==============================================================================
target=Qwen/Qwen2.5-1.5B-Instruct  drafter=Qwen/Qwen2.5-0.5B-Instruct  dataset=humaneval  mode=greedy
commit=3e38251  n_samples=10
------------------------------------------------------------------------------
[arm_composite-g4]
  workload_tokens_per_second = 76.46375542673032
  tokens_per_second_median = 79.55222429574867
  tokens_per_second_p25 = 69.85021485155741
  tokens_per_second_p75 = 83.9947095650734
  tokens_per_second_mean = 77.65441220301864
  acceptance_rate_mean = 0.5109862676726286
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_ngram-g2_median = 1.209200028509544
  speedup_vs_arm_ngram-g4_median = 1.161011628265475
  speedup_vs_arm_ngram-g6_median = 1.126736540206736
  speedup_vs_arm_qwen-g2_median = 1.1721936061513822
  speedup_vs_arm_qwen-g4_median = 1.1074321067953634
  speedup_vs_arm_qwen-g6_median = 1.119002396743674
  speedup_vs_bandit_thompson_median = 1.1211510959572473
  speedup_vs_bandit_ucb1_median = 1.0558175047946516
  speedup_vs_hf_assisted_median = 3.424878726821996
  speedup_vs_hf_greedy_compiled_median = 1.6336698480217855
  speedup_vs_hf_greedy_median = 3.218814241226999
  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 = 0.0100150709310151
  step_profile.bookkeeping_ms_mean = 0.3846131496388337
  step_profile.bookkeeping_ms_total = 167.30672009289265
  step_profile.draft_fraction = 0.44738493052459494
  step_profile.draft_ms_mean = 17.181119176813926
  step_profile.draft_ms_total = 7473.786841914058
  step_profile.n_steps = 435.0
  step_profile.sample_fraction = 0.011299840962884443
  step_profile.sample_ms_mean = 0.43395273514177607
  step_profile.sample_ms_total = 188.7694397866726
  step_profile.step_ms_mean = 38.40343740829106
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.942528735632184
  step_profile.verify_fraction = 0.5313001575815055
  step_profile.verify_ms_mean = 20.403752346696525
  step_profile.verify_ms_total = 8875.632270812988
  host_syncs.by_kind = {'cuda': 45}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 40
  host_syncs.per_step = 1.125
  host_syncs.prompt_id = HumanEval/23
  host_syncs.total = 45
  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_ngram-g2]
  workload_tokens_per_second = 63.45036436267508
  tokens_per_second_median = 66.84784131066232
  tokens_per_second_p25 = 56.744218913520655
  tokens_per_second_p75 = 69.9624690267307
  tokens_per_second_mean = 64.33934662579648
  acceptance_rate_mean = 0.319459638231722
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.826993034069013
  speedup_vs_arm_ngram-g4_median = 0.9506681099244605
  speedup_vs_arm_ngram-g6_median = 0.9332705422504608
  speedup_vs_arm_qwen-g2_median = 0.955244528749868
  speedup_vs_arm_qwen-g4_median = 0.9257622006928928
  speedup_vs_arm_qwen-g6_median = 0.9173940336162689
  speedup_vs_bandit_thompson_median = 0.9163521858094513
  speedup_vs_bandit_ucb1_median = 0.8850580374180221
  speedup_vs_hf_assisted_median = 2.809310271878428
  speedup_vs_hf_greedy_compiled_median = 1.3733131470099311
  speedup_vs_hf_greedy_median = 2.705659383955138
  forward_counts.emitted_tokens_per_step = 1.3333333333333333
  forward_counts.n_generations = 10
  forward_counts.n_steps = 960
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 970
  forward_counts.target_forwards_per_emitted_token = 0.7578125
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.018645336118626288
  step_profile.bookkeeping_ms_mean = 0.3907371670628587
  step_profile.bookkeeping_ms_total = 375.1076803803444
  step_profile.draft_fraction = 0.004600512602599414
  step_profile.draft_ms_mean = 0.09640969998824099
  step_profile.draft_ms_total = 92.55331198871136
  step_profile.n_steps = 960.0
  step_profile.sample_fraction = 0.02016970703814803
  step_profile.sample_ms_mean = 0.42268233398596444
  step_profile.sample_ms_total = 405.7750406265259
  step_profile.step_ms_mean = 20.95629516018865
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 1.3333333333333333
  step_profile.verify_fraction = 0.9565844442406263
  step_profile.verify_ms_mean = 20.046465959151586
  step_profile.verify_ms_total = 19244.607320785522
  host_syncs.by_kind = {'cuda': 92}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 90
  host_syncs.per_step = 1.0222222222222221
  host_syncs.prompt_id = HumanEval/23
  host_syncs.total = 92
  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_ngram-g4]
  workload_tokens_per_second = 66.42243963127225
  tokens_per_second_median = 70.05104131642989
  tokens_per_second_p25 = 58.21218285588789
  tokens_per_second_p75 = 75.46789045105838
  tokens_per_second_mean = 67.91835659272508
  acceptance_rate_mean = 0.21773360382965218
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.8613758858636632
  speedup_vs_arm_ngram-g2_median = 1.0518934895534264
  speedup_vs_arm_ngram-g6_median = 0.9876396541540635
  speedup_vs_arm_qwen-g2_median = 1.0047961531118568
  speedup_vs_arm_qwen-g4_median = 0.9756850080933754
  speedup_vs_arm_qwen-g6_median = 0.9541701477173078
  speedup_vs_bandit_thompson_median = 0.9566384604255371
  speedup_vs_bandit_ucb1_median = 0.9353636760734767
  speedup_vs_hf_assisted_median = 2.9144179660817247
  speedup_vs_hf_greedy_compiled_median = 1.44051185660832
  speedup_vs_hf_greedy_median = 2.878626955227327
  forward_counts.emitted_tokens_per_step = 1.41749723145072
  forward_counts.n_generations = 10
  forward_counts.n_steps = 903
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 913
  forward_counts.target_forwards_per_emitted_token = 0.71328125
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.019486661318673795
  step_profile.bookkeeping_ms_mean = 0.41476667160013586
  step_profile.bookkeeping_ms_total = 374.5343044549227
  step_profile.draft_fraction = 0.004600161549918617
  step_profile.draft_ms_mean = 0.09791280628735836
  step_profile.draft_ms_total = 88.41526407748461
  step_profile.n_steps = 903.0
  step_profile.sample_fraction = 0.02020882108569596
  step_profile.sample_ms_mean = 0.430137585992961
  step_profile.sample_ms_total = 388.41424015164375
  step_profile.step_ms_mean = 21.284645164057466
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 1.41749723145072
  step_profile.verify_fraction = 0.9557043560457116
  step_profile.verify_ms_mean = 20.341828100177008
  step_profile.verify_ms_total = 18368.67077445984
  host_syncs.by_kind = {'cuda': 85}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 83
  host_syncs.per_step = 1.0240963855421688
  host_syncs.prompt_id = HumanEval/23
  host_syncs.total = 85
  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_ngram-g6]
  workload_tokens_per_second = 67.18331562955943
  tokens_per_second_median = 72.0054835701662
  tokens_per_second_p25 = 58.44619047814492
  tokens_per_second_p75 = 75.94142059149232
  tokens_per_second_mean = 68.98838043177065
  acceptance_rate_mean = 0.16325668379711872
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.8875771592143998
  speedup_vs_arm_ngram-g2_median = 1.071500670592608
  speedup_vs_arm_ngram-g4_median = 1.0125184546625365
  speedup_vs_arm_qwen-g2_median = 1.0280032894671502
  speedup_vs_arm_qwen-g4_median = 0.9980351118667179
  speedup_vs_arm_qwen-g6_median = 0.9666808017576305
  speedup_vs_bandit_thompson_median = 0.9751605513271736
  speedup_vs_bandit_ucb1_median = 0.9524121259497353
  speedup_vs_hf_assisted_median = 2.93911399128529
  speedup_vs_hf_greedy_compiled_median = 1.4806937038083317
  speedup_vs_hf_greedy_median = 2.92933069579576
  forward_counts.emitted_tokens_per_step = 1.4446952595936795
  forward_counts.n_generations = 10
  forward_counts.n_steps = 886
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 896
  forward_counts.target_forwards_per_emitted_token = 0.7
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.01933840409642531
  step_profile.bookkeeping_ms_mean = 0.4147448489614857
  step_profile.bookkeeping_ms_total = 367.4639361798763
  step_profile.draft_fraction = 0.004544975808798174
  step_profile.draft_ms_mean = 0.09747470866543984
  step_profile.draft_ms_total = 86.36259187757969
  step_profile.n_steps = 886.0
  step_profile.sample_fraction = 0.020387139978088206
  step_profile.sample_ms_mean = 0.43723676725381116
  step_profile.sample_ms_total = 387.3917757868767
  step_profile.step_ms_mean = 21.446694716558902
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 1.4446952595936795
  step_profile.verify_fraction = 0.9557294801166883
  step_profile.verify_ms_mean = 20.497238391678167
  step_profile.verify_ms_total = 18160.553215026855
  host_syncs.by_kind = {'cuda': 83}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 81
  host_syncs.per_step = 1.0246913580246915
  host_syncs.prompt_id = HumanEval/23
  host_syncs.total = 83
  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
[arm_qwen-g2]
  workload_tokens_per_second = 66.76880461803883
  tokens_per_second_median = 67.34153565781634
  tokens_per_second_p25 = 64.12038418371955
  tokens_per_second_p75 = 69.77332912738335
  tokens_per_second_mean = 66.93771687250549
  acceptance_rate_mean = 0.8476525252803826
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.8533600645891612
  speedup_vs_arm_ngram-g2_median = 1.0470956796508792
  speedup_vs_arm_ngram-g4_median = 0.9954213035872793
  speedup_vs_arm_ngram-g6_median = 0.9732031717334486
  speedup_vs_arm_qwen-g4_median = 0.9476407674749002
  speedup_vs_arm_qwen-g6_median = 0.9664187726512938
  speedup_vs_bandit_thompson_median = 0.9439637633694145
  speedup_vs_bandit_ucb1_median = 0.9287173703151013
  speedup_vs_hf_assisted_median = 2.9075625918706143
  speedup_vs_hf_greedy_compiled_median = 1.3792922691429672
  speedup_vs_hf_greedy_median = 2.7717879564036716
  forward_counts.emitted_tokens_per_step = 2.7176220806794054
  forward_counts.n_generations = 10
  forward_counts.n_steps = 471
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 481
  forward_counts.target_forwards_per_emitted_token = 0.37578125
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.006999275615533069
  step_profile.bookkeeping_ms_mean = 0.28455806392266747
  step_profile.bookkeeping_ms_total = 134.02684810757637
  step_profile.draft_fraction = 0.5105699578830687
  step_profile.draft_ms_mean = 20.757405007720752
  step_profile.draft_ms_total = 9776.737758636475
  step_profile.n_steps = 471.0
  step_profile.sample_fraction = 0.003763373660560579
  step_profile.sample_ms_mean = 0.15300130777677912
  step_profile.sample_ms_total = 72.06361596286297
  step_profile.step_ms_mean = 40.65535914761878
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.7176220806794054
  step_profile.verify_fraction = 0.4786673928408377
  step_profile.verify_ms_mean = 19.460394768198583
  step_profile.verify_ms_total = 9165.845935821533
  host_syncs.by_kind = {'cuda': 55}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 51
  host_syncs.per_step = 1.0784313725490196
  host_syncs.prompt_id = HumanEval/23
  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, 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_qwen-g4]
  workload_tokens_per_second = 69.73968151775765
  tokens_per_second_median = 72.0525172430655
  tokens_per_second_p25 = 65.80254707821253
  tokens_per_second_p75 = 74.40605150272695
  tokens_per_second_mean = 70.12845579720118
  acceptance_rate_mean = 0.7637703903145825
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.9033449444544823
  speedup_vs_arm_ngram-g2_median = 1.0802593108884069
  speedup_vs_arm_ngram-g4_median = 1.0250476036237586
  speedup_vs_arm_ngram-g6_median = 1.0019828015021526
  speedup_vs_arm_qwen-g2_median = 1.055252686095979
  speedup_vs_arm_qwen-g6_median = 1.037492052749065
  speedup_vs_bandit_thompson_median = 1.000847144975118
  speedup_vs_bandit_ucb1_median = 0.934109816361641
  speedup_vs_hf_assisted_median = 3.0456719751984176
