = the untimed probe generation on the first prompt
  nonfinite_logits.target_entries = 0
  nonfinite_logits.target_forwards_checked = 128

==============================================================================
target=Qwen/Qwen2.5-1.5B-Instruct  drafter=Qwen/Qwen2.5-0.5B-Instruct  dataset=humaneval  mode=sample T=0.7
commit=36134c0  n_samples=10
------------------------------------------------------------------------------
[flashspec_ucb]
  tokens_per_second_median = 22.994806388294748
  tokens_per_second_p25 = 21.758612773754045
  tokens_per_second_p75 = 25.17813966076342
  tokens_per_second_mean = 23.24264491488757
  acceptance_rate_mean = 0.7036410816413052
  gpu_memory_gb = 3.8405613899230957
  speedup_vs_hf_assisted_mean = 1.16744261671329
  speedup_vs_hf_assisted_median = 1.1413036574151052
  speedup_vs_hf_assisted_std = 0.21279009976448587
  speedup_vs_hf_sample_mean = 0.9179432212998455
  speedup_vs_hf_sample_median = 0.9143376283741449
  speedup_vs_hf_sample_std = 0.08702499002024035
  forward_counts.drafter_forwards = 1346
  forward_counts.drafter_forwards_per_step = 4.0
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 3.8323353293413174
  forward_counts.n_generations = 10
  forward_counts.n_steps = 334
  forward_counts.target_forwards = 344
  forward_counts.target_forwards_per_emitted_token = 0.26875
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.002100541471240699
  step_profile.bookkeeping_ms_mean = 0.3492640007488028
  step_profile.bookkeeping_ms_total = 116.65417625010014
  step_profile.draft_fraction = 0.7715404287822512
  step_profile.draft_ms_mean = 128.28658733253707
  step_profile.draft_ms_total = 42847.72016906738
  step_profile.n_steps = 334.0
  step_profile.sample_fraction = 0.002229785413964431
  step_profile.sample_ms_mean = 0.3707538199817349
  step_profile.sample_ms_total = 123.83177587389946
  step_profile.step_ms_mean = 166.27331834705825
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 3.8323353293413174
  step_profile.verify_fraction = 0.2241292443325437
  step_profile.verify_ms_mean = 37.26671319379064
  step_profile.verify_ms_total = 12447.082206726074
  host_syncs.by_kind = {'cuda': 37}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 35
  host_syncs.per_step = 1.0571428571428572
  host_syncs.per_target_forward = 1.0571428571428572
  host_syncs.prompt_id = HumanEval/23
  host_syncs.target_forwards_after_prefill = 35
  host_syncs.total = 37
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 141
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = the untimed probe generation on the first prompt
  nonfinite_logits.target_entries = 0
  nonfinite_logits.target_forwards_checked = 36
[hf_assisted]
  tokens_per_second_median = 19.75492333430464
  tokens_per_second_p25 = 17.68724539958827
  tokens_per_second_p75 = 22.51555521260074
  tokens_per_second_mean = 20.475610667931075
  gpu_memory_gb = 3.9224438667297363
  forward_counts.drafter_forwards = 1633
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 6.2745098039215685
  forward_counts.n_generations = 10
  forward_counts.target_forwards = 214
  forward_counts.target_forwards_per_emitted_token = 0.1671875
  forward_counts.tokens_emitted = 1280
  host_syncs.by_kind = {'cuda': 736}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.per_target_forward = 25.379310344827587
  host_syncs.prompt_id = HumanEval/23
  host_syncs.target_forwards_after_prefill = 29
  host_syncs.total = 736
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 136
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = the untimed probe generation on the first prompt
  nonfinite_logits.target_entries = 0
  nonfinite_logits.target_forwards_checked = 30
[hf_sample]
  tokens_per_second_median = 25.401692277734487
  tokens_per_second_p25 = 25.188006318909956
  tokens_per_second_p75 = 25.52414674099085
  tokens_per_second_mean = 25.326890820725243
  gpu_memory_gb = 3.840813636779785
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 1.0078740157480315
  forward_counts.n_generations = 10
  forward_counts.target_forwards = 1280
