==============================================================================
target=Qwen/Qwen2.5-1.5B-Instruct  drafter=Qwen/Qwen2.5-0.5B-Instruct  dataset=humaneval  mode=greedy
commit=3c43e08  n_samples=10
------------------------------------------------------------------------------
[flashspec_ucb]
  tokens_per_second_median = 35.695688670935155
  tokens_per_second_p25 = 29.562405679239028
  tokens_per_second_p75 = 38.56945102463008
  tokens_per_second_mean = 34.693976023298305
  acceptance_rate_mean = 0.3909112890710105
  gpu_memory_gb = 3.9214916229248047
  speedup_vs_flashspec_compiled_mean = 0.41954120051805266
  speedup_vs_flashspec_compiled_median = 0.4188055136119598
  speedup_vs_flashspec_compiled_std = 0.014263844879067605
  speedup_vs_hf_assisted_mean = 1.5306402428314296
  speedup_vs_hf_assisted_median = 1.5587746305864147
  speedup_vs_hf_assisted_std = 0.22114121642604315
  speedup_vs_hf_greedy_compiled_mean = 0.6828941972895749
  speedup_vs_hf_greedy_compiled_median = 0.7066183063738005
  speedup_vs_hf_greedy_compiled_std = 0.09748214401861131
  speedup_vs_hf_greedy_mean = 1.4193835501504974
  speedup_vs_hf_greedy_median = 1.451343369376347
  speedup_vs_hf_greedy_std = 0.20333073301959725
  forward_counts.drafter_forwards = 490
  forward_counts.drafter_forwards_per_step = 0.9266409266409267
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 2.471042471042471
  forward_counts.n_generations = 10
  forward_counts.n_steps = 518
  forward_counts.source = measured with a forward hook
  forward_counts.target_forwards = 528
  forward_counts.target_forwards_per_emitted_token = 0.4125
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.004806087629717654
  step_profile.bookkeeping_ms_mean = 0.349341466950979
  step_profile.bookkeeping_ms_total = 180.95887988060713
  step_profile.draft_fraction = 0.4405833432950997
  step_profile.draft_ms_mean = 32.02480756055607
  step_profile.draft_ms_total = 16588.850316368043
  step_profile.n_steps = 518.0
  step_profile.sample_fraction = 0.006188729845484279
  step_profile.sample_ms_mean = 0.44984197737397374
  step_profile.sample_ms_total = 233.0181442797184
  step_profile.step_ms_mean = 72.68728618073533
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.471042471042471
  step_profile.verify_fraction = 0.5484218392296983
  step_profile.verify_ms_mean = 39.863295175854304
  step_profile.verify_ms_total = 20649.18690109253
  host_syncs.by_kind = {'cuda': 52}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 49
  host_syncs.per_step = 1.0612244897959184
  host_syncs.per_target_forward = 1.0612244897959184
  host_syncs.prompt_id = HumanEval/23
  host_syncs.target_forwards_after_prefill = 49
  host_syncs.total = 52
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 49
  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
  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
[flashspec_compiled]
  tokens_per_second_median = 84.15121158912345
  tokens_per_second_p25 = 72.74887834456933
  tokens_per_second_p75 = 89.15351141840875
  tokens_per_second_mean = 82.64249752375056
  acceptance_rate_mean = 0.3919555827630965
  gpu_memory_gb = 3.935239791870117
  speedup_vs_flashspec_ucb_mean = 2.3863124085599825
  speedup_vs_flashspec_ucb_median = 2.3887997196486594
  speedup_vs_flashspec_ucb_std = 0.08107965548071398
  speedup_vs_hf_assisted_mean = 3.652505968831382
  speedup_vs_hf_assisted_median = 3.6799192622210697
  speedup_vs_hf_assisted_std = 0.5408645301201851
  speedup_vs_hf_greedy_compiled_mean = 1.62676877269628
  speedup_vs_hf_greedy_compiled_median = 1.6703915213699534
  speedup_vs_hf_greedy_compiled_std = 0.22149501744928646
  speedup_vs_hf_greedy_mean = 3.3806395396919258
  speedup_vs_hf_greedy_median = 3.4316137147436416
  speedup_vs_hf_greedy_std = 0.45821943040021557
  forward_counts.drafter_forwards = 490
  forward_counts.drafter_forwards_per_step = 0.9284332688588007
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 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; drafter forwards from the per-step arm records