  speedup_vs_hf_greedy_compiled_median = 1.4665351644632312
  speedup_vs_hf_greedy_median = 2.9721827283085664
  forward_counts.emitted_tokens_per_step = 4.169381107491857
  forward_counts.n_generations = 10
  forward_counts.n_steps = 307
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 317
  forward_counts.target_forwards_per_emitted_token = 0.24765625
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.0060429912929673545
  step_profile.bookkeeping_ms_mean = 0.3609827228303064
  step_profile.bookkeeping_ms_total = 110.82169590890408
  step_profile.draft_fraction = 0.6566978264488779
  step_profile.draft_ms_mean = 39.22834867297244
  step_profile.draft_ms_total = 12043.103042602539
  step_profile.n_steps = 307.0
  step_profile.sample_fraction = 0.002696511807402415
  step_profile.sample_ms_mean = 0.1610782023652757
  step_profile.sample_ms_total = 49.45100812613964
  step_profile.step_ms_mean = 59.73576749157439
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 4.169381107491857
  step_profile.verify_fraction = 0.33456267045075244
  step_profile.verify_ms_mean = 19.985357893406373
  step_profile.verify_ms_total = 6135.504873275757
  host_syncs.by_kind = {'cuda': 39}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 36
  host_syncs.per_step = 1.0833333333333333
  host_syncs.prompt_id = HumanEval/23
  host_syncs.total = 39
  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_qwen-g6]
  workload_tokens_per_second = 66.88096892103184
  tokens_per_second_median = 70.377656575358
  tokens_per_second_p25 = 61.62004042549512
  tokens_per_second_p75 = 73.02696674083388
  tokens_per_second_mean = 67.87436327217897
  acceptance_rate_mean = 0.6879125684268674
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.8936762295037795
  speedup_vs_arm_ngram-g2_median = 1.0901370572285969
  speedup_vs_arm_ngram-g4_median = 1.0480423866040844
  speedup_vs_arm_ngram-g6_median = 1.0345000154620396
  speedup_vs_arm_qwen-g2_median = 1.0347481473595648
  speedup_vs_arm_qwen-g4_median = 0.9639307890395676
  speedup_vs_bandit_thompson_median = 0.9809033752941556
  speedup_vs_bandit_ucb1_median = 0.8991175976722829
  speedup_vs_hf_assisted_median = 3.0085938136382877
  speedup_vs_hf_greedy_compiled_median = 1.4401620338571082
  speedup_vs_hf_greedy_median = 2.8966694981910583
  forward_counts.emitted_tokens_per_step = 5.267489711934156
  forward_counts.n_generations = 10
  forward_counts.n_steps = 243
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 253
  forward_counts.target_forwards_per_emitted_token = 0.19765625
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.005582842520669357
  step_profile.bookkeeping_ms_mean = 0.4393825188831047
  step_profile.bookkeeping_ms_total = 106.76995208859444
  step_profile.draft_fraction = 0.7323795474503949
  step_profile.draft_ms_mean = 57.63995118003814
  step_profile.draft_ms_total = 14006.508136749268
  step_profile.n_steps = 243.0
  step_profile.sample_fraction = 0.0022353450028251244
  step_profile.sample_ms_mean = 0.17592678179220897
  step_profile.sample_ms_total = 42.75020797550678
  step_profile.step_ms_mean = 78.70229497901451
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 5.267489711934156
  step_profile.verify_fraction = 0.2598022650261106
  step_profile.verify_ms_mean = 20.44703449830106
  step_profile.verify_ms_total = 4968.629383087158
  host_syncs.by_kind = {'cuda': 35}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 32
  host_syncs.per_step = 1.09375
  host_syncs.prompt_id = HumanEval/23
  host_syncs.total = 35
  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
[bandit_thompson]
  workload_tokens_per_second = 68.90401703893421
  tokens_per_second_median = 71.09938862169275
  tokens_per_second_p25 = 64.20817091596366
  tokens_per_second_p75 = 75.34895686563753
  tokens_per_second_mean = 70.11761424660179
  acceptance_rate_mean = 0.4229770921014879
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.8919485740452465
  speedup_vs_arm_ngram-g2_median = 1.0913676935716488
  speedup_vs_arm_ngram-g4_median = 1.045333609401764
  speedup_vs_arm_ngram-g6_median = 1.0255447743971184
  speedup_vs_arm_qwen-g2_median = 1.059427486441753
  speedup_vs_arm_qwen-g4_median = 0.9992934777918729
  speedup_vs_arm_qwen-g6_median = 1.0196458148221987
  speedup_vs_bandit_ucb1_median = 0.9452608407067249
  speedup_vs_hf_assisted_median = 3.119925954441981
  speedup_vs_hf_greedy_compiled_median = 1.4471674934708139
  speedup_vs_hf_greedy_median = 2.93611763567821
  forward_counts.emitted_tokens_per_step = 2.3443223443223444
  forward_counts.n_generations = 10
  forward_counts.n_steps = 546
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 556
  forward_counts.target_forwards_per_emitted_token = 0.434375
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.012132952032133316
  step_profile.bookkeeping_ms_mean = 0.4119486588534418
  step_profile.bookkeeping_ms_total = 224.92396773397923
  step_profile.draft_fraction = 0.3708571896454205
  step_profile.draft_ms_mean = 12.591669487852208
  step_profile.draft_ms_total = 6875.051540367305
  step_profile.n_steps = 546.0
  step_profile.sample_fraction = 0.012776831694353537
  step_profile.sample_ms_mean = 0.43381022746528025
  step_profile.sample_ms_total = 236.86038419604301
  step_profile.step_ms_mean = 33.952879543446905
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.3443223443223444
  step_profile.verify_fraction = 0.6042330266280926
  step_profile.verify_ms_mean = 20.515451169275977
  step_profile.verify_ms_total = 11201.436338424683
  host_syncs.by_kind = {'cuda': 69}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 65
  host_syncs.per_step = 1.0615384615384615
  host_syncs.prompt_id = HumanEval/23
  host_syncs.total = 69
  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
[bandit_ucb1]
  workload_tokens_per_second = 72.67941480685818
  tokens_per_second_median = 72.3107138721505
  tokens_per_second_p25 = 67.47902688279427
  tokens_per_second_p75 = 82.6769826503066
  tokens_per_second_mean = 73.84568842857585
  acceptance_rate_mean = 0.5318662998514808
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.9472912681240115
  speedup_vs_arm_ngram-g2_median = 1.129870014188402
  speedup_vs_arm_ngram-g4_median = 1.0692174375132026
  speedup_vs_arm_ngram-g6_median = 1.0499796510738038
  speedup_vs_arm_qwen-g2_median = 1.076769270049365
  speedup_vs_arm_qwen-g4_median = 1.0712356157218799
  speedup_vs_arm_qwen-g6_median = 1.1122155653878092
  speedup_vs_bandit_thompson_median = 1.057916357581401
  speedup_vs_hf_assisted_median = 3.350340550221213
  speedup_vs_hf_greedy_compiled_median = 1.489713287635102
  speedup_vs_hf_greedy_median = 2.9545035739994896
  forward_counts.emitted_tokens_per_step = 2.8256070640176603
  forward_counts.n_generations = 10
  forward_counts.n_steps = 453
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 463
  forward_counts.target_forwards_per_emitted_token = 0.36171875
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.00994454578849
  step_profile.bookkeeping_ms_mean = 0.3858436376974809
  step_profile.bookkeeping_ms_total = 174.78716787695885
  step_profile.draft_fraction = 0.45001843114850637
  step_profile.draft_ms_mean = 17.460500680304914
  step_profile.draft_ms_total = 7909.606808178127
  step_profile.n_steps = 453.0
  step_profile.sample_fraction = 0.011257899356393172
  step_profile.sample_ms_mean = 0.43680113027690526
  step_profile.sample_ms_total = 197.87091201543808
  step_profile.step_ms_mean = 38.79952346783534
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.8256070640176603
  step_profile.verify_fraction = 0.5287791237066105
  step_profile.verify_ms_mean = 20.51637801955604
  step_profile.verify_ms_total = 9293.919242858887
  host_syncs.by_kind = {'cuda': 53}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 49
  host_syncs.per_step = 1.0816326530612246
  host_syncs.prompt_id = HumanEval/23
  host_syncs.total = 53
  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_assisted]
  workload_tokens_per_second = 22.64158454982198
  tokens_per_second_median = 23.601277457517536
  tokens_per_second_p25 = 21.31846600755457
  tokens_per_second_p75 = 24.412638219956563
  tokens_per_second_mean = 22.87569727878533
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.29209921158796753
  speedup_vs_arm_ngram-g2_median = 0.3562484443820021
  speedup_vs_arm_ngram-g4_median = 0.343185978401946
  speedup_vs_arm_ngram-g6_median = 0.3402456488827364
  speedup_vs_arm_qwen-g2_median = 0.3439322476359914
  speedup_vs_arm_qwen-g4_median = 0.32857936375379393
  speedup_vs_arm_qwen-g6_median = 0.33238244898686464
  speedup_vs_bandit_thompson_median = 0.32052150402879487
  speedup_vs_bandit_ucb1_median = 0.298633981333441
  speedup_vs_hf_greedy_compiled_median = 0.4869074080107677
  speedup_vs_hf_greedy_median = 0.9610985137724622
  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, 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 = 24.445733157543984
  tokens_per_second_median = 24.409788089906908
  tokens_per_second_p25 = 24.254548477068294
  tokens_per_second_p75 = 24.637155697267822
  tokens_per_second_mean = 24.44814534699301
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.31081846100438715
  speedup_vs_arm_ngram-g2_median = 0.36975160964186954
  speedup_vs_arm_ngram-g4_median = 0.3478859312333811
  speedup_vs_arm_ngram-g6_median = 0.34169789099446124
  speedup_vs_arm_qwen-g2_median = 0.36078019017636254
  speedup_vs_arm_qwen-g4_median = 0.3364667204981497
  speedup_vs_arm_qwen-g6_median = 0.34523604816784936
  speedup_vs_bandit_thompson_median = 0.340698423958426
  speedup_vs_bandit_ucb1_median = 0.338521061063163
  speedup_vs_hf_assisted_median = 1.040481078024873
  speedup_vs_hf_greedy_compiled_median = 0.5021068962602442
  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 = 48.75843727359648
  tokens_per_second_median = 48.75937613455268
  tokens_per_second_p25 = 48.528514050212934
  tokens_per_second_p75 = 49.036691271256906
  tokens_per_second_mean = 48.76005356775071
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.612365424574636
  speedup_vs_arm_ngram-g2_median = 0.7285208826719412
  speedup_vs_arm_ngram-g4_median = 0.6946533225826274
  speedup_vs_arm_ngram-g6_median = 0.6758982862096388
  speedup_vs_arm_qwen-g2_median = 0.7251111752060195
  speedup_vs_arm_qwen-g4_median = 0.6819157892172729
  speedup_vs_arm_qwen-g6_median = 0.6943722934058733
  speedup_vs_bandit_thompson_median = 0.6912369979125954
  speedup_vs_bandit_ucb1_median = 0.6714841180875957
  speedup_vs_hf_assisted_median = 2.0540785860951596
  speedup_vs_hf_greedy_median = 1.9916098580274422
  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=3e38251  n_samples=10
------------------------------------------------------------------------------
[arm_composite-g4]
  workload_tokens_per_second = 64.90609714173813
  tokens_per_second_median = 61.38411957244732
  tokens_per_second_p25 = 59.37850893945911
  tokens_per_second_p75 = 73.67784335943738
  tokens_per_second_mean = 66.41082359210698
  acceptance_rate_mean = 0.4941384849358883
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_ngram-g2_median = 1.0455319751029923
  speedup_vs_arm_ngram-g4_median = 1.0833400489849747
  speedup_vs_arm_ngram-g6_median = 1.0386703833861828
  speedup_vs_arm_qwen-g2_median = 0.9617002192383768
  speedup_vs_arm_qwen-g4_median = 0.9510551078698127
  speedup_vs_arm_qwen-g6_median = 1.022439658149655
  speedup_vs_bandit_thompson_median = 0.9474942530887054
  speedup_vs_bandit_ucb1_median = 0.9793276622376283
  speedup_vs_hf_assisted_median = 2.936118106765345
  speedup_vs_hf_sample_compiled_median = 1.27595439492818
  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 = 0.009819241018112325
  step_profile.bookkeeping_ms_mean = 0.4229722878930344
  step_profile.bookkeeping_ms_total = 193.29833556711674