  forward_counts.target_forwards_per_emitted_token = 1.0
  forward_counts.tokens_emitted = 1280
  host_syncs.by_kind = {'cuda': 130}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.per_target_forward = 1.0236220472440944
  host_syncs.prompt_id = HumanEval/23
  host_syncs.target_forwards_after_prefill = 127
  host_syncs.total = 130
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 0
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = the untimed probe generation on the first prompt
  nonfinite_logits.target_entries = 0
  nonfinite_logits.target_forwards_checked = 128

==============================================================================
target=Qwen/Qwen2.5-1.5B-Instruct  drafter=Qwen/Qwen2.5-0.5B-Instruct  dataset=mt_bench  mode=greedy
commit=36134c0  n_samples=10
------------------------------------------------------------------------------
[flashspec_ucb]
  tokens_per_second_median = 19.263895987077582
  tokens_per_second_p25 = 17.876720942167275
  tokens_per_second_p75 = 19.932220650397593
  tokens_per_second_mean = 19.76241945847483
  acceptance_rate_mean = 0.5308389674229258
  gpu_memory_gb = 3.8356237411499023
  speedup_vs_hf_assisted_mean = 1.105130151690302
  speedup_vs_hf_assisted_median = 1.073667645549547
  speedup_vs_hf_assisted_std = 0.23292512860744252
  speedup_vs_hf_greedy_mean = 0.7726392992912692
  speedup_vs_hf_greedy_median = 0.7528549865839097
  speedup_vs_hf_greedy_std = 0.12364162411755802
  forward_counts.drafter_forwards = 1622
  forward_counts.drafter_forwards_per_step = 4.0
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 3.176178660049628
  forward_counts.n_generations = 10
  forward_counts.n_steps = 403
  forward_counts.target_forwards = 413
  forward_counts.target_forwards_per_emitted_token = 0.32265625
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.0025918279974730144
  step_profile.bookkeeping_ms_mean = 0.4269040603347807
  step_profile.bookkeeping_ms_total = 172.0423363149166
  step_profile.draft_fraction = 0.7700769859268011
  step_profile.draft_ms_mean = 126.84058987828105
  step_profile.draft_ms_total = 51116.757720947266
  step_profile.n_steps = 403.0
  step_profile.sample_fraction = 0.0007701899485053786
  step_profile.sample_ms_mean = 0.12685919612202395
  step_profile.sample_ms_total = 51.124256037175655
  step_profile.step_ms_mean = 164.71157065631067
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 3.176178660049628
  step_profile.verify_fraction = 0.22656099612722058
  step_profile.verify_ms_mean = 37.31721752157282
  step_profile.verify_ms_total = 15038.838661193848
  host_syncs.by_kind = {'cuda': 47}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 44
  host_syncs.per_step = 1.0681818181818181
  host_syncs.per_target_forward = 1.0681818181818181
  host_syncs.prompt_id = 81
  host_syncs.target_forwards_after_prefill = 44
  host_syncs.total = 47
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 177
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = the untimed probe generation on the first prompt
  nonfinite_logits.target_entries = 0
  nonfinite_logits.target_forwards_checked = 45
  greedy_token_match.first_divergence = [None, None, None, None, None, None, None, None, 97, None]
  greedy_token_match.n_prompts = 10
  greedy_token_match.prompts_matched = 9
  greedy_token_match.reference = hf_greedy
[hf_assisted]
  tokens_per_second_median = 19.857969448236993
  tokens_per_second_p25 = 15.308813592486892
  tokens_per_second_p75 = 22.27020365580778
  tokens_per_second_mean = 18.666133868443797
  gpu_memory_gb = 3.916429042816162
  forward_counts.drafter_forwards = 1803
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 3.8905775075987843
  forward_counts.n_generations = 10
  forward_counts.target_forwards = 339
  forward_counts.target_forwards_per_emitted_token = 0.26484375
  forward_counts.tokens_emitted = 1280
  host_syncs.by_kind = {'cuda': 1184}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.per_target_forward = 28.878048780487806
  host_syncs.prompt_id = 81