  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 = 0.012167873534829846
  step_profile.bookkeeping_ms_mean = 0.3703920777807844
  step_profile.bookkeeping_ms_total = 191.49270421266556
  step_profile.draft_fraction = 0.3362076086791934
  step_profile.draft_ms_mean = 10.234215073647766
  step_profile.draft_ms_total = 5291.089193075895
  step_profile.n_steps = 517.0
  step_profile.sample_fraction = 0.012955168864541169
  step_profile.sample_ms_mean = 0.3943574775003142
  step_profile.sample_ms_total = 203.88281586766243
  step_profile.step_ms_mean = 30.440164973818817
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.4758220502901356
  step_profile.verify_fraction = 0.6386693489214356
  step_profile.verify_ms_mean = 19.441200344889953
  step_profile.verify_ms_total = 10051.100578308105
  host_syncs.by_kind = {'cuda': 55}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 49
  host_syncs.per_step = 1.1224489795918366
  host_syncs.per_target_forward = 1.1224489795918366
  host_syncs.prompt_id = HumanEval/23
  host_syncs.target_forwards_after_prefill = 49
  host_syncs.target_forwards_source = derived: one per step / per decoded token
  host_syncs.total = 55
  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_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
[hf_assisted]
  tokens_per_second_median = 23.302891492129334
  tokens_per_second_p25 = 21.183524392621223
  tokens_per_second_p75 = 24.424857044996198
  tokens_per_second_mean = 22.804913149518683
  gpu_memory_gb = 4.010546684265137
  forward_counts.drafter_forwards = 1385
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 7.231638418079096
  forward_counts.n_generations = 10
  forward_counts.source = measured with a forward hook
  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 = 24.429024644223723
  tokens_per_second_p25 = 24.328362833297252
  tokens_per_second_p75 = 24.54774113031622
  tokens_per_second_mean = 24.451439555297668
  gpu_memory_gb = 3.9279322624206543
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 1.0078740157480315
  forward_counts.n_generations = 10
  forward_counts.source = measured with a forward hook
  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
[hf_greedy_compiled]
  tokens_per_second_median = 50.84919550812026
  tokens_per_second_p25 = 50.817140359183256
  tokens_per_second_p75 = 51.01212214615007
  tokens_per_second_mean = 50.817803955260814
  gpu_memory_gb = 3.9221267700195312
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 1.0078740157480315
  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.by_kind = {'cuda': 132}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.per_target_forward = 1.0393700787401574
  host_syncs.prompt_id = HumanEval/23
  host_syncs.target_forwards_after_prefill = 127
  host_syncs.target_forwards_source = derived: one per step / per decoded token
  host_syncs.total = 132
  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

==============================================================================
target=Qwen/Qwen2.5-1.5B-Instruct  drafter=Qwen/Qwen2.5-0.5B-Instruct  dataset=humaneval  mode=sample T=0.7
commit=3c43e08  n_samples=10
------------------------------------------------------------------------------
[flashspec_ucb]
  tokens_per_second_median = 28.096687974443004
  tokens_per_second_p25 = 25.489319336463833
  tokens_per_second_p75 = 37.08217864854619
  tokens_per_second_mean = 30.604296376997105
  acceptance_rate_mean = 0.3852359399073896
  gpu_memory_gb = 3.942927360534668
  speedup_vs_flashspec_compiled_mean = 0.3978888521087383
  speedup_vs_flashspec_compiled_median = 0.3928919140166261
  speedup_vs_flashspec_compiled_std = 0.02650422343652674
  speedup_vs_hf_assisted_mean = 1.569289280913687
  speedup_vs_hf_assisted_median = 1.5315892372032842
  speedup_vs_hf_assisted_std = 0.31261723322137347
  speedup_vs_hf_sample_compiled_mean = 0.6062646965059815
  speedup_vs_hf_sample_compiled_median = 0.5554736867969583
  speedup_vs_hf_sample_compiled_std = 0.12191114176173863
  speedup_vs_hf_sample_mean = 1.2521521094667833
  speedup_vs_hf_sample_median = 1.1429797583795867