  step_profile.draft_fraction = 0.503126395593414
  step_profile.draft_ms_mean = 21.67260404862056
  step_profile.draft_ms_total = 9904.380050219595
  step_profile.n_steps = 457.0
  step_profile.sample_fraction = 0.015877229086763128
  step_profile.sample_ms_mean = 0.6839253563327915
  step_profile.sample_ms_total = 312.5538878440857
  step_profile.step_ms_mean = 43.07586371623126
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.800875273522976
  step_profile.verify_fraction = 0.47117713430171054
  step_profile.verify_ms_mean = 20.29636202338488
  step_profile.verify_ms_total = 9275.43744468689
  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_ngram-g2]
  workload_tokens_per_second = 58.65310551171174
  tokens_per_second_median = 58.67232306034959
  tokens_per_second_p25 = 56.61859042731339
  tokens_per_second_p75 = 62.88100844076779
  tokens_per_second_mean = 59.00175018422495
  acceptance_rate_mean = 0.280431371633048
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.9564742210778587
  speedup_vs_arm_ngram-g4_median = 0.9641578695882684
  speedup_vs_arm_ngram-g6_median = 0.9744121495881803
  speedup_vs_arm_qwen-g2_median = 0.8795957604863619
  speedup_vs_arm_qwen-g4_median = 0.9803083852466049
  speedup_vs_arm_qwen-g6_median = 0.9215869661549685
  speedup_vs_bandit_thompson_median = 0.9576535797005039
  speedup_vs_bandit_ucb1_median = 0.8403227577284977
  speedup_vs_hf_assisted_median = 3.050969506028364
  speedup_vs_hf_sample_compiled_median = 1.207058880363171
  forward_counts.emitted_tokens_per_step = 1.2512218963831867
  forward_counts.n_generations = 10
  forward_counts.n_steps = 1023
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1033
  forward_counts.target_forwards_per_emitted_token = 0.80703125
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.018805921624985216
  step_profile.bookkeeping_ms_mean = 0.40019797472572743
  step_profile.bookkeeping_ms_total = 409.4025281444192
  step_profile.draft_fraction = 0.0045391816162181796
  step_profile.draft_ms_mean = 0.09659570671130019
  step_profile.draft_ms_total = 98.8174079656601
  step_profile.n_steps = 1023.0
  step_profile.sample_fraction = 0.03154657471033294
  step_profile.sample_ms_mean = 0.6713244668549456
  step_profile.sample_ms_total = 686.7649295926094
  step_profile.step_ms_mean = 21.280423406318544
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 1.2512218963831867
  step_profile.verify_fraction = 0.9451083220484636
  step_profile.verify_ms_mean = 20.11230525802657
  step_profile.verify_ms_total = 20574.88827896118
  host_syncs.by_kind = {'cuda': 97}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 95
  host_syncs.per_step = 1.0210526315789474
  host_syncs.prompt_id = HumanEval/23
  host_syncs.total = 97
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[arm_ngram-g4]
  workload_tokens_per_second = 61.691986617704465
  tokens_per_second_median = 63.928925339301884
  tokens_per_second_p25 = 56.82248640439265
  tokens_per_second_p75 = 68.04992639257938
  tokens_per_second_mean = 62.38694236908039
  acceptance_rate_mean = 0.21218822774959728
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.9238796471824042
  speedup_vs_arm_ngram-g2_median = 1.0372047953746297
  speedup_vs_arm_ngram-g6_median = 0.9906585090945623
  speedup_vs_arm_qwen-g2_median = 0.963683511572009
  speedup_vs_arm_qwen-g4_median = 1.0162238949986082
  speedup_vs_arm_qwen-g6_median = 0.942772796242722
  speedup_vs_bandit_thompson_median = 1.0380831919104616
  speedup_vs_bandit_ucb1_median = 0.9157700766280967
  speedup_vs_hf_assisted_median = 3.0643224104564837
  speedup_vs_hf_sample_compiled_median = 1.3252224720423307
  forward_counts.emitted_tokens_per_step = 1.3375130616509927
  forward_counts.n_generations = 10
  forward_counts.n_steps = 957
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 967
  forward_counts.target_forwards_per_emitted_token = 0.75546875
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.019067982040089807
  step_profile.bookkeeping_ms_mean = 0.41237928535372737
  step_profile.bookkeeping_ms_total = 394.6469760835171
  step_profile.draft_fraction = 0.004443389780706912
  step_profile.draft_ms_mean = 0.09609626747410872
  step_profile.draft_ms_total = 91.96412797272205
  step_profile.n_steps = 957.0
  step_profile.sample_fraction = 0.03174966746963608
  step_profile.sample_ms_mean = 0.6866434609503831
  step_profile.sample_ms_total = 657.1177921295166
  step_profile.step_ms_mean = 21.626792205211512
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 1.3375130616509927
  step_profile.verify_fraction = 0.9447389607095672
  step_profile.verify_ms_mean = 20.43167319143329
  step_profile.verify_ms_total = 19553.11124420166
  host_syncs.by_kind = {'cuda': 98}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 96
  host_syncs.per_step = 1.0208333333333333
  host_syncs.prompt_id = HumanEval/23
  host_syncs.total = 98
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[arm_ngram-g6]
  workload_tokens_per_second = 62.22049690271739
  tokens_per_second_median = 63.69478573406319
  tokens_per_second_p25 = 56.58972003894397
  tokens_per_second_p75 = 70.27111966446452
  tokens_per_second_mean = 63.19241243571332
  acceptance_rate_mean = 0.14082987362074137
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.9630897572852135
  speedup_vs_arm_ngram-g2_median = 1.026282266532098
  speedup_vs_arm_ngram-g4_median = 1.0094542297307847
  speedup_vs_arm_qwen-g2_median = 0.9572959531709487
  speedup_vs_arm_qwen-g4_median = 1.0362596916545992
  speedup_vs_arm_qwen-g6_median = 0.9637263453593173
  speedup_vs_bandit_thompson_median = 1.0018063272599322
  speedup_vs_bandit_ucb1_median = 0.9383767293696408
  speedup_vs_hf_assisted_median = 3.116477471715709
  speedup_vs_hf_sample_compiled_median = 1.3190885640113046
  forward_counts.emitted_tokens_per_step = 1.3631522896698616
  forward_counts.n_generations = 10
  forward_counts.n_steps = 939
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 949
  forward_counts.target_forwards_per_emitted_token = 0.74140625
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.019101198017491612
  step_profile.bookkeeping_ms_mean = 0.41744115903449897
  step_profile.bookkeeping_ms_total = 391.9772483333945
  step_profile.draft_fraction = 0.004406744656218483
  step_profile.draft_ms_mean = 0.09630582307855161
  step_profile.draft_ms_total = 90.43116787075996
  step_profile.n_steps = 939.0
  step_profile.sample_fraction = 0.03197760449202805
  step_profile.sample_ms_mean = 0.698844557816482
  step_profile.sample_ms_total = 656.2150397896767
  step_profile.step_ms_mean = 21.854187295070915
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 1.3631522896698616
  step_profile.verify_fraction = 0.9445144528342618
  step_profile.verify_ms_mean = 20.641595755141385
  step_profile.verify_ms_total = 19382.45841407776
  host_syncs.by_kind = {'cuda': 108}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 106
  host_syncs.per_step = 1.0188679245283019
  host_syncs.prompt_id = HumanEval/23
  host_syncs.total = 108
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[arm_qwen-g2]
  workload_tokens_per_second = 66.01641305074533
  tokens_per_second_median = 66.7044283572397
  tokens_per_second_p25 = 65.0786613358389
  tokens_per_second_p75 = 68.6547455219794
  tokens_per_second_mean = 66.21765690869812
  acceptance_rate_mean = 0.8910045754674323
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 1.0402527412844202
  speedup_vs_arm_ngram-g2_median = 1.136939108540783
  speedup_vs_arm_ngram-g4_median = 1.0378546797458905
  speedup_vs_arm_ngram-g6_median = 1.0468358126524453
  speedup_vs_arm_qwen-g4_median = 1.0726597806203944
  speedup_vs_arm_qwen-g6_median = 1.0326401943098875
  speedup_vs_bandit_thompson_median = 1.0653517461673745
  speedup_vs_bandit_ucb1_median = 1.016175083265721
  speedup_vs_hf_assisted_median = 3.086841631122099
  speedup_vs_hf_sample_compiled_median = 1.3955609719896538
  forward_counts.emitted_tokens_per_step = 2.735042735042735
  forward_counts.n_generations = 10
  forward_counts.n_steps = 468
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 478
  forward_counts.target_forwards_per_emitted_token = 0.3734375
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.00688087737397132
  step_profile.bookkeeping_ms_mean = 0.2847448884167223
  step_profile.bookkeeping_ms_total = 133.26060777902603
  step_profile.draft_fraction = 0.5113058105819529
  step_profile.draft_ms_mean = 21.158888331845276
  step_profile.draft_ms_total = 9902.359739303589
  step_profile.n_steps = 468.0
  step_profile.sample_fraction = 0.0096216182024805
  step_profile.sample_ms_mean = 0.3981623930426744
  step_profile.sample_ms_total = 186.33999994397163
  step_profile.step_ms_mean = 41.382061173454815
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.735042735042735
  step_profile.verify_fraction = 0.4721916938415953
  step_profile.verify_ms_mean = 19.540265560150146
  step_profile.verify_ms_total = 9144.844282150269
  host_syncs.by_kind = {'cuda': 51}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 48
  host_syncs.per_step = 1.0625
  host_syncs.prompt_id = HumanEval/23
  host_syncs.total = 51
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[arm_qwen-g4]
  workload_tokens_per_second = 61.98276395241417
  tokens_per_second_median = 63.71127692658077
  tokens_per_second_p25 = 58.11534419474769
  tokens_per_second_p75 = 66.13787521133267
  tokens_per_second_mean = 62.35630962897553
  acceptance_rate_mean = 0.6973848833657661
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 1.0518028663206136
  speedup_vs_arm_ngram-g2_median = 1.020087728461225
  speedup_vs_arm_ngram-g4_median = 0.9840509137353408
  speedup_vs_arm_ngram-g6_median = 0.9663259911011961
  speedup_vs_arm_qwen-g2_median = 0.9322620783669084
  speedup_vs_arm_qwen-g6_median = 0.9665816301645699
  speedup_vs_bandit_thompson_median = 0.9813704685936069
  speedup_vs_bandit_ucb1_median = 0.9184574373016158
  speedup_vs_hf_assisted_median = 2.882064158719155
  speedup_vs_hf_sample_compiled_median = 1.315751794259138
  forward_counts.emitted_tokens_per_step = 3.764705882352941
  forward_counts.n_generations = 10
  forward_counts.n_steps = 340
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 350
  forward_counts.target_forwards_per_emitted_token = 0.2734375
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.006952291966913971
  step_profile.bookkeeping_ms_mean = 0.42191011708448917
  step_profile.bookkeeping_ms_total = 143.4494398087263
  step_profile.draft_fraction = 0.6567551230463572
  step_profile.draft_ms_mean = 39.85615566478056
  step_profile.draft_ms_total = 13551.09292602539
  step_profile.n_steps = 340.0
  step_profile.sample_fraction = 0.0066760408875970135
  step_profile.sample_ms_mean = 0.40514541189460196
  step_profile.sample_ms_total = 137.74944004416466
  step_profile.step_ms_mean = 60.68647851562675
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 3.764705882352941
  step_profile.verify_fraction = 0.3296165440991318
  step_profile.verify_ms_mean = 20.003267321867103
  step_profile.verify_ms_total = 6801.110889434814
  host_syncs.by_kind = {'cuda': 38}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 35
  host_syncs.per_step = 1.0857142857142856
  host_syncs.prompt_id = HumanEval/23
  host_syncs.total = 38
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[arm_qwen-g6]
  workload_tokens_per_second = 62.229707087363245
  tokens_per_second_median = 63.97123185805204
  tokens_per_second_p25 = 57.52343781233637
  tokens_per_second_p75 = 70.90075900472786
  tokens_per_second_mean = 63.42893619086946
  acceptance_rate_mean = 0.6458299873131749
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.9786183545933012
  speedup_vs_arm_ngram-g2_median = 1.0851159001120503
  speedup_vs_arm_ngram-g4_median = 1.0609035995475815