  host_syncs.target_forwards_after_prefill = 41
  host_syncs.total = 1184
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 310
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = the untimed probe generation on the first prompt
  nonfinite_logits.target_entries = 0
  nonfinite_logits.target_forwards_checked = 42
  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
[hf_greedy]
  tokens_per_second_median = 25.487991963425753
  tokens_per_second_p25 = 25.439671247202842
  tokens_per_second_p75 = 25.65189575863387
  tokens_per_second_mean = 25.543520835402468
  gpu_memory_gb = 3.845881938934326
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 1.0078740157480315
  forward_counts.n_generations = 10
  forward_counts.target_forwards = 1280
  forward_counts.target_forwards_per_emitted_token = 1.0
  forward_counts.tokens_emitted = 1280
  host_syncs.by_kind = {'cuda': 130}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.per_target_forward = 1.0236220472440944
  host_syncs.prompt_id = 81
  host_syncs.target_forwards_after_prefill = 127
  host_syncs.total = 130
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 0
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = the untimed probe generation on the first prompt
  nonfinite_logits.target_entries = 0
  nonfinite_logits.target_forwards_checked = 128

==============================================================================
target=Qwen/Qwen2.5-1.5B-Instruct  drafter=Qwen/Qwen2.5-0.5B-Instruct  dataset=mt_bench  mode=sample T=0.7
commit=36134c0  n_samples=10
------------------------------------------------------------------------------
[flashspec_ucb]
  tokens_per_second_median = 19.856912897628682
  tokens_per_second_p25 = 18.059678676289927
  tokens_per_second_p75 = 21.612539675023445
  tokens_per_second_mean = 20.213364580924342
  acceptance_rate_mean = 0.5357908259547473
  gpu_memory_gb = 3.8409838676452637
  speedup_vs_hf_assisted_mean = 1.4976773693251928
  speedup_vs_hf_assisted_median = 1.1933850262734405
  speedup_vs_hf_assisted_std = 0.8122564404927818
  speedup_vs_hf_sample_mean = 0.7876554123365764
  speedup_vs_hf_sample_median = 0.7713409325667324
  speedup_vs_hf_sample_std = 0.1077078432640804
  forward_counts.drafter_forwards = 1566
  forward_counts.drafter_forwards_per_step = 4.0
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 3.290488431876607
  forward_counts.n_generations = 10
  forward_counts.n_steps = 389
  forward_counts.target_forwards = 399
  forward_counts.target_forwards_per_emitted_token = 0.31171875
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.0024941167364164857
  step_profile.bookkeeping_ms_mean = 0.413076153451189
  step_profile.bookkeeping_ms_total = 160.6866236925125
  step_profile.draft_fraction = 0.7718171620907522
  step_profile.draft_ms_mean = 127.82852535689092
  step_profile.draft_ms_total = 49725.29636383057
  step_profile.n_steps = 389.0
  step_profile.sample_fraction = 0.002202673364079771
  step_profile.sample_ms_mean = 0.36480723907524637
  step_profile.sample_ms_total = 141.91001600027084
  step_profile.step_ms_mean = 165.62021633545967
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 3.290488431876607
  step_profile.verify_fraction = 0.22348604780875153
  step_profile.verify_ms_mean = 37.01380758604231
  step_profile.verify_ms_total = 14398.371150970459
  host_syncs.by_kind = {'cuda': 48}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 46
  host_syncs.per_step = 1.0434782608695652
  host_syncs.per_target_forward = 1.0434782608695652
  host_syncs.prompt_id = 81
  host_syncs.target_forwards_after_prefill = 46
  host_syncs.total = 48
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 185
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = the untimed probe generation on the first prompt
  nonfinite_logits.target_entries = 0
  nonfinite_logits.target_forwards_checked = 47
[hf_assisted]
  tokens_per_second_median = 14.715778555192315
  tokens_per_second_p25 = 12.989302254851275
  tokens_per_second_p75 = 19.21646429804001
  tokens_per_second_mean = 15.848071423813462