  speedup_vs_hf_sample_std = 0.25603553851070976
  forward_counts.drafter_forwards = 600
  forward_counts.drafter_forwards_per_step = 1.104868913857678
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 2.397003745318352
  forward_counts.n_generations = 10
  forward_counts.n_steps = 534
  forward_counts.source = measured with a forward hook
  forward_counts.target_forwards = 544
  forward_counts.target_forwards_per_emitted_token = 0.425
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.004433473129645359
  step_profile.bookkeeping_ms_mean = 0.36042385033509705
  step_profile.bookkeeping_ms_total = 192.46633607894182
  step_profile.draft_fraction = 0.4790506348248191
  step_profile.draft_ms_mean = 38.94492404938637
  step_profile.draft_ms_total = 20796.589442372322
  step_profile.n_steps = 534.0
  step_profile.sample_fraction = 0.014898809868698282
  step_profile.sample_ms_mean = 1.2112143823270047
  step_profile.sample_ms_total = 646.7884801626205
  step_profile.step_ms_mean = 81.29604934899604
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.397003745318352
  step_profile.verify_fraction = 0.5016170821768373
  step_profile.verify_ms_mean = 40.77948706694757
  step_profile.verify_ms_total = 21776.24609375
  host_syncs.by_kind = {'cuda': 61}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 58
  host_syncs.per_step = 1.0517241379310345
  host_syncs.per_target_forward = 1.0517241379310345
  host_syncs.prompt_id = HumanEval/23
  host_syncs.target_forwards_after_prefill = 58
  host_syncs.total = 61
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 73
  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 = 59
[flashspec_compiled]
  tokens_per_second_median = 75.73985864164916
  tokens_per_second_p25 = 66.31949556957646
  tokens_per_second_p75 = 86.96726957235924
  tokens_per_second_mean = 76.60502008939667
  acceptance_rate_mean = 0.3852643082609558
  gpu_memory_gb = 3.943270206451416
  speedup_vs_flashspec_ucb_mean = 2.5247021501499236
  speedup_vs_flashspec_ucb_median = 2.5453047977199414
  speedup_vs_flashspec_ucb_std = 0.17231765454179995
  speedup_vs_hf_assisted_mean = 3.9263771569747545
  speedup_vs_hf_assisted_median = 3.9158860079861073
  speedup_vs_hf_assisted_std = 0.6016504639607045
  speedup_vs_hf_sample_compiled_mean = 1.5176170977757173
  speedup_vs_hf_sample_compiled_median = 1.5026896386679338
  speedup_vs_hf_sample_compiled_std = 0.24766793632026046
  speedup_vs_hf_sample_mean = 3.1330361948583154
  speedup_vs_hf_sample_median = 3.066243803639478
  speedup_vs_hf_sample_std = 0.5153358793984777
  forward_counts.drafter_forwards = 602
  forward_counts.drafter_forwards_per_step = 1.1106941838649156
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 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; drafter forwards from the per-step arm records
  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 = 0.011474232736719525
  step_profile.bookkeeping_ms_mean = 0.3683031292782492
  step_profile.bookkeeping_ms_total = 196.30556790530682
  step_profile.draft_fraction = 0.3691206163172414
  step_profile.draft_ms_mean = 11.848136706840366
  step_profile.draft_ms_total = 6315.056864745915
  step_profile.n_steps = 533.0
  step_profile.sample_fraction = 0.01884200796581996
  step_profile.sample_ms_mean = 0.6047960377768326
  step_profile.sample_ms_total = 322.3562881350517
  step_profile.step_ms_mean = 32.09827948666368
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.401500938086304
  step_profile.verify_fraction = 0.6005631429802191
  step_profile.verify_ms_mean = 19.277043612768235
  step_profile.verify_ms_total = 10274.664245605469
  host_syncs.by_kind = {'cuda': 63}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 58
  host_syncs.per_step = 1.0862068965517242
  host_syncs.per_target_forward = 1.0862068965517242
  host_syncs.prompt_id = HumanEval/23
  host_syncs.target_forwards_after_prefill = 58
  host_syncs.target_forwards_source = derived: one per step / per decoded token
  host_syncs.total = 63
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[hf_assisted]