  speedup_vs_arm_ngram-g6_median = 1.037953968384869
  speedup_vs_arm_qwen-g2_median = 0.9684715321346467
  speedup_vs_arm_qwen-g4_median = 1.0356266668397354
  speedup_vs_bandit_thompson_median = 1.0143568388008104
  speedup_vs_bandit_ucb1_median = 1.01404504387099
  speedup_vs_hf_assisted_median = 3.0873308586019617
  speedup_vs_hf_sample_compiled_median = 1.321506744742217
  forward_counts.emitted_tokens_per_step = 5.0
  forward_counts.n_generations = 10
  forward_counts.n_steps = 256
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 266
  forward_counts.target_forwards_per_emitted_token = 0.2078125
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.005609945308185002
  step_profile.bookkeeping_ms_mean = 0.45044350012904033
  step_profile.bookkeeping_ms_total = 115.31353603303432
  step_profile.draft_fraction = 0.7311701421979092
  step_profile.draft_ms_mean = 58.70838661491871
  step_profile.draft_ms_total = 15029.34697341919
  step_profile.n_steps = 256.0
  step_profile.sample_fraction = 0.005138762878433383
  step_profile.sample_ms_mean = 0.41261049977038056
  step_profile.sample_ms_total = 105.62828794121742
  step_profile.step_ms_mean = 80.29374180739978
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 5.0
  step_profile.verify_fraction = 0.2580811496154724
  step_profile.verify_ms_mean = 20.722301192581654
  step_profile.verify_ms_total = 5304.909105300903
  host_syncs.by_kind = {'cuda': 37}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 34
  host_syncs.per_step = 1.088235294117647
  host_syncs.prompt_id = HumanEval/23
  host_syncs.total = 37
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[bandit_thompson]
  workload_tokens_per_second = 62.23088804278899
  tokens_per_second_median = 62.731496813748585
  tokens_per_second_p25 = 61.024757808559194
  tokens_per_second_p75 = 65.1290368779959
  tokens_per_second_mean = 62.46083239324331
  acceptance_rate_mean = 0.6881358856734927
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 1.0554202838781526
  speedup_vs_arm_ngram-g2_median = 1.0442196091241338
  speedup_vs_arm_ngram-g4_median = 0.9645191034180076
  speedup_vs_arm_ngram-g6_median = 0.9983324566644495
  speedup_vs_arm_qwen-g2_median = 0.9386618985423247
  speedup_vs_arm_qwen-g4_median = 1.0190211663438187
  speedup_vs_arm_qwen-g6_median = 0.987120084203958
  speedup_vs_bandit_ucb1_median = 0.9222310558355244
  speedup_vs_hf_assisted_median = 2.9934316537912498
  speedup_vs_hf_sample_compiled_median = 1.2983044723871024
  forward_counts.emitted_tokens_per_step = 3.1604938271604937
  forward_counts.n_generations = 10
  forward_counts.n_steps = 405
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 415
  forward_counts.target_forwards_per_emitted_token = 0.32421875
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.00650760841816977
  step_profile.bookkeeping_ms_mean = 0.3299729378135116
  step_profile.bookkeeping_ms_total = 133.6390398144722
  step_profile.draft_fraction = 0.582898457549746
  step_profile.draft_ms_mean = 29.556283065161587
  step_profile.draft_ms_total = 11970.294641390443
  step_profile.n_steps = 405.0
  step_profile.sample_fraction = 0.013344143844436347
  step_profile.sample_ms_mean = 0.6766243547569086
  step_profile.sample_ms_total = 274.032863676548
  step_profile.step_ms_mean = 50.70571500463986
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 3.1604938271604937
  step_profile.verify_fraction = 0.3972497901876479
  step_profile.verify_ms_mean = 20.142834646907854
  step_profile.verify_ms_total = 8157.848031997681
  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 = HumanEval/23
  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
[bandit_ucb1]
  workload_tokens_per_second = 65.92538042819774
  tokens_per_second_median = 66.33631628287489
  tokens_per_second_p25 = 62.74199531303853
  tokens_per_second_p75 = 72.87954147479579
  tokens_per_second_mean = 66.71165570998342
  acceptance_rate_mean = 0.5867531752421452
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 1.0214617597987723
  speedup_vs_arm_ngram-g2_median = 1.190203038504812
  speedup_vs_arm_ngram-g4_median = 1.0921839456896882
  speedup_vs_arm_ngram-g6_median = 1.066175924156921
  speedup_vs_arm_qwen-g2_median = 0.984196177603291
  speedup_vs_arm_qwen-g4_median = 1.0887857860878098
  speedup_vs_arm_qwen-g6_median = 0.9874061143690123
  speedup_vs_bandit_thompson_median = 1.0843569862861617
  speedup_vs_hf_assisted_median = 3.2848890197488845
  speedup_vs_hf_sample_compiled_median = 1.3771251397170552
  forward_counts.emitted_tokens_per_step = 2.694736842105263
  forward_counts.n_generations = 10
  forward_counts.n_steps = 475
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 485
  forward_counts.target_forwards_per_emitted_token = 0.37890625
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.009680615760300704
  step_profile.bookkeeping_ms_mean = 0.3949881265351647
  step_profile.bookkeeping_ms_total = 187.61936010420322
  step_profile.draft_fraction = 0.4703353280191564
  step_profile.draft_ms_mean = 19.190604674078916
  step_profile.draft_ms_total = 9115.537220187485
  step_profile.n_steps = 475.0
  step_profile.sample_fraction = 0.016920439121435954
  step_profile.sample_ms_mean = 0.6903871317913658
  step_profile.sample_ms_total = 327.93388760089874
  step_profile.step_ms_mean = 40.8019630481538
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.694736842105263
  step_profile.verify_fraction = 0.503063617099107
  step_profile.verify_ms_mean = 20.525983115748357
  step_profile.verify_ms_total = 9749.841979980469
  host_syncs.by_kind = {'cuda': 47}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 43
  host_syncs.per_step = 1.0930232558139534
  host_syncs.prompt_id = HumanEval/23
  host_syncs.total = 47
  nonfinite_logits.prompt_id = HumanEval/23
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[hf_assisted]
  workload_tokens_per_second = 17.86888152245691
  tokens_per_second_median = 20.857748180247196
  tokens_per_second_p25 = 19.589787170718004
  tokens_per_second_p75 = 22.853278889534657
  tokens_per_second_mean = 20.078133450969734
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.3406177223904664
  speedup_vs_arm_ngram-g2_median = 0.32776615096494227
  speedup_vs_arm_ngram-g4_median = 0.32661431013663467
  speedup_vs_arm_ngram-g6_median = 0.3209179900022727
  speedup_vs_arm_qwen-g2_median = 0.32401062353156057
  speedup_vs_arm_qwen-g4_median = 0.3469832385991153
  speedup_vs_arm_qwen-g6_median = 0.3240047191122416
  speedup_vs_bandit_thompson_median = 0.3343156566594857
  speedup_vs_bandit_ucb1_median = 0.3045711377582715
  speedup_vs_hf_sample_compiled_median = 0.4467569874192664
  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_sample_compiled]
  workload_tokens_per_second = 47.91703061211116
  tokens_per_second_median = 48.14217810350597
  tokens_per_second_p25 = 47.90795300294086
  tokens_per_second_p75 = 48.43661782755615
  tokens_per_second_mean = 47.93407489267826
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.7837277161429781
  speedup_vs_arm_ngram-g2_median = 0.8284647984037778
  speedup_vs_arm_ngram-g4_median = 0.7553876408264177
  speedup_vs_arm_ngram-g6_median = 0.7601762839808305
  speedup_vs_arm_qwen-g2_median = 0.7166658799934875
  speedup_vs_arm_qwen-g4_median = 0.7600222101481766
  speedup_vs_arm_qwen-g6_median = 0.7567127421184678
  speedup_vs_bandit_thompson_median = 0.7702822534488123
  speedup_vs_bandit_ucb1_median = 0.7261529161173583
  speedup_vs_hf_assisted_median = 2.241145050164943
  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=3e38251  n_samples=10
------------------------------------------------------------------------------
[arm_composite-g4]
  workload_tokens_per_second = 135.4730685282579
  tokens_per_second_median = 156.40152187902373
  tokens_per_second_p25 = 138.2962045298195
  tokens_per_second_p75 = 171.48758909319045
  tokens_per_second_mean = 145.18966881822317
  acceptance_rate_mean = 0.6083112276000873
  gpu_memory_gb = 4.036861896514893
  speedup_vs_arm_ngram-g2_median = 1.2810704936821962
  speedup_vs_arm_ngram-g4_median = 0.9486437114662862
  speedup_vs_arm_ngram-g6_median = 0.8079764070557491
  speedup_vs_arm_qwen-g2_median = 2.202086398391023
  speedup_vs_arm_qwen-g4_median = 2.0096379006419323
  speedup_vs_arm_qwen-g6_median = 1.9329901384179908
  speedup_vs_bandit_thompson_median = 0.9892250665912172
  speedup_vs_bandit_ucb1_median = 1.119419374819651
  speedup_vs_hf_assisted_median = 5.857172095712367
  speedup_vs_hf_greedy_compiled_median = 3.2192940820341276
  speedup_vs_hf_greedy_median = 6.39236365929918
  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 = 0.01086157605215803
  step_profile.bookkeeping_ms_mean = 0.2897665263826381
  step_profile.bookkeeping_ms_total = 102.28758381307125
  step_profile.draft_fraction = 0.16927458815945787
  step_profile.draft_ms_mean = 4.51592928874003
  step_profile.draft_ms_total = 1594.1230389252305
  step_profile.n_steps = 353.0
  step_profile.sample_fraction = 0.016792703705558902
  step_profile.sample_ms_mean = 0.4479979146641307
  step_profile.sample_ms_total = 158.14326387643814
  step_profile.step_ms_mean = 26.678128937380677
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 3.6260623229461757
  step_profile.verify_fraction = 0.8030711320828252
  step_profile.verify_ms_mean = 21.42443520759388
  step_profile.verify_ms_total = 7562.82562828064
  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_ngram-g2]
  workload_tokens_per_second = 107.22288740551457
  tokens_per_second_median = 118.94347347325188
  tokens_per_second_p25 = 108.03224976796909
  tokens_per_second_p75 = 124.98444909046255
  tokens_per_second_mean = 110.97962059361437
  acceptance_rate_mean = 0.646846040861779
  gpu_memory_gb = 4.036861896514893
  speedup_vs_arm_composite-g4_median = 0.7805972386982294
  speedup_vs_arm_ngram-g4_median = 0.7334895586787906
  speedup_vs_arm_ngram-g6_median = 0.5837709234989132
  speedup_vs_arm_qwen-g2_median = 1.6565918151182137
  speedup_vs_arm_qwen-g4_median = 1.4856476286900415
  speedup_vs_arm_qwen-g6_median = 1.473401827892109
  speedup_vs_bandit_thompson_median = 0.7515333469601206
  speedup_vs_bandit_ucb1_median = 0.8555290246151935
  speedup_vs_hf_assisted_median = 4.5721995564250015
  speedup_vs_hf_greedy_compiled_median = 2.451265638400241
  speedup_vs_hf_greedy_median = 4.899876131437265
  forward_counts.emitted_tokens_per_step = 2.3104693140794224
  forward_counts.n_generations = 10
  forward_counts.n_steps = 554
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 564
  forward_counts.target_forwards_per_emitted_token = 0.440625
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.012034436584437825
  step_profile.bookkeeping_ms_mean = 0.2584757259122301
  step_profile.bookkeeping_ms_total = 143.19555215537548
  step_profile.draft_fraction = 0.004440992947241143
  step_profile.draft_ms_mean = 0.095383682298316
  step_profile.draft_ms_total = 52.84255999326706
  step_profile.n_steps = 554.0
  step_profile.sample_fraction = 0.020319347946046385
  step_profile.sample_ms_mean = 0.43641911888165597
  step_profile.sample_ms_total = 241.7761918604374
  step_profile.step_ms_mean = 21.478008056187242
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.3104693140794224
  step_profile.verify_fraction = 0.9632052225222747
  step_profile.verify_ms_mean = 20.68772952909504
  step_profile.verify_ms_total = 11461.002159118652
  host_syncs.by_kind = {'cuda': 81}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 79
  host_syncs.per_step = 1.0253164556962024
  host_syncs.prompt_id = HumanEval/0:docstring
  host_syncs.total = 81
  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_ngram-g4]
  workload_tokens_per_second = 140.16742181403353
  tokens_per_second_median = 165.6045867977912
  tokens_per_second_p25 = 147.16558897515807