  gpu_memory_gb = 3.9214768409729004
  forward_counts.drafter_forwards = 2423
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 3.78698224852071
  forward_counts.n_generations = 10
  forward_counts.target_forwards = 348
  forward_counts.target_forwards_per_emitted_token = 0.271875
  forward_counts.tokens_emitted = 1280
  host_syncs.by_kind = {'cuda': 1017}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.per_target_forward = 21.638297872340427
  host_syncs.prompt_id = 81
  host_syncs.target_forwards_after_prefill = 47
  host_syncs.total = 1017
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 160
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = the untimed probe generation on the first prompt
  nonfinite_logits.target_entries = 0
  nonfinite_logits.target_forwards_checked = 48
[hf_sample]
  tokens_per_second_median = 25.708033881186918
  tokens_per_second_p25 = 25.52976546475353
  tokens_per_second_p75 = 25.845993306023857
  tokens_per_second_mean = 25.65710328209583
  gpu_memory_gb = 3.8480911254882812
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 1.0078740157480315
  forward_counts.n_generations = 10
  forward_counts.target_forwards = 1280
  forward_counts.target_forwards_per_emitted_token = 1.0
  forward_counts.tokens_emitted = 1280
  host_syncs.by_kind = {'cuda': 130}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.per_target_forward = 1.0236220472440944
  host_syncs.prompt_id = 81
  host_syncs.target_forwards_after_prefill = 127
  host_syncs.total = 130
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 0
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = the untimed probe generation on the first prompt
  nonfinite_logits.target_entries = 0
  nonfinite_logits.target_forwards_checked = 128

==============================================================================
target=TinyLlama/TinyLlama-1.1B-Chat-v1.0  drafter=Felladrin/Llama-160M-Chat-v1  dataset=humaneval  mode=greedy
commit=36134c0  n_samples=10
------------------------------------------------------------------------------
[flashspec_ucb]
  tokens_per_second_median = 31.250871186065634
  tokens_per_second_p25 = 29.67332081324305
  tokens_per_second_p75 = 31.730403496731114
  tokens_per_second_mean = 30.70618152609352
  acceptance_rate_mean = 0.3428825495568735
  gpu_memory_gb = 3.3315539360046387
  speedup_vs_hf_assisted_mean = 1.4125325297734626
  speedup_vs_hf_assisted_median = 1.3627982299701538
  speedup_vs_hf_assisted_std = 0.27062147055215363
  speedup_vs_hf_greedy_mean = 0.9105102756558912
  speedup_vs_hf_greedy_median = 0.920824455046579
  speedup_vs_hf_greedy_std = 0.08475385335040325
  forward_counts.drafter_forwards = 2058
  forward_counts.drafter_forwards_per_step = 4.0
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 2.5
  forward_counts.n_generations = 10
  forward_counts.n_steps = 512
  forward_counts.target_forwards = 522
  forward_counts.target_forwards_per_emitted_token = 0.4078125
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.0046463518595364324
  step_profile.bookkeeping_ms_mean = 0.38098749960772693
  step_profile.bookkeeping_ms_total = 195.0655997991562
  step_profile.draft_fraction = 0.6380801436644502
  step_profile.draft_ms_mean = 52.32073803991079
  step_profile.draft_ms_total = 26788.217876434326
  step_profile.n_steps = 512.0
  step_profile.sample_fraction = 0.0016820581915732947
  step_profile.sample_ms_mean = 0.13792393774201628
  step_profile.sample_ms_total = 70.61705612391233
  step_profile.step_ms_mean = 81.99712615947647
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.5
  step_profile.verify_fraction = 0.35559144628444006
  step_profile.verify_ms_mean = 29.15747668221593
  step_profile.verify_ms_total = 14928.628061294556
  host_syncs.by_kind = {'cuda': 66}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 64
  host_syncs.per_step = 1.03125
  host_syncs.per_target_forward = 1.03125
  host_syncs.prompt_id = HumanEval/23
  host_syncs.target_forwards_after_prefill = 64
  host_syncs.total = 66
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 257