  tokens_per_second_median = 18.844078044479534
  tokens_per_second_p25 = 18.04090535750411
  tokens_per_second_p75 = 21.24417385471367
  tokens_per_second_mean = 19.63495972218567
  gpu_memory_gb = 4.024676322937012
  forward_counts.drafter_forwards = 1565
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 6.037735849056604
  forward_counts.n_generations = 10
  forward_counts.source = measured with a forward hook
  forward_counts.target_forwards = 222
  forward_counts.target_forwards_per_emitted_token = 0.1734375
  forward_counts.tokens_emitted = 1280
  host_syncs.by_kind = {'cuda': 911}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.per_target_forward = 29.387096774193548
  host_syncs.prompt_id = HumanEval/23
  host_syncs.target_forwards_after_prefill = 31
  host_syncs.total = 911
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 185
  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 = 32
[hf_sample]
  tokens_per_second_median = 24.497491036510233
  tokens_per_second_p25 = 24.352074445631605
  tokens_per_second_p75 = 24.59023809444537
  tokens_per_second_mean = 24.45979545298337
  gpu_memory_gb = 3.9425158500671387
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 1.0078740157480315
  forward_counts.n_generations = 10
  forward_counts.source = measured with a forward hook
  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
[hf_sample_compiled]
  tokens_per_second_median = 50.42612714053795
  tokens_per_second_p25 = 50.38888740804433
  tokens_per_second_p75 = 50.596619340479606
  tokens_per_second_mean = 50.48580054918255
  gpu_memory_gb = 3.9360790252685547
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 1.0078740157480315
  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.by_kind = {'cuda': 132}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.per_target_forward = 1.0393700787401574
  host_syncs.prompt_id = HumanEval/23
  host_syncs.target_forwards_after_prefill = 127
  host_syncs.target_forwards_source = derived: one per step / per decoded token
  host_syncs.total = 132
  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=mt_bench  mode=greedy
commit=3c43e08  n_samples=10
------------------------------------------------------------------------------
[flashspec_ucb]
  tokens_per_second_median = 24.83938465726492
  tokens_per_second_p25 = 22.961833475251858
  tokens_per_second_p75 = 27.021801893823184
  tokens_per_second_mean = 25.670913943695915
  acceptance_rate_mean = 0.2620603532680512
  gpu_memory_gb = 3.9019665718078613
  speedup_vs_flashspec_compiled_mean = 0.38795171348549673
  speedup_vs_flashspec_compiled_median = 0.3833218503217505
  speedup_vs_flashspec_compiled_std = 0.017973596368473305
  speedup_vs_hf_assisted_mean = 1.552765151105294
  speedup_vs_hf_assisted_median = 1.5451784248202172
  speedup_vs_hf_assisted_std = 0.31642436256421863
  speedup_vs_hf_greedy_compiled_mean = 0.49801931172996394
  speedup_vs_hf_greedy_compiled_median = 0.4828425422653919
  speedup_vs_hf_greedy_compiled_std = 0.061584696771344455
  speedup_vs_hf_greedy_mean = 1.0546085257521942
  speedup_vs_hf_greedy_median = 1.0179685727219971
  speedup_vs_hf_greedy_std = 0.13683557531973306
  forward_counts.drafter_forwards = 708
  forward_counts.drafter_forwards_per_step = 1.0889235569422777
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 1.9968798751950079
  forward_counts.n_generations = 10
  forward_counts.n_steps = 641
  forward_counts.source = measured with a forward hook
  forward_counts.target_forwards = 651
  forward_counts.target_forwards_per_emitted_token = 0.50859375
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.0051596999625681235
  step_profile.bookkeeping_ms_mean = 0.40682793084340235
  step_profile.bookkeeping_ms_total = 260.7767036706209
  step_profile.draft_fraction = 0.4785017060963863
  step_profile.draft_ms_mean = 37.72852305530944
  step_profile.draft_ms_total = 24183.98327845335
  step_profile.n_steps = 641.0
  step_profile.sample_fraction = 0.006193558787204519
  step_profile.sample_ms_mean = 0.4883448115656045