  tokens_per_second_p75 = 174.42453258264564
  tokens_per_second_mean = 151.9288701767842
  acceptance_rate_mean = 0.5346198877724118
  gpu_memory_gb = 4.036861896514893
  speedup_vs_arm_composite-g4_median = 1.0541398633431585
  speedup_vs_arm_ngram-g2_median = 1.3633495551911534
  speedup_vs_arm_ngram-g6_median = 0.8405527541523126
  speedup_vs_arm_qwen-g2_median = 2.3055026651360313
  speedup_vs_arm_qwen-g4_median = 2.0679928555985194
  speedup_vs_arm_qwen-g6_median = 2.044641448030932
  speedup_vs_bandit_thompson_median = 1.019657232324922
  speedup_vs_bandit_ucb1_median = 1.1765458520678385
  speedup_vs_hf_assisted_median = 6.370395596070457
  speedup_vs_hf_greedy_compiled_median = 3.4246697957203653
  speedup_vs_hf_greedy_median = 6.821336198027975
  forward_counts.emitted_tokens_per_step = 3.12958435207824
  forward_counts.n_generations = 10
  forward_counts.n_steps = 409
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 419
  forward_counts.target_forwards_per_emitted_token = 0.32734375
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.013281348540127854
  step_profile.bookkeeping_ms_mean = 0.2954315112170497
  step_profile.bookkeeping_ms_total = 120.83148808777332
  step_profile.draft_fraction = 0.004340993386481409
  step_profile.draft_ms_mean = 0.09656144724134127
  step_profile.draft_ms_total = 39.493631921708584
  step_profile.n_steps = 409.0
  step_profile.sample_fraction = 0.020166371131946113
  step_profile.sample_ms_mean = 0.44858257286706005
  step_profile.sample_ms_total = 183.47027230262756
  step_profile.step_ms_mean = 22.244089922378144
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 3.12958435207824
  step_profile.verify_fraction = 0.9622112869414446
  step_profile.verify_ms_mean = 21.403514391052695
  step_profile.verify_ms_total = 8754.037385940552
  host_syncs.by_kind = {'cuda': 71}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 69
  host_syncs.per_step = 1.0289855072463767
  host_syncs.prompt_id = HumanEval/0:docstring
  host_syncs.total = 71
  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_ngram-g6]
  workload_tokens_per_second = 161.05633283319682
  tokens_per_second_median = 196.21647394033454
  tokens_per_second_p25 = 174.74587446960984
  tokens_per_second_p75 = 223.1332216463316
  tokens_per_second_mean = 183.7334737938549
  acceptance_rate_mean = 0.453086699125582
  gpu_memory_gb = 4.036861896514893
  speedup_vs_arm_composite-g4_median = 1.2376620037074113
  speedup_vs_arm_ngram-g2_median = 1.7130036927991474
  speedup_vs_arm_ngram-g4_median = 1.1897193417362895
  speedup_vs_arm_qwen-g2_median = 2.731303368041459
  speedup_vs_arm_qwen-g4_median = 2.4827439656260104
  speedup_vs_arm_qwen-g6_median = 2.495728899151784
  speedup_vs_bandit_thompson_median = 1.2128593580413793
  speedup_vs_bandit_ucb1_median = 1.3906509347193705
  speedup_vs_hf_assisted_median = 7.918515214864707
  speedup_vs_hf_greedy_compiled_median = 4.049257910430214
  speedup_vs_hf_greedy_median = 8.081969791836464
  forward_counts.emitted_tokens_per_step = 3.646723646723647
  forward_counts.n_generations = 10
  forward_counts.n_steps = 351
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 361
  forward_counts.target_forwards_per_emitted_token = 0.28203125
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.014360251754224737
  step_profile.bookkeeping_ms_mean = 0.3239112026840873
  step_profile.bookkeeping_ms_total = 113.69283214211464
  step_profile.draft_fraction = 0.004074957238874327
  step_profile.draft_ms_mean = 0.09191512257030547
  step_profile.draft_ms_total = 32.26220802217722
  step_profile.n_steps = 351.0
  step_profile.sample_fraction = 0.020183906341539364
  step_profile.sample_ms_mean = 0.4552701087588938
  step_profile.sample_ms_total = 159.79980817437172
  step_profile.step_ms_mean = 22.556094992470708
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 3.646723646723647
  step_profile.verify_fraction = 0.9613808846653615
  step_profile.verify_ms_mean = 21.68499855845742
  step_profile.verify_ms_total = 7611.434494018555
  host_syncs.by_kind = {'cuda': 68}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 66
  host_syncs.per_step = 1.0303030303030303
  host_syncs.prompt_id = HumanEval/0:docstring
  host_syncs.total = 68
  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_qwen-g2]
  workload_tokens_per_second = 70.2850088986192
  tokens_per_second_median = 71.01497473207606
  tokens_per_second_p25 = 70.79375799978682
  tokens_per_second_p75 = 71.3513080044732
  tokens_per_second_mean = 70.3668844309468
  acceptance_rate_mean = 0.9374125321488126
  gpu_memory_gb = 4.036861896514893
  speedup_vs_arm_composite-g4_median = 0.4546316949418633
  speedup_vs_arm_ngram-g2_median = 0.6036542593883256
  speedup_vs_arm_ngram-g4_median = 0.4340241953756651
  speedup_vs_arm_ngram-g6_median = 0.366562810569962
  speedup_vs_arm_qwen-g4_median = 0.9243322810289648
  speedup_vs_arm_qwen-g6_median = 0.912775345717926
  speedup_vs_bandit_thompson_median = 0.44284541100285724
  speedup_vs_bandit_ucb1_median = 0.5164269106684851
  speedup_vs_hf_assisted_median = 2.8218030302057637
  speedup_vs_hf_greedy_compiled_median = 1.4639920961656951
  speedup_vs_hf_greedy_median = 2.9164701571886322
  forward_counts.emitted_tokens_per_step = 2.889390519187359
  forward_counts.n_generations = 10
  forward_counts.n_steps = 443
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 453
  forward_counts.target_forwards_per_emitted_token = 0.35390625
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.006136481064441366
  step_profile.bookkeeping_ms_mean = 0.2519519638194457
  step_profile.bookkeeping_ms_total = 111.61471997201443
  step_profile.draft_fraction = 0.5114953536865975
  step_profile.draft_ms_mean = 21.001003261075333
  step_profile.draft_ms_total = 9303.444444656372
  step_profile.n_steps = 443.0
  step_profile.sample_fraction = 0.0037505606621601874
  step_profile.sample_ms_mean = 0.15399071786123797
  step_profile.sample_ms_total = 68.21788801252842
  step_profile.step_ms_mean = 41.05805284390331
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.889390519187359
  step_profile.verify_fraction = 0.478617604586801
  step_profile.verify_ms_mean = 19.651106901147294
  step_profile.verify_ms_total = 8705.440357208252
  host_syncs.by_kind = {'cuda': 49}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 46
  host_syncs.per_step = 1.065217391304348
  host_syncs.prompt_id = HumanEval/0:docstring
  host_syncs.total = 49
  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_qwen-g4]
  workload_tokens_per_second = 75.402769143059
  tokens_per_second_median = 76.81742210217215
  tokens_per_second_p25 = 76.37863610357228
  tokens_per_second_p75 = 78.97961442266129
  tokens_per_second_mean = 75.76200084347465
  acceptance_rate_mean = 0.8972649506653644
  gpu_memory_gb = 4.036861896514893
  speedup_vs_arm_composite-g4_median = 0.49921509173212375
  speedup_vs_arm_ngram-g2_median = 0.6731168793710888
  speedup_vs_arm_ngram-g4_median = 0.4840600653983943
  speedup_vs_arm_ngram-g6_median = 0.4030868901337754
  speedup_vs_arm_qwen-g2_median = 1.081866958547543
  speedup_vs_arm_qwen-g6_median = 0.9861643667228526
  speedup_vs_bandit_thompson_median = 0.4992953058683742
  speedup_vs_bandit_ucb1_median = 0.575895242526753
  speedup_vs_hf_assisted_median = 3.0799822909068766
  speedup_vs_hf_greedy_compiled_median = 1.586580329509323
  speedup_vs_hf_greedy_median = 3.155964314330828
  forward_counts.emitted_tokens_per_step = 4.63768115942029
  forward_counts.n_generations = 10
  forward_counts.n_steps = 276
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 286
  forward_counts.target_forwards_per_emitted_token = 0.2234375
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.004877206876897133
  step_profile.bookkeeping_ms_mean = 0.2997189563145672
  step_profile.bookkeeping_ms_total = 82.72243194282055
  step_profile.draft_fraction = 0.6589531336312708
  step_profile.draft_ms_mean = 40.494641801585324
  step_profile.draft_ms_total = 11176.521137237549
  step_profile.n_steps = 276.0
  step_profile.sample_fraction = 0.0026620465278363935
  step_profile.sample_ms_mean = 0.16359072459348734
  step_profile.sample_ms_total = 45.151039987802505
  step_profile.step_ms_mean = 61.45299223092371
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 4.63768115942029
  step_profile.verify_fraction = 0.3335076129639956
  step_profile.verify_ms_mean = 20.495040748430334
  step_profile.verify_ms_total = 5656.6312465667725
  host_syncs.by_kind = {'cuda': 34}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 31
  host_syncs.per_step = 1.096774193548387
  host_syncs.prompt_id = HumanEval/0:docstring
  host_syncs.total = 34
  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_qwen-g6]
  workload_tokens_per_second = 75.37199894833218
  tokens_per_second_median = 78.6346030970967
  tokens_per_second_p25 = 76.21173815139213
  tokens_per_second_p75 = 80.82350103513608
  tokens_per_second_mean = 76.18621405348891
  acceptance_rate_mean = 0.8530619266436459
  gpu_memory_gb = 4.036861896514893
  speedup_vs_arm_composite-g4_median = 0.5187525139172283
  speedup_vs_arm_ngram-g2_median = 0.6792803298723868
  speedup_vs_arm_ngram-g4_median = 0.4901588139845191
  speedup_vs_arm_ngram-g6_median = 0.40133047002554983
  speedup_vs_arm_qwen-g2_median = 1.095567778168658
  speedup_vs_arm_qwen-g4_median = 1.0140686747619814
  speedup_vs_bandit_thompson_median = 0.5159174958093276
  speedup_vs_bandit_ucb1_median = 0.5811523721778086
  speedup_vs_hf_assisted_median = 3.0874184900837784
  speedup_vs_hf_greedy_compiled_median = 1.617475519780239
  speedup_vs_hf_greedy_median = 3.226762382049218
  forward_counts.emitted_tokens_per_step = 6.124401913875598
  forward_counts.n_generations = 10
  forward_counts.n_steps = 209
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 219
  forward_counts.target_forwards_per_emitted_token = 0.17109375
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.004098205821138892
  step_profile.bookkeeping_ms_mean = 0.3327516169068916
  step_profile.bookkeeping_ms_total = 69.54508793354034
  step_profile.draft_fraction = 0.73251094453661
  step_profile.draft_ms_mean = 59.475832067608266
  step_profile.draft_ms_total = 12430.448902130127
  step_profile.n_steps = 209.0
  step_profile.sample_fraction = 0.0021513070718793916
  step_profile.sample_ms_mean = 0.17467422034751856
  step_profile.sample_ms_total = 36.50691205263138
  step_profile.step_ms_mean = 81.19446202299811
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 6.124401913875598
  step_profile.verify_fraction = 0.2612395425703717
  step_profile.verify_ms_mean = 21.21120411813544
  step_profile.verify_ms_total = 4433.141660690308
  host_syncs.by_kind = {'cuda': 27}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 24
  host_syncs.per_step = 1.125
  host_syncs.prompt_id = HumanEval/0:docstring
  host_syncs.total = 27
  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
[bandit_thompson]
  workload_tokens_per_second = 135.98638800139773
  tokens_per_second_median = 160.55168910381917
  tokens_per_second_p25 = 142.94596639685577
  tokens_per_second_p75 = 170.21935027115316
  tokens_per_second_mean = 147.18879317069758
  acceptance_rate_mean = 0.598211778778353
  gpu_memory_gb = 4.036861896514893
  speedup_vs_arm_composite-g4_median = 1.0112843113709773
  speedup_vs_arm_ngram-g2_median = 1.330613516432635
  speedup_vs_arm_ngram-g4_median = 0.9807281156449761
  speedup_vs_arm_ngram-g6_median = 0.8244979883318773
  speedup_vs_arm_qwen-g2_median = 2.2592259469681535
  speedup_vs_arm_qwen-g4_median = 2.003449810367812
  speedup_vs_arm_qwen-g6_median = 1.938998485773831
  speedup_vs_bandit_ucb1_median = 1.142979760796985
  speedup_vs_hf_assisted_median = 6.0215290316677645
  speedup_vs_hf_greedy_compiled_median = 3.319774662049733