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = the untimed probe generation on the first prompt
  nonfinite_logits.target_entries = 0
  nonfinite_logits.target_forwards_checked = 65
  greedy_token_match.first_divergence = [None, None, None, None, None, None, None, 61, None, None]
  greedy_token_match.n_prompts = 10
  greedy_token_match.prompts_matched = 9
  greedy_token_match.reference = hf_greedy
[hf_assisted]
  tokens_per_second_median = 22.801485829108483
  tokens_per_second_p25 = 20.90997221766804
  tokens_per_second_p75 = 23.29082343261765
  tokens_per_second_mean = 22.38428621036846
  gpu_memory_gb = 3.357449531555176
  forward_counts.drafter_forwards = 3088
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 2.644628099173554
  forward_counts.n_generations = 10
  forward_counts.target_forwards = 494
  forward_counts.target_forwards_per_emitted_token = 0.3859375
  forward_counts.tokens_emitted = 1280
  host_syncs.by_kind = {'cuda': 1249}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.per_target_forward = 19.215384615384615
  host_syncs.prompt_id = HumanEval/23
  host_syncs.target_forwards_after_prefill = 65
  host_syncs.total = 1249
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 285
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = the untimed probe generation on the first prompt
  nonfinite_logits.target_entries = 0
  nonfinite_logits.target_forwards_checked = 66
  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
[hf_greedy]
  tokens_per_second_median = 33.80787372488116
  tokens_per_second_p25 = 33.588002304518476
  tokens_per_second_p75 = 34.12714768557471
  tokens_per_second_mean = 33.75510455641334
  gpu_memory_gb = 3.339426040649414
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 1.0078740157480315
  forward_counts.n_generations = 10
  forward_counts.target_forwards = 1280
  forward_counts.target_forwards_per_emitted_token = 1.0
  forward_counts.tokens_emitted = 1280
  host_syncs.by_kind = {'cuda': 130}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.per_target_forward = 1.0236220472440944
  host_syncs.prompt_id = HumanEval/23
  host_syncs.target_forwards_after_prefill = 127
  host_syncs.total = 130
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 0
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = the untimed probe generation on the first prompt
  nonfinite_logits.target_entries = 0
  nonfinite_logits.target_forwards_checked = 128

==============================================================================
target=TinyLlama/TinyLlama-1.1B-Chat-v1.0  drafter=Felladrin/Llama-160M-Chat-v1  dataset=humaneval  mode=sample T=0.7
commit=36134c0  n_samples=10
------------------------------------------------------------------------------
[flashspec_ucb]
  tokens_per_second_median = 29.79554888577421
  tokens_per_second_p25 = 28.57163363897569
  tokens_per_second_p75 = 30.722737408248435
  tokens_per_second_mean = 29.421510727038346
  acceptance_rate_mean = 0.37814445408204117
  gpu_memory_gb = 3.3389954566955566
  speedup_vs_hf_assisted_mean = 1.7173833778410283
  speedup_vs_hf_assisted_median = 1.5265792438087513
  speedup_vs_hf_assisted_std = 0.5660489163262552
  speedup_vs_hf_sample_mean = 0.8729098015834653
  speedup_vs_hf_sample_median = 0.8781563928644487
  speedup_vs_hf_sample_std = 0.07445042674784437
  forward_counts.drafter_forwards = 2098
  forward_counts.drafter_forwards_per_step = 4.0
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 2.4521072796934864
  forward_counts.n_generations = 10
  forward_counts.n_steps = 522
  forward_counts.target_forwards = 532
  forward_counts.target_forwards_per_emitted_token = 0.415625
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.004620182475910055
  step_profile.bookkeeping_ms_mean = 0.38737802301672686
  step_profile.bookkeeping_ms_total = 202.2113280147314
  step_profile.draft_fraction = 0.6358721630932631
  step_profile.draft_ms_mean = 53.314539569547804
  step_profile.draft_ms_total = 27830.189655303955
  step_profile.n_steps = 522.0
  step_profile.sample_fraction = 0.009565242730701707