  step_profile.sample_ms_total = 313.0290242135525
  step_profile.step_ms_mean = 78.84720696838988
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 1.9968798751950079
  step_profile.verify_fraction = 0.510145035153841
  step_profile.verify_ms_mean = 40.223511170671436
  step_profile.verify_ms_total = 25783.27066040039
  host_syncs.by_kind = {'cuda': 67}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 63
  host_syncs.per_step = 1.0634920634920635
  host_syncs.per_target_forward = 1.0634920634920635
  host_syncs.prompt_id = 81
  host_syncs.target_forwards_after_prefill = 63
  host_syncs.total = 67
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 87
  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 = 64
  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
[flashspec_compiled]
  tokens_per_second_median = 65.5665446517103
  tokens_per_second_p25 = 59.93265391596037
  tokens_per_second_p75 = 71.89633242799991
  tokens_per_second_mean = 66.06565866300903
  acceptance_rate_mean = 0.2868673245081164
  gpu_memory_gb = 3.9116411209106445
  speedup_vs_flashspec_ucb_mean = 2.5830904698586523
  speedup_vs_flashspec_ucb_median = 2.6090134422745286
  speedup_vs_flashspec_ucb_std = 0.11778598240757707
  speedup_vs_hf_assisted_mean = 4.0204623827915835
  speedup_vs_hf_assisted_median = 3.9853712833572263
  speedup_vs_hf_assisted_std = 0.8930301058692843
  speedup_vs_hf_greedy_compiled_mean = 1.2816431927301075
  speedup_vs_hf_greedy_compiled_median = 1.2745206859837361
  speedup_vs_hf_greedy_compiled_std = 0.12505549558718806
  speedup_vs_hf_greedy_mean = 2.713430536441695
  speedup_vs_hf_greedy_median = 2.6870360558531057
  speedup_vs_hf_greedy_std = 0.27848931860219095
  forward_counts.drafter_forwards = 696
  forward_counts.drafter_forwards_per_step = 1.0837282780410742
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 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; drafter forwards from the per-step arm records
  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 = 0.013448296516422575
  step_profile.bookkeeping_ms_mean = 0.4144613198335702
  step_profile.bookkeeping_ms_total = 262.3540154546499
  step_profile.draft_fraction = 0.35598297869772716
  step_profile.draft_ms_mean = 10.970993613144515
  step_profile.draft_ms_total = 6944.638957120478
  step_profile.n_steps = 633.0
  step_profile.sample_fraction = 0.01293836748752573
  step_profile.sample_ms_mean = 0.39874588270887945
  step_profile.sample_ms_total = 252.4061437547207
  step_profile.step_ms_mean = 30.818871321542098
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.022116903633491
  step_profile.verify_fraction = 0.6176303572983245
  step_profile.verify_ms_mean = 19.034670505855132
  step_profile.verify_ms_total = 12048.946430206299
  host_syncs.by_kind = {'cuda': 72}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 67
  host_syncs.per_step = 1.0746268656716418
  host_syncs.per_target_forward = 1.0746268656716418
  host_syncs.prompt_id = 81
  host_syncs.target_forwards_after_prefill = 67
  host_syncs.target_forwards_source = derived: one per step / per decoded token
  host_syncs.total = 72
  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_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
[hf_assisted]
  tokens_per_second_median = 18.834736736908617
  tokens_per_second_p25 = 14.102116189597641
  tokens_per_second_p75 = 20.858267222930852
  tokens_per_second_mean = 17.35488273659592
  gpu_memory_gb = 3.982469081878662
  forward_counts.drafter_forwards = 1803
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 3.8905775075987843
  forward_counts.n_generations = 10
  forward_counts.source = measured with a forward hook
  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 = 24.358718936743088
  tokens_per_second_p25 = 24.18066265396688
  tokens_per_second_p75 = 24.491882521750334
  tokens_per_second_mean = 24.360896424368537
  gpu_memory_gb = 3.9123830795288086
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 1.0078740157480315
  forward_counts.n_generations = 10
  forward_counts.source = measured with a forward hook