  speedup_vs_hf_greedy_median = 6.5885090105944215
  forward_counts.emitted_tokens_per_step = 3.2160804020100504
  forward_counts.n_generations = 10
  forward_counts.n_steps = 398
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 408
  forward_counts.target_forwards_per_emitted_token = 0.31875
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.012468125637197031
  step_profile.bookkeeping_ms_mean = 0.29384948768133495
  step_profile.bookkeeping_ms_total = 116.9520960971713
  step_profile.draft_fraction = 0.05288847396323477
  step_profile.draft_ms_mean = 1.2464785349915675
  step_profile.draft_ms_total = 496.09845692664385
  step_profile.n_steps = 398.0
  step_profile.sample_fraction = 0.019034100058668707
  step_profile.sample_ms_mean = 0.44859674288399853
  step_profile.sample_ms_total = 178.54150366783142
  step_profile.step_ms_mean = 23.568056356817035
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 3.2160804020100504
  step_profile.verify_fraction = 0.9156093003408995
  step_profile.verify_ms_mean = 21.579131591260133
  step_profile.verify_ms_total = 8588.494373321533
  host_syncs.by_kind = {'cuda': 55}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 51
  host_syncs.per_step = 1.0784313725490196
  host_syncs.prompt_id = HumanEval/0:docstring
  host_syncs.total = 55
  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
[bandit_ucb1]
  workload_tokens_per_second = 122.92800505386738
  tokens_per_second_median = 139.06426478860618
  tokens_per_second_p25 = 120.52435037865631
  tokens_per_second_p75 = 156.02928457588297
  tokens_per_second_mean = 132.03335694863608
  acceptance_rate_mean = 0.7128091818688113
  gpu_memory_gb = 4.036861896514893
  speedup_vs_arm_composite-g4_median = 0.8933405458003466
  speedup_vs_arm_ngram-g2_median = 1.1690100139192152
  speedup_vs_arm_ngram-g4_median = 0.849973460133596
  speedup_vs_arm_ngram-g6_median = 0.7191289858677279
  speedup_vs_arm_qwen-g2_median = 1.9365093148192636
  speedup_vs_arm_qwen-g4_median = 1.7368099845502671
  speedup_vs_arm_qwen-g6_median = 1.7235522827440812
  speedup_vs_bandit_thompson_median = 0.8749410166962479
  speedup_vs_hf_assisted_median = 5.379776317177139
  speedup_vs_hf_greedy_compiled_median = 2.875693064449667
  speedup_vs_hf_greedy_median = 5.728508220821256
  forward_counts.emitted_tokens_per_step = 4.050632911392405
  forward_counts.n_generations = 10
  forward_counts.n_steps = 316
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 326
  forward_counts.target_forwards_per_emitted_token = 0.2546875
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.009151182417378027
  step_profile.bookkeeping_ms_mean = 0.30067058265963686
  step_profile.bookkeeping_ms_total = 95.01190412044525
  step_profile.draft_fraction = 0.3183855911609641
  step_profile.draft_ms_mean = 10.460853782458868
  step_profile.draft_ms_total = 3305.629795257002
  step_profile.n_steps = 316.0
  step_profile.sample_fraction = 0.013777212462155691
  step_profile.sample_ms_mean = 0.4526630887125112
  step_profile.sample_ms_total = 143.04153603315353
  step_profile.step_ms_mean = 32.85592713010132
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 4.050632911392405
  step_profile.verify_fraction = 0.6586860139595022
  step_profile.verify_ms_mean = 21.641739676270305
  step_profile.verify_ms_total = 6838.789737701416
  host_syncs.by_kind = {'cuda': 48}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 44
  host_syncs.per_step = 1.0909090909090908
  host_syncs.prompt_id = HumanEval/0:docstring
  host_syncs.total = 48
  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_assisted]
  workload_tokens_per_second = 24.294584749008994
  tokens_per_second_median = 25.795452853006076
  tokens_per_second_p25 = 24.00299348874271
  tokens_per_second_p75 = 27.004788299700607
  tokens_per_second_mean = 24.69346989487415
  gpu_memory_gb = 4.036861896514893
  speedup_vs_arm_composite-g4_median = 0.17079924237356303
  speedup_vs_arm_ngram-g2_median = 0.21879524116874838
  speedup_vs_arm_ngram-g4_median = 0.15725368155521208
  speedup_vs_arm_ngram-g6_median = 0.1263650782090281
  speedup_vs_arm_qwen-g2_median = 0.3544971168640578
  speedup_vs_arm_qwen-g4_median = 0.32502888721678524
  speedup_vs_arm_qwen-g6_median = 0.3239017185929489
  speedup_vs_bandit_thompson_median = 0.166082999680698
  speedup_vs_bandit_ucb1_median = 0.18621459761484227
  speedup_vs_hf_greedy_compiled_median = 0.5345302311151681
  speedup_vs_hf_greedy_median = 1.0592687212546208
  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
  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 = 24.38033522643374
  tokens_per_second_median = 24.34964439355547
  tokens_per_second_p25 = 24.22611841156104
  tokens_per_second_p75 = 24.503973595778284
  tokens_per_second_mean = 24.381621492090236
  gpu_memory_gb = 4.036861896514893
  speedup_vs_arm_composite-g4_median = 0.15654368900763838
  speedup_vs_arm_ngram-g2_median = 0.20409974348036158
  speedup_vs_arm_ngram-g4_median = 0.14679199281077637
  speedup_vs_arm_ngram-g6_median = 0.12398817964930187
  speedup_vs_arm_qwen-g2_median = 0.342883612937618
  speedup_vs_arm_qwen-g4_median = 0.31687063896539874
  speedup_vs_arm_qwen-g6_median = 0.309931692820972
  speedup_vs_bandit_thompson_median = 0.15183392488759367
  speedup_vs_bandit_ucb1_median = 0.17462861713295175
  speedup_vs_hf_assisted_median = 0.9443276626455602
  speedup_vs_hf_greedy_compiled_median = 0.5036274582188257
  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 = 48.45432208206881
  tokens_per_second_median = 48.450781242990644
  tokens_per_second_p25 = 48.25337363700696
  tokens_per_second_p75 = 48.597782079850404
  tokens_per_second_mean = 48.45552520979821
  gpu_memory_gb = 4.036861896514893
  speedup_vs_arm_composite-g4_median = 0.3110222823499171
  speedup_vs_arm_ngram-g2_median = 0.4080025449384693
  speedup_vs_arm_ngram-g4_median = 0.2925323590871758
  speedup_vs_arm_ngram-g6_median = 0.24757418441757434
  speedup_vs_arm_qwen-g2_median = 0.683065374268339
  speedup_vs_arm_qwen-g4_median = 0.6302920030926732
  speedup_vs_arm_qwen-g6_median = 0.6182922190379585
  speedup_vs_bandit_thompson_median = 0.3014790373867705
  speedup_vs_bandit_ucb1_median = 0.3479675157683503
  speedup_vs_hf_assisted_median = 1.87129046130773
  speedup_vs_hf_greedy_median = 1.9855963933470049
  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=3e38251  n_samples=10
------------------------------------------------------------------------------
[arm_composite-g4]
  workload_tokens_per_second = 138.21937030673996
  tokens_per_second_median = 143.86705435979758
  tokens_per_second_p25 = 141.1382699071676
  tokens_per_second_p75 = 166.93318163213127
  tokens_per_second_mean = 146.02420766659534
  acceptance_rate_mean = 0.6123546770190365
  gpu_memory_gb = 4.036861896514893
  speedup_vs_arm_ngram-g2_median = 1.371584710953472
  speedup_vs_arm_ngram-g4_median = 0.95838517836358
  speedup_vs_arm_ngram-g6_median = 0.7496989623611447
  speedup_vs_arm_qwen-g2_median = 2.0841719313986298
  speedup_vs_arm_qwen-g4_median = 1.9942583405222851
  speedup_vs_arm_qwen-g6_median = 1.9918363082671275
  speedup_vs_bandit_thompson_median = 0.7487722156786364
  speedup_vs_bandit_ucb1_median = 1.0336888495638035
  speedup_vs_hf_assisted_median = 5.992993339765638
  speedup_vs_hf_sample_compiled_median = 3.0116797401462323
  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 = 0.010891823180847412
  step_profile.bookkeeping_ms_mean = 0.28885167864975575
  step_profile.bookkeeping_ms_total = 100.520384170115
  step_profile.draft_fraction = 0.14916238384111608
  step_profile.draft_ms_mean = 3.9557936489153867
  step_profile.draft_ms_total = 1376.6161898225546
  step_profile.n_steps = 348.0
  step_profile.sample_fraction = 0.026750513678372154
  step_profile.sample_ms_mean = 0.7094249192668104
  step_profile.sample_ms_total = 246.87987190485
  step_profile.step_ms_mean = 26.520048466971378
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 3.67816091954023
  step_profile.verify_fraction = 0.8131952792996644
  step_profile.verify_ms_mean = 21.565978220139428
  step_profile.verify_ms_total = 7504.9604206085205
  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_ngram-g2]
  workload_tokens_per_second = 107.43825740553294
  tokens_per_second_median = 115.15961248286163
  tokens_per_second_p25 = 100.38606591585163
  tokens_per_second_p75 = 120.10661775087738
  tokens_per_second_mean = 109.47733758248884
  acceptance_rate_mean = 0.6999190021566685
  gpu_memory_gb = 4.036861896514893
  speedup_vs_arm_composite-g4_median = 0.7297232833069834
  speedup_vs_arm_ngram-g4_median = 0.6804499394618797
  speedup_vs_arm_ngram-g6_median = 0.5367031282123302
  speedup_vs_arm_qwen-g2_median = 1.6682428765176829
  speedup_vs_arm_qwen-g4_median = 1.5318420162370923
  speedup_vs_arm_qwen-g6_median = 1.5251086995453722
  speedup_vs_bandit_thompson_median = 0.5388188625575547
  speedup_vs_bandit_ucb1_median = 0.8887254917786906
  speedup_vs_hf_assisted_median = 4.407488985491135
  speedup_vs_hf_sample_compiled_median = 2.4043230036992247
  forward_counts.emitted_tokens_per_step = 2.3572744014732967
  forward_counts.n_generations = 10
  forward_counts.n_steps = 543
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 553
  forward_counts.target_forwards_per_emitted_token = 0.43203125
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.011868957466498256
  step_profile.bookkeeping_ms_mean = 0.2595616057377933
  step_profile.bookkeeping_ms_total = 140.94195191562176
  step_profile.draft_fraction = 0.004388983003719339
  step_profile.draft_ms_mean = 0.09598243815573979
  step_profile.draft_ms_total = 52.118463918566704
  step_profile.n_steps = 543.0
  step_profile.sample_fraction = 0.031627096142159875
  step_profile.sample_ms_mean = 0.6916512998428134
  step_profile.sample_ms_total = 375.5666558146477
  step_profile.step_ms_mean = 21.86894733344872
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.3572744014732967
  step_profile.verify_fraction = 0.9521149633876226
  step_profile.verify_ms_mean = 20.82175198971237
  step_profile.verify_ms_total = 11306.211330413818
  host_syncs.by_kind = {'cuda': 74}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 72
  host_syncs.per_step = 1.0277777777777777
  host_syncs.prompt_id = HumanEval/0:docstring
  host_syncs.total = 74
  nonfinite_logits.prompt_id = HumanEval/0:docstring
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[arm_ngram-g4]
  workload_tokens_per_second = 156.03475447515817
  tokens_per_second_median = 161.65814536875087
  tokens_per_second_p25 = 147.08846802773996
  tokens_per_second_p75 = 183.27108026338033
  tokens_per_second_mean = 160.22042597885556
  acceptance_rate_mean = 0.7191036064773366
  gpu_memory_gb = 4.036861896514893
  speedup_vs_arm_composite-g4_median = 1.043434049439441
  speedup_vs_arm_ngram-g2_median = 1.4696700829384879
  speedup_vs_arm_ngram-g6_median = 0.8569547075781242
  speedup_vs_arm_qwen-g2_median = 2.5449686290635496
  speedup_vs_arm_qwen-g4_median = 2.3069521944456364
  speedup_vs_arm_qwen-g6_median = 2.241957934978326
  speedup_vs_bandit_thompson_median = 0.858909159432546
  speedup_vs_bandit_ucb1_median = 1.286640587715051
  speedup_vs_hf_assisted_median = 6.619471969950551
  speedup_vs_hf_sample_compiled_median = 3.3739158906008586
  forward_counts.emitted_tokens_per_step = 3.5457063711911356
  forward_counts.n_generations = 10
  forward_counts.n_steps = 361
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 371
  forward_counts.target_forwards_per_emitted_token = 0.28984375
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.011832369235612551
  step_profile.bookkeeping_ms_mean = 0.26783477062994093
  step_profile.bookkeeping_ms_total = 96.68835219740868
  step_profile.draft_fraction = 0.004064716329669389
  step_profile.draft_ms_mean = 0.09200797778995413