  step_profile.sample_ms_mean = 0.8019953406633545
  step_profile.sample_ms_total = 418.64156782627106
  step_profile.step_ms_mean = 83.84474531829471
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.4521072796934864
  step_profile.verify_fraction = 0.3499424117001251
  step_profile.verify_ms_mean = 29.340832385066825
  step_profile.verify_ms_total = 15315.914505004883
  host_syncs.by_kind = {'cuda': 56}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 54
  host_syncs.per_step = 1.037037037037037
  host_syncs.per_target_forward = 1.037037037037037
  host_syncs.prompt_id = HumanEval/23
  host_syncs.target_forwards_after_prefill = 54
  host_syncs.total = 56
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 217
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = the untimed probe generation on the first prompt
  nonfinite_logits.target_entries = 0
  nonfinite_logits.target_forwards_checked = 55
[hf_assisted]
  tokens_per_second_median = 17.487019025333403
  tokens_per_second_p25 = 13.7530429740147
  tokens_per_second_p75 = 23.568704422712038
  tokens_per_second_mean = 18.82751246626244
  gpu_memory_gb = 3.3656387329101562
  forward_counts.drafter_forwards = 4085
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 2.633744855967078
  forward_counts.n_generations = 10
  forward_counts.target_forwards = 496
  forward_counts.target_forwards_per_emitted_token = 0.3875
  forward_counts.tokens_emitted = 1280
  host_syncs.by_kind = {'cuda': 1427}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.per_target_forward = 29.122448979591837
  host_syncs.prompt_id = HumanEval/23
  host_syncs.target_forwards_after_prefill = 49
  host_syncs.total = 1427
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 285
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = the untimed probe generation on the first prompt
  nonfinite_logits.target_entries = 0
  nonfinite_logits.target_forwards_checked = 50
[hf_sample]
  tokens_per_second_median = 33.78956769887036
  tokens_per_second_p25 = 33.30730619276562
  tokens_per_second_p75 = 34.0416340840893
  tokens_per_second_mean = 33.73008909850701
  gpu_memory_gb = 3.3462777137756348
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 1.0078740157480315
  forward_counts.n_generations = 10
  forward_counts.target_forwards = 1280
  forward_counts.target_forwards_per_emitted_token = 1.0
  forward_counts.tokens_emitted = 1280
  host_syncs.by_kind = {'cuda': 130}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.per_target_forward = 1.0236220472440944
  host_syncs.prompt_id = HumanEval/23
  host_syncs.target_forwards_after_prefill = 127
  host_syncs.total = 130
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 0
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = the untimed probe generation on the first prompt
  nonfinite_logits.target_entries = 0
  nonfinite_logits.target_forwards_checked = 128

==============================================================================
target=TinyLlama/TinyLlama-1.1B-Chat-v1.0  drafter=Felladrin/Llama-160M-Chat-v1  dataset=mt_bench  mode=greedy
commit=36134c0  n_samples=10
------------------------------------------------------------------------------
[flashspec_ucb]
  tokens_per_second_median = 30.210543463312913
  tokens_per_second_p25 = 26.763347847207847
  tokens_per_second_p75 = 31.938117072957773
  tokens_per_second_mean = 29.791384453537624
  acceptance_rate_mean = 0.3849257507706609
  gpu_memory_gb = 3.339047431945801
  speedup_vs_hf_assisted_mean = 1.2764131727949497
  speedup_vs_hf_assisted_median = 1.2096862194202025
  speedup_vs_hf_assisted_std = 0.30763088989920623
  speedup_vs_hf_greedy_mean = 0.8806707816983652
  speedup_vs_hf_greedy_median = 0.9063770613006314
  speedup_vs_hf_greedy_std = 0.08803359161877829
  forward_counts.drafter_forwards = 2122
  forward_counts.drafter_forwards_per_step = 4.0
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 2.4242424242424243
  forward_counts.n_generations = 10
  forward_counts.n_steps = 528
  forward_counts.target_forwards = 538
  forward_counts.target_forwards_per_emitted_token = 0.4203125
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.004899589128484578
  step_profile.bookkeeping_ms_mean = 0.4025797573002902
  step_profile.bookkeeping_ms_total = 212.56211185455322
  step_profile.draft_fraction = 0.6389217178062622
  step_profile.draft_ms_mean = 52.497657118421614
  step_profile.draft_ms_total = 27718.76295852661
  step_profile.n_steps = 528.0
  step_profile.sample_fraction = 0.001804451513599971
  step_profile.sample_ms_mean = 0.1482646061446179
  step_profile.sample_ms_total = 78.28371204435825
  step_profile.step_ms_mean = 82.1660238732725
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.4242424242424243
  step_profile.verify_fraction = 0.35437424155165326
  step_profile.verify_ms_mean = 29.117522391405974
  step_profile.verify_ms_total = 15374.051822662354
  host_syncs.by_kind = {'cuda': 50}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 48
  host_syncs.per_step = 1.0416666666666667
  host_syncs.per_target_forward = 1.0416666666666667
  host_syncs.prompt_id = 81
  host_syncs.target_forwards_after_prefill = 48
  host_syncs.total = 50
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 193
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = the untimed probe generation on the first prompt
  nonfinite_logits.target_entries = 0
  nonfinite_logits.target_forwards_checked = 49
  greedy_token_match.first_divergence = [None, None, None, None, None, 0, None, None, None, None]
  greedy_token_match.n_prompts = 10
  greedy_token_match.prompts_matched = 9
  greedy_token_match.reference = hf_greedy
[hf_assisted]
  tokens_per_second_median = 24.551502847019417
  tokens_per_second_p25 = 20.15396423641451
  tokens_per_second_p75 = 28.619996148281555
  tokens_per_second_mean = 24.388425831710688
  gpu_memory_gb = 3.3552961349487305
  forward_counts.drafter_forwards = 2676
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 2.549800796812749
  forward_counts.n_generations = 10
  forward_counts.target_forwards = 512
  forward_counts.target_forwards_per_emitted_token = 0.4
  forward_counts.tokens_emitted = 1280
  host_syncs.by_kind = {'cuda': 968}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.per_target_forward = 21.043478260869566
  host_syncs.prompt_id = 81
  host_syncs.target_forwards_after_prefill = 46
  host_syncs.total = 968
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 228
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = the untimed probe generation on the first prompt
  nonfinite_logits.target_entries = 0
  nonfinite_logits.target_forwards_checked = 47
  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
[hf_greedy]
  tokens_per_second_median = 33.96586860489057
  tokens_per_second_p25 = 33.82290284657868
  tokens_per_second_p75 = 34.12758268923611
  tokens_per_second_mean = 33.830355383115716
  gpu_memory_gb = 3.3504180908203125
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 1.0078740157480315
  forward_counts.n_generations = 10
  forward_counts.target_forwards = 1280
  forward_counts.target_forwards_per_emitted_token = 1.0
  forward_counts.tokens_emitted = 1280
  host_syncs.by_kind = {'cuda': 130}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.per_target_forward = 1.0236220472440944
  host_syncs.prompt_id = 81
  host_syncs.target_forwards_after_prefill = 127
  host_syncs.total = 130
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 0
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = the untimed probe generation on the first prompt
  nonfinite_logits.target_entries = 0
  nonfinite_logits.target_forwards_checked = 128

==============================================================================
target=TinyLlama/TinyLlama-1.1B-Chat-v1.0  drafter=Felladrin/Llama-160M-Chat-v1  dataset=mt_bench  mode=sample T=0.7
commit=36134c0  n_samples=10
------------------------------------------------------------------------------
[flashspec_ucb]
  tokens_per_second_median = 27.690370534753658
  tokens_per_second_p25 = 26.354477514108684
  tokens_per_second_p75 = 31.13884479526491