  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
[hf_greedy_compiled]
  tokens_per_second_median = 51.56689852526201
  tokens_per_second_p25 = 51.44110439771817
  tokens_per_second_p75 = 51.6446975154745
  tokens_per_second_mean = 51.546810222699335
  gpu_memory_gb = 3.9070162773132324
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 1.0078740157480315
  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.by_kind = {'cuda': 132}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.per_target_forward = 1.0393700787401574
  host_syncs.prompt_id = 81
  host_syncs.target_forwards_after_prefill = 127
  host_syncs.target_forwards_source = derived: one per step / per decoded token
  host_syncs.total = 132
  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

==============================================================================
target=Qwen/Qwen2.5-1.5B-Instruct  drafter=Qwen/Qwen2.5-0.5B-Instruct  dataset=mt_bench  mode=sample T=0.7
commit=3c43e08  n_samples=10
------------------------------------------------------------------------------
[flashspec_ucb]
  tokens_per_second_median = 25.26133407738635
  tokens_per_second_p25 = 23.020854972868243
  tokens_per_second_p75 = 28.43236543070436
  tokens_per_second_mean = 27.023215157122003
  acceptance_rate_mean = 0.2884380538950996
  gpu_memory_gb = 3.9193735122680664
  speedup_vs_flashspec_compiled_mean = 0.40295436676352525
  speedup_vs_flashspec_compiled_median = 0.398933860867385
  speedup_vs_flashspec_compiled_std = 0.02145496019125291
  speedup_vs_hf_assisted_mean = 2.297109002161638
  speedup_vs_hf_assisted_median = 1.784691816669535
  speedup_vs_hf_assisted_std = 1.2337829416012864
  speedup_vs_hf_sample_compiled_mean = 0.5378824437702476
  speedup_vs_hf_sample_compiled_median = 0.5027379844644687
  speedup_vs_hf_sample_compiled_std = 0.12618979818858142
  speedup_vs_hf_sample_mean = 1.1087222394251202
  speedup_vs_hf_sample_median = 1.0434498871962403
  speedup_vs_hf_sample_std = 0.269462038860216
  forward_counts.drafter_forwards = 696
  forward_counts.drafter_forwards_per_step = 1.104669887278583
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 2.0611916264090175
  forward_counts.n_generations = 10
  forward_counts.n_steps = 621
  forward_counts.source = measured with a forward hook
  forward_counts.target_forwards = 631
  forward_counts.target_forwards_per_emitted_token = 0.49296875
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.004883445780903506
  step_profile.bookkeeping_ms_mean = 0.38871358356614044
  step_profile.bookkeeping_ms_total = 241.3911353945732
  step_profile.draft_fraction = 0.4803419806832038
  step_profile.draft_ms_mean = 38.234365860837094
  step_profile.draft_ms_total = 23743.541199579835
  step_profile.n_steps = 621.0
  step_profile.sample_fraction = 0.014224917689339391
  step_profile.sample_ms_mean = 1.1322781042392318
  step_profile.sample_ms_total = 703.144702732563
  step_profile.step_ms_mean = 79.59821834946696
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.0611916264090175
  step_profile.verify_fraction = 0.5005496558465533
  step_profile.verify_ms_mean = 39.842860800824496
  step_profile.verify_ms_total = 24742.41655731201
  host_syncs.by_kind = {'cuda': 73}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 70
  host_syncs.per_step = 1.042857142857143
  host_syncs.per_target_forward = 1.042857142857143
  host_syncs.prompt_id = 81
  host_syncs.target_forwards_after_prefill = 70
  host_syncs.total = 73
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 93
  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 = 71
[flashspec_compiled]
  tokens_per_second_median = 63.11715269934044
  tokens_per_second_p25 = 59.5644483083085
  tokens_per_second_p75 = 69.07556580010277
  tokens_per_second_mean = 66.59285605438832
  acceptance_rate_mean = 0.30095750082855754
  gpu_memory_gb = 3.9206857681274414
  speedup_vs_flashspec_ucb_mean = 2.488609808542554
  speedup_vs_flashspec_ucb_median = 2.5080845996363026
  speedup_vs_flashspec_ucb_std = 0.13041137513906798
  speedup_vs_hf_assisted_mean = 5.686528196854407