  step_profile.draft_ms_total = 33.21487998217344
  step_profile.n_steps = 361.0
  step_profile.sample_fraction = 0.031222722420465854
  step_profile.sample_ms_mean = 0.7067503161377524
  step_profile.sample_ms_total = 255.1368641257286
  step_profile.step_ms_mean = 22.635768483612182
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 3.5457063711911356
  step_profile.verify_fraction = 0.9528801920142522
  step_profile.verify_ms_mean = 21.569175419054535
  step_profile.verify_ms_total = 7786.4723262786865
  host_syncs.by_kind = {'cuda': 53}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 51
  host_syncs.per_step = 1.0392156862745099
  host_syncs.prompt_id = HumanEval/0:docstring
  host_syncs.total = 53
  nonfinite_logits.prompt_id = HumanEval/0:docstring
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[arm_ngram-g6]
  workload_tokens_per_second = 168.50972198168637
  tokens_per_second_median = 205.15008878498514
  tokens_per_second_p25 = 183.52202513648754
  tokens_per_second_p75 = 232.74479594691326
  tokens_per_second_mean = 193.1417518588236
  acceptance_rate_mean = 0.5298703231451246
  gpu_memory_gb = 4.036861896514893
  speedup_vs_arm_composite-g4_median = 1.3383299548046579
  speedup_vs_arm_ngram-g2_median = 1.8672188938160519
  speedup_vs_arm_ngram-g4_median = 1.1680562965611547
  speedup_vs_arm_qwen-g2_median = 3.0654301745736294
  speedup_vs_arm_qwen-g4_median = 3.0384563883531923
  speedup_vs_arm_qwen-g6_median = 2.876509085910745
  speedup_vs_bandit_thompson_median = 1.0077997109857508
  speedup_vs_bandit_ucb1_median = 1.423905695778715
  speedup_vs_hf_assisted_median = 7.987908593076419
  speedup_vs_hf_sample_compiled_median = 4.289859975541859
  forward_counts.emitted_tokens_per_step = 3.914373088685015
  forward_counts.n_generations = 10
  forward_counts.n_steps = 327
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 337
  forward_counts.target_forwards_per_emitted_token = 0.26328125
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.013225618234867727
  step_profile.bookkeeping_ms_mean = 0.30597274527181545
  step_profile.bookkeeping_ms_total = 100.05308770388365
  step_profile.draft_fraction = 0.004006703811000379
  step_profile.draft_ms_mean = 0.0926945071883887
  step_profile.draft_ms_total = 30.311103850603104
  step_profile.n_steps = 327.0
  step_profile.sample_fraction = 0.031114598071048032
  step_profile.sample_ms_mean = 0.7198316797568528
  step_profile.sample_ms_total = 235.38495928049088
  step_profile.step_ms_mean = 23.134853875122122
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 3.914373088685015
  step_profile.verify_fraction = 0.9516530798830839
  step_profile.verify_ms_mean = 22.016354942905064
  step_profile.verify_ms_total = 7199.348066329956
  host_syncs.by_kind = {'cuda': 68}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 66
  host_syncs.per_step = 1.0303030303030303
  host_syncs.prompt_id = HumanEval/0:docstring
  host_syncs.total = 68
  nonfinite_logits.prompt_id = HumanEval/0:docstring
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[arm_qwen-g2]
  workload_tokens_per_second = 67.62688823251747
  tokens_per_second_median = 68.31355578247965
  tokens_per_second_p25 = 65.44514516895929
  tokens_per_second_p75 = 69.95189800982499
  tokens_per_second_mean = 67.75312230960546
  acceptance_rate_mean = 0.9267043614054128
  gpu_memory_gb = 4.036861896514893
  speedup_vs_arm_composite-g4_median = 0.4798160100858667
  speedup_vs_arm_ngram-g2_median = 0.5995272625794922
  speedup_vs_arm_ngram-g4_median = 0.3929487002475369
  speedup_vs_arm_ngram-g6_median = 0.32623200443834605
  speedup_vs_arm_qwen-g4_median = 0.9180504915770016
  speedup_vs_arm_qwen-g6_median = 0.9003842584286963
  speedup_vs_bandit_thompson_median = 0.32225969636354745
  speedup_vs_bandit_ucb1_median = 0.5215752240364364
  speedup_vs_hf_assisted_median = 2.7904788333399146
  speedup_vs_hf_sample_compiled_median = 1.4284960644835842
  forward_counts.emitted_tokens_per_step = 2.831858407079646
  forward_counts.n_generations = 10
  forward_counts.n_steps = 452
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 462
  forward_counts.target_forwards_per_emitted_token = 0.3609375
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.006346521175561295
  step_profile.bookkeeping_ms_mean = 0.26544814161231034
  step_profile.bookkeeping_ms_total = 119.98256000876427
  step_profile.draft_fraction = 0.5131857684757387
  step_profile.draft_ms_mean = 21.46439045509406
  step_profile.draft_ms_total = 9701.904485702515
  step_profile.n_steps = 452.0
  step_profile.sample_fraction = 0.009563261743648944
  step_profile.sample_ms_mean = 0.3999907961197659
  step_profile.sample_ms_total = 180.7958398461342
  step_profile.step_ms_mean = 41.825771043588105
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.831858407079646
  step_profile.verify_fraction = 0.4709044486050511
  step_profile.verify_ms_mean = 19.695941650761966
  step_profile.verify_ms_total = 8902.56562614441
  host_syncs.by_kind = {'cuda': 48}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 45
  host_syncs.per_step = 1.0666666666666667
  host_syncs.prompt_id = HumanEval/0:docstring
  host_syncs.total = 48
  nonfinite_logits.prompt_id = HumanEval/0:docstring
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[arm_qwen-g4]
  workload_tokens_per_second = 71.59995638048053
  tokens_per_second_median = 76.60710812497459
  tokens_per_second_p25 = 67.2483394354154
  tokens_per_second_p75 = 79.44644546378693
  tokens_per_second_mean = 72.61139965370124
  acceptance_rate_mean = 0.8685268019282573
  gpu_memory_gb = 4.036861896514893
  speedup_vs_arm_composite-g4_median = 0.5039183090599058
  speedup_vs_arm_ngram-g2_median = 0.652927218578051
  speedup_vs_arm_ngram-g4_median = 0.43359753826183534
  speedup_vs_arm_ngram-g6_median = 0.3291252601266067
  speedup_vs_arm_qwen-g2_median = 1.0894790309258806
  speedup_vs_arm_qwen-g6_median = 1.0159229375629413
  speedup_vs_bandit_thompson_median = 0.33350325610027987
  speedup_vs_bandit_ucb1_median = 0.5840216960628223
  speedup_vs_hf_assisted_median = 3.0731154568393255
  speedup_vs_hf_sample_compiled_median = 1.5980283896029839
  forward_counts.emitted_tokens_per_step = 4.413793103448276
  forward_counts.n_generations = 10
  forward_counts.n_steps = 290
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 300
  forward_counts.target_forwards_per_emitted_token = 0.234375
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.005407445186926798
  step_profile.bookkeeping_ms_mean = 0.33304673047928973
  step_profile.bookkeeping_ms_total = 96.58355183899403
  step_profile.draft_fraction = 0.6560432256677649
  step_profile.draft_ms_mean = 40.40596692315464
  step_profile.draft_ms_total = 11717.730407714844
  step_profile.n_steps = 290.0
  step_profile.sample_fraction = 0.006700506471543665
  step_profile.sample_ms_mean = 0.4126868966324576
  step_profile.sample_ms_total = 119.6792000234127
  step_profile.step_ms_mean = 61.59040340981607
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 4.413793103448276
  step_profile.verify_fraction = 0.33184882267376475
  step_profile.verify_ms_mean = 20.438702859549686
  step_profile.verify_ms_total = 5927.223829269409
  host_syncs.by_kind = {'cuda': 34}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 31
  host_syncs.per_step = 1.096774193548387
  host_syncs.prompt_id = HumanEval/0:docstring
  host_syncs.total = 34
  nonfinite_logits.prompt_id = HumanEval/0:docstring
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[arm_qwen-g6]
  workload_tokens_per_second = 72.96606526586656
  tokens_per_second_median = 76.40494178866957
  tokens_per_second_p25 = 70.11554898576523
  tokens_per_second_p75 = 78.00712188094558
  tokens_per_second_mean = 73.7398495717925
  acceptance_rate_mean = 0.8327743329477355
  gpu_memory_gb = 4.036861896514893
  speedup_vs_arm_composite-g4_median = 0.5035557152703912
  speedup_vs_arm_ngram-g2_median = 0.6557582448074506
  speedup_vs_arm_ngram-g4_median = 0.4465816095529663
  speedup_vs_arm_ngram-g6_median = 0.3493873698125703
  speedup_vs_arm_qwen-g2_median = 1.1106379975794933
  speedup_vs_arm_qwen-g4_median = 0.9843382619240453
  speedup_vs_bandit_thompson_median = 0.352379294534543
  speedup_vs_bandit_ucb1_median = 0.5863317695107686
  speedup_vs_hf_assisted_median = 3.039690177342499
  speedup_vs_hf_sample_compiled_median = 1.5955621637181818
  forward_counts.emitted_tokens_per_step = 5.953488372093023
  forward_counts.n_generations = 10
  forward_counts.n_steps = 215
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 225
  forward_counts.target_forwards_per_emitted_token = 0.17578125
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.0043318015057777545
  step_profile.bookkeeping_ms_mean = 0.35318042823048523
  step_profile.bookkeeping_ms_total = 75.93379206955433
  step_profile.draft_fraction = 0.7301073360778328
  step_profile.draft_ms_mean = 59.527109279189
  step_profile.draft_ms_total = 12798.328495025635
  step_profile.n_steps = 215.0
  step_profile.sample_fraction = 0.005229970873528621
  step_profile.sample_ms_mean = 0.42640997060509617
  step_profile.sample_ms_total = 91.67814368009567
  step_profile.step_ms_mean = 81.53199719779714
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 5.953488372093023
  step_profile.verify_fraction = 0.26033089154286077
  step_profile.verify_ms_mean = 21.225297519772553
  step_profile.verify_ms_total = 4563.438966751099
  host_syncs.by_kind = {'cuda': 26}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 23
  host_syncs.per_step = 1.1304347826086956
  host_syncs.prompt_id = HumanEval/0:docstring
  host_syncs.total = 26
  nonfinite_logits.prompt_id = HumanEval/0:docstring
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[bandit_thompson]
  workload_tokens_per_second = 167.40160993296038
  tokens_per_second_median = 203.66717638785948
  tokens_per_second_p25 = 181.6829900694417
  tokens_per_second_p75 = 232.29725235397345
  tokens_per_second_mean = 191.85016285364281
  acceptance_rate_mean = 0.633296239744747
  gpu_memory_gb = 4.036861896514893
  speedup_vs_arm_composite-g4_median = 1.339234646237349
  speedup_vs_arm_ngram-g2_median = 1.8582998524691932
  speedup_vs_arm_ngram-g4_median = 1.166057090264208
  speedup_vs_arm_ngram-g6_median = 0.9922608518569243
  speedup_vs_arm_qwen-g2_median = 3.103087576517329
  speedup_vs_arm_qwen-g4_median = 2.99850564457868
  speedup_vs_arm_qwen-g6_median = 2.8525606226879976
  speedup_vs_bandit_ucb1_median = 1.4100776805438122
  speedup_vs_hf_assisted_median = 7.923337867206955
  speedup_vs_hf_sample_compiled_median = 4.255098313003652
  forward_counts.emitted_tokens_per_step = 4.142394822006472
  forward_counts.n_generations = 10
  forward_counts.n_steps = 309
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 319
  forward_counts.target_forwards_per_emitted_token = 0.24921875
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.012539172426228354
  step_profile.bookkeeping_ms_mean = 0.3090173729897317
  step_profile.bookkeeping_ms_total = 95.4863682538271
  step_profile.draft_fraction = 0.06201927763653762
  step_profile.draft_ms_mean = 1.5284130083318668
  step_profile.draft_ms_total = 472.2796195745468
  step_profile.n_steps = 309.0
  step_profile.sample_fraction = 0.029163966192746103
  step_profile.sample_ms_mean = 0.7187214524614772
  step_profile.sample_ms_total = 222.08492881059647
  step_profile.step_ms_mean = 24.644160115651328
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 4.142394822006472
  step_profile.verify_fraction = 0.896277583744488
  step_profile.verify_ms_mean = 22.08800828186825
  step_profile.verify_ms_total = 6825.19455909729
  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 = HumanEval/0:docstring
  host_syncs.total = 52
  nonfinite_logits.prompt_id = HumanEval/0:docstring