  tokens_per_second_mean = 28.771765580434668
  acceptance_rate_mean = 0.35637230927831254
  gpu_memory_gb = 3.3453612327575684
  speedup_vs_hf_assisted_mean = 1.3328745241210602
  speedup_vs_hf_assisted_median = 1.2956752804023388
  speedup_vs_hf_assisted_std = 0.2891887401669252
  speedup_vs_hf_sample_mean = 0.8558642353829311
  speedup_vs_hf_sample_median = 0.8500740058459157
  speedup_vs_hf_sample_std = 0.0884066033095477
  forward_counts.drafter_forwards = 2106
  forward_counts.drafter_forwards_per_step = 4.0
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 2.4427480916030535
  forward_counts.n_generations = 10
  forward_counts.n_steps = 524
  forward_counts.target_forwards = 534
  forward_counts.target_forwards_per_emitted_token = 0.4171875
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.004790681528979338
  step_profile.bookkeeping_ms_mean = 0.41111108382012096
  step_profile.bookkeeping_ms_total = 215.4222079217434
  step_profile.draft_fraction = 0.636094785375766
  step_profile.draft_ms_mean = 54.586307824112986
  step_profile.draft_ms_total = 28603.225299835205
  step_profile.n_steps = 524.0
  step_profile.sample_fraction = 0.010435328358961597
  step_profile.sample_ms_mean = 0.8955049768431496
  step_profile.sample_ms_total = 469.2446078658104
  step_profile.step_ms_mean = 85.81473874505467
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.4427480916030535
  step_profile.verify_fraction = 0.348679204736293
  step_profile.verify_ms_mean = 29.921814860278413
  step_profile.verify_ms_total = 15679.030986785889
  host_syncs.by_kind = {'cuda': 55}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 53
  host_syncs.per_step = 1.0377358490566038
  host_syncs.per_target_forward = 1.0377358490566038
  host_syncs.prompt_id = 81
  host_syncs.target_forwards_after_prefill = 53
  host_syncs.total = 55
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 213
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = the untimed probe generation on the first prompt
  nonfinite_logits.target_entries = 0
  nonfinite_logits.target_forwards_checked = 54
[hf_assisted]
  tokens_per_second_median = 22.920757315705643
  tokens_per_second_p25 = 20.03353302426056
  tokens_per_second_p75 = 24.297143585330673
  tokens_per_second_mean = 22.41828973717221
  gpu_memory_gb = 3.361243724822998
  forward_counts.drafter_forwards = 2877
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 2.4758220502901356
  forward_counts.n_generations = 10
  forward_counts.target_forwards = 527
  forward_counts.target_forwards_per_emitted_token = 0.41171875
  forward_counts.tokens_emitted = 1280
  host_syncs.by_kind = {'cuda': 2085}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.per_target_forward = 46.333333333333336
  host_syncs.prompt_id = 81
  host_syncs.target_forwards_after_prefill = 45
  host_syncs.total = 2085
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 517
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = the untimed probe generation on the first prompt
  nonfinite_logits.target_entries = 0
  nonfinite_logits.target_forwards_checked = 46
[hf_sample]
  tokens_per_second_median = 33.636622481082696
  tokens_per_second_p25 = 33.08990594716871
  tokens_per_second_p75 = 34.08290521661548
  tokens_per_second_mean = 33.60892267012456
  gpu_memory_gb = 3.3561997413635254
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 1.0078740157480315
  forward_counts.n_generations = 10
  forward_counts.target_forwards = 1280
  forward_counts.target_forwards_per_emitted_token = 1.0
  forward_counts.tokens_emitted = 1280
  host_syncs.by_kind = {'cuda': 130}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.per_target_forward = 1.0236220472440944
  host_syncs.prompt_id = 81
  host_syncs.target_forwards_after_prefill = 127
  host_syncs.total = 130
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 0
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = the untimed probe generation on the first prompt
  nonfinite_logits.target_entries = 0
  nonfinite_logits.target_forwards_checked = 128