  speedup_vs_hf_assisted_median = 4.474878921310762
  speedup_vs_hf_assisted_std = 2.9017388173754006
  speedup_vs_hf_sample_compiled_mean = 1.3254296549193796
  speedup_vs_hf_sample_compiled_median = 1.2630258753083248
  speedup_vs_hf_sample_compiled_std = 0.2452748624439106
  speedup_vs_hf_sample_mean = 2.730970476863263
  speedup_vs_hf_sample_median = 2.5801491971418065
  speedup_vs_hf_sample_std = 0.5264049164831425
  forward_counts.drafter_forwards = 716
  forward_counts.drafter_forwards_per_step = 1.1423948220064726
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 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; drafter forwards from the per-step arm records
  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 = 0.01251376030775333
  step_profile.bookkeeping_ms_mean = 0.39932380590141786
  step_profile.bookkeeping_ms_total = 246.78211204707623
  step_profile.draft_fraction = 0.36903386977357544
  step_profile.draft_ms_mean = 11.776157266909443
  step_profile.draft_ms_total = 7277.665190950036
  step_profile.n_steps = 618.0
  step_profile.sample_fraction = 0.018938116275148777
  step_profile.sample_ms_mean = 0.6043299920736394
  step_profile.sample_ms_total = 373.4759351015091
  step_profile.step_ms_mean = 31.91077630390627
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.071197411003236
  step_profile.verify_fraction = 0.5995142536435225
  step_profile.verify_ms_mean = 19.13096523902177
  step_profile.verify_ms_total = 11822.936517715454
  host_syncs.by_kind = {'cuda': 73}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 68
  host_syncs.per_step = 1.0735294117647058
  host_syncs.per_target_forward = 1.0735294117647058
  host_syncs.prompt_id = 81
  host_syncs.target_forwards_after_prefill = 68
  host_syncs.target_forwards_source = derived: one per step / per decoded token
  host_syncs.total = 73
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[hf_assisted]
  tokens_per_second_median = 14.341218875882541
  tokens_per_second_p25 = 9.085275562353358
  tokens_per_second_p75 = 18.19088437853941
  tokens_per_second_mean = 14.066216700106787
  gpu_memory_gb = 4.005721092224121
  forward_counts.drafter_forwards = 2675
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 3.9751552795031055
  forward_counts.n_generations = 10
  forward_counts.source = measured with a forward hook
  forward_counts.target_forwards = 332
  forward_counts.target_forwards_per_emitted_token = 0.259375
  forward_counts.tokens_emitted = 1280
  host_syncs.by_kind = {'cuda': 1603}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.per_target_forward = 42.18421052631579
  host_syncs.prompt_id = 81
  host_syncs.target_forwards_after_prefill = 38
  host_syncs.total = 1603
  nonfinite_logits.drafter_entries = 0
  nonfinite_logits.drafter_forwards_checked = 388
  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 = 39
[hf_sample]
  tokens_per_second_median = 24.371632087092596
  tokens_per_second_p25 = 24.137529875735552
  tokens_per_second_p75 = 24.643657960930078
  tokens_per_second_mean = 24.415557756138078
  gpu_memory_gb = 3.9260549545288086
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 1.0078740157480315
  forward_counts.n_generations = 10
  forward_counts.source = measured with a forward hook
  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
[hf_sample_compiled]
  tokens_per_second_median = 50.24734945285092
  tokens_per_second_p25 = 50.13961412887575
  tokens_per_second_p75 = 50.398525858279534
  tokens_per_second_mean = 50.2443149822003
  gpu_memory_gb = 3.9206881523132324
  forward_counts.emitted_tokens_per_target_forward_after_prefill = 1.0078740157480315
  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.by_kind = {'cuda': 132}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.per_target_forward = 1.0393700787401574
  host_syncs.prompt_id = 81
  host_syncs.target_forwards_after_prefill = 127
  host_syncs.target_forwards_source = derived: one per step / per decoded token
  host_syncs.total = 132
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile