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[bandit_ucb1]
  workload_tokens_per_second = 119.255063830093
  tokens_per_second_median = 130.50616069568002
  tokens_per_second_p25 = 104.26656401273506
  tokens_per_second_p75 = 160.1149324923341
  tokens_per_second_mean = 133.51698436161934
  acceptance_rate_mean = 0.7228578531152006
  gpu_memory_gb = 4.036861896514893
  speedup_vs_arm_composite-g4_median = 0.9678686922829292
  speedup_vs_arm_ngram-g2_median = 1.1255748768463454
  speedup_vs_arm_ngram-g4_median = 0.7777944120620531
  speedup_vs_arm_ngram-g6_median = 0.7023131703994507
  speedup_vs_arm_qwen-g2_median = 1.9177386028934453
  speedup_vs_arm_qwen-g4_median = 1.7122725129211214
  speedup_vs_arm_qwen-g6_median = 1.705859625336001
  speedup_vs_bandit_thompson_median = 0.7092657247404757
  speedup_vs_hf_assisted_median = 5.265615243833418
  speedup_vs_hf_sample_compiled_median = 2.7328818264516563
  forward_counts.emitted_tokens_per_step = 3.9263803680981595
  forward_counts.n_generations = 10
  forward_counts.n_steps = 326
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 336
  forward_counts.target_forwards_per_emitted_token = 0.2625
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.009616119132541606
  step_profile.bookkeeping_ms_mean = 0.31562277327285954
  step_profile.bookkeeping_ms_total = 102.89302408695221
  step_profile.draft_fraction = 0.31119362522284455
  step_profile.draft_ms_mean = 10.214078430589174
  step_profile.draft_ms_total = 3329.789568372071
  step_profile.n_steps = 326.0
  step_profile.sample_fraction = 0.021736057402806368
  step_profile.sample_ms_mean = 0.7134265521186992
  step_profile.sample_ms_total = 232.57705599069595
  step_profile.step_ms_mean = 32.82226113492817
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 3.9263803680981595
  step_profile.verify_fraction = 0.6574541982418075
  step_profile.verify_ms_mean = 21.57913337894744
  step_profile.verify_ms_total = 7034.797481536865
  host_syncs.by_kind = {'cuda': 41}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 37
  host_syncs.per_step = 1.1081081081081081
  host_syncs.prompt_id = HumanEval/0:docstring
  host_syncs.total = 41
  nonfinite_logits.prompt_id = HumanEval/0:docstring
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[hf_assisted]
  workload_tokens_per_second = 24.072729686721566
  tokens_per_second_median = 24.844427476805066
  tokens_per_second_p25 = 23.06484466461187
  tokens_per_second_p75 = 26.043548572038723
  tokens_per_second_mean = 24.352869469238048
  gpu_memory_gb = 4.036861896514893
  speedup_vs_arm_composite-g4_median = 0.16691571571738661
  speedup_vs_arm_ngram-g2_median = 0.22688887243115002
  speedup_vs_arm_ngram-g4_median = 0.1512547731093976
  speedup_vs_arm_ngram-g6_median = 0.12518921494222096
  speedup_vs_arm_qwen-g2_median = 0.35844379003639426
  speedup_vs_arm_qwen-g4_median = 0.3254958275359417
  speedup_vs_arm_qwen-g6_median = 0.32907507553046267
  speedup_vs_bandit_thompson_median = 0.12620968966391582
  speedup_vs_bandit_ucb1_median = 0.189957281930916
  speedup_vs_hf_sample_compiled_median = 0.5196786147122127
  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_sample_compiled]
  workload_tokens_per_second = 47.85235447567032
  tokens_per_second_median = 47.877821426380606
  tokens_per_second_p25 = 47.76174074930799
  tokens_per_second_p75 = 47.93425928006561
  tokens_per_second_mean = 47.85271231110988
  gpu_memory_gb = 4.036861896514893
  speedup_vs_arm_composite-g4_median = 0.33206469695999746
  speedup_vs_arm_ngram-g2_median = 0.41625624721503746
  speedup_vs_arm_ngram-g4_median = 0.2965569868201643
  speedup_vs_arm_ngram-g6_median = 0.23314928068394222
  speedup_vs_arm_qwen-g2_median = 0.700062898887752
  speedup_vs_arm_qwen-g4_median = 0.625779931618349
  speedup_vs_arm_qwen-g6_median = 0.6269680597709087
  speedup_vs_bandit_thompson_median = 0.23504512923664167
  speedup_vs_bandit_ucb1_median = 0.36641355087669114
  speedup_vs_hf_assisted_median = 1.925014535819263
  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=3e38251  n_samples=30
------------------------------------------------------------------------------
[arm_composite-g4]
  workload_tokens_per_second = 78.89803748151297
  tokens_per_second_median = 78.41882484248795
  tokens_per_second_p25 = 62.90277077807063
  tokens_per_second_p75 = 126.67678909284638
  tokens_per_second_mean = 93.7167403540511
  acceptance_rate_mean = 0.4929740645944779
  gpu_memory_gb = 4.039843559265137
  speedup_vs_arm_ngram-g2_median = 1.2112623023667615
  speedup_vs_arm_ngram-g4_median = 1.0757934005358534
  speedup_vs_arm_ngram-g6_median = 1.0695058769421073
  speedup_vs_arm_qwen-g2_median = 1.1502410070049571
  speedup_vs_arm_qwen-g4_median = 1.152032955679196
  speedup_vs_arm_qwen-g6_median = 1.2568350069139176
  speedup_vs_bandit_thompson_median = 1.025899638893427
  speedup_vs_bandit_ucb1_legacy_median = 1.0619613052665215
  speedup_vs_bandit_ucb1_median = 1.071750345995203
  speedup_vs_flashspec_compiled_legacy_median = 1.1463563965843342
  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 = 0.01069233148366362
  step_profile.bookkeeping_ms_mean = 0.393376897128694
  step_profile.bookkeeping_ms_total = 519.2575042098761
  step_profile.draft_fraction = 0.4171417316274402
  step_profile.draft_ms_mean = 15.346879237816866
  step_profile.draft_ms_total = 20257.880593918264
  step_profile.n_steps = 1320.0
  step_profile.sample_fraction = 0.011989978468505066
  step_profile.sample_ms_mean = 0.44111806052652275
  step_profile.sample_ms_total = 582.27583989501
  step_profile.step_ms_mean = 36.790563192856354
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 2.909090909090909
  step_profile.verify_fraction = 0.5601759584203911
  step_profile.verify_ms_mean = 20.609188997384273
  step_profile.verify_ms_total = 27204.12947654724
  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_ngram-g2]
  workload_tokens_per_second = 68.65957746912103
  tokens_per_second_median = 66.32981627631827
  tokens_per_second_p25 = 54.25506658090488
  tokens_per_second_p75 = 99.02021621472403
  tokens_per_second_mean = 76.32338657588535
  acceptance_rate_mean = 0.36256636440491286
  gpu_memory_gb = 4.039843559265137
  speedup_vs_arm_composite-g4_median = 0.8255882349798365
  speedup_vs_arm_ngram-g4_median = 0.9508823104699687
  speedup_vs_arm_ngram-g6_median = 0.9262932141309537
  speedup_vs_arm_qwen-g2_median = 1.0365306569498542
  speedup_vs_arm_qwen-g4_median = 1.002028407611375
  speedup_vs_arm_qwen-g6_median = 1.1417689465989234
  speedup_vs_bandit_thompson_median = 0.8629194829953832
  speedup_vs_bandit_ucb1_legacy_median = 0.909047488289551
  speedup_vs_bandit_ucb1_median = 0.9198675477066229
  speedup_vs_flashspec_compiled_legacy_median = 0.9994509592919123
  forward_counts.emitted_tokens_per_step = 1.4556482183472328
  forward_counts.n_generations = 30
  forward_counts.n_steps = 2638
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 2668
  forward_counts.target_forwards_per_emitted_token = 0.6947916666666667
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 3840
  step_profile.bookkeeping_fraction = 0.01785722022895576
  step_profile.bookkeeping_ms_mean = 0.37750663245386296
  step_profile.bookkeeping_ms_total = 995.8624964132905
  step_profile.draft_fraction = 0.004814645962634213
  step_profile.draft_ms_mean = 0.10178296288603901
  step_profile.draft_ms_total = 268.5034560933709
  step_profile.n_steps = 2638.0
  step_profile.sample_fraction = 0.020520929403012138
  step_profile.sample_ms_mean = 0.4338181897534666
  step_profile.sample_ms_total = 1144.412384569645
  step_profile.step_ms_mean = 21.140279820356927
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 1.4556482183472328
  step_profile.verify_fraction = 0.9568072044053979
  step_profile.verify_ms_mean = 20.22717203526356
  step_profile.verify_ms_total = 53359.27982902527
  host_syncs.by_kind = {'cuda': 109}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 107
  host_syncs.per_step = 1.0186915887850467
  host_syncs.prompt_id = mt_bench/121
  host_syncs.total = 109
  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_ngram-g4]
  workload_tokens_per_second = 73.75676263701732
  tokens_per_second_median = 69.73998101185336
  tokens_per_second_p25 = 54.65960386618752
  tokens_per_second_p75 = 130.3801934303072
  tokens_per_second_mean = 91.64073617234105
  acceptance_rate_mean = 0.2551398005711606
  gpu_memory_gb = 4.039843559265137
  speedup_vs_arm_composite-g4_median = 0.9296377808675671
  speedup_vs_arm_ngram-g2_median = 1.0516653203812423
  speedup_vs_arm_ngram-g6_median = 0.9856573225228941
  speedup_vs_arm_qwen-g2_median = 1.0841361175158086
  speedup_vs_arm_qwen-g4_median = 1.0597997708092717
  speedup_vs_arm_qwen-g6_median = 1.1589619984314972
  speedup_vs_bandit_thompson_median = 1.0115987086122304
  speedup_vs_bandit_ucb1_legacy_median = 0.989555782914316
  speedup_vs_bandit_ucb1_median = 1.001527557119792
  speedup_vs_flashspec_compiled_legacy_median = 1.0448949147656943
  forward_counts.emitted_tokens_per_step = 1.5894039735099337
  forward_counts.n_generations = 30
  forward_counts.n_steps = 2416
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 2446
  forward_counts.target_forwards_per_emitted_token = 0.6369791666666667
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 3840
  step_profile.bookkeeping_fraction = 0.018629054333503474
  step_profile.bookkeeping_ms_mean = 0.4003217087762541
  step_profile.bookkeeping_ms_total = 967.17724840343
  step_profile.draft_fraction = 0.004699596909638161
  step_profile.draft_ms_mean = 0.10099013249655038
  step_profile.draft_ms_total = 243.99216011166573
  step_profile.n_steps = 2416.0
  step_profile.sample_fraction = 0.02045791391851401
  step_profile.sample_ms_mean = 0.4396222648365016
  step_profile.sample_ms_total = 1062.1273918449879
  step_profile.step_ms_mean = 21.489105222925595
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 1.5894039735099337
  step_profile.verify_fraction = 0.9562134348383443
  step_profile.verify_ms_mean = 20.548171116816288
  step_profile.verify_ms_total = 49644.38141822815
  host_syncs.by_kind = {'cuda': 104}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 102
  host_syncs.per_step = 1.0196078431372548
  host_syncs.prompt_id = mt_bench/121
  host_syncs.total = 104
  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_ngram-g6]
  workload_tokens_per_second = 75.96419755365469
  tokens_per_second_median = 72.27049793370898
  tokens_per_second_p25 = 54.35384366148586
  tokens_per_second_p75 = 149.98900119262436
  tokens_per_second_mean = 102.3815385940849
  acceptance_rate_mean = 0.19553943356347409
  gpu_memory_gb = 4.039843559265137
  speedup_vs_arm_composite-g4_median = 0.9350832529442696
  speedup_vs_arm_ngram-g2_median = 1.0796116324947787
  speedup_vs_arm_ngram-g4_median = 1.0145540104869195
  speedup_vs_arm_qwen-g2_median = 1.1126481495328584
  speedup_vs_arm_qwen-g4_median = 1.0877612045675982
  speedup_vs_arm_qwen-g6_median = 1.1528483794250146
  speedup_vs_bandit_thompson_median = 1.0227600966201333
  speedup_vs_bandit_ucb1_legacy_median = 1.0054080203306786
  speedup_vs_bandit_ucb1_median = 1.0414567025441936
  speedup_vs_flashspec_compiled_legacy_median = 1.0722172221705808
  forward_counts.emitted_tokens_per_step = 1.6516129032258065
  forward_counts.n_generations = 30
  forward_counts.n_steps = 2325
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 2355
  forward_counts.target_forwards_per_emitted_token = 0.61328125
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 3840
  step_profile.bookkeeping_fraction = 0.018961493411466794
  step_profile.bookkeeping_ms_mean = 0.41111641466938037
  step_profile.bookkeeping_ms_total = 955.8456641063094
  step_profile.draft_fraction = 0.004672423907229899
  step_profile.draft_ms_mean = 0.10130584774478789
  step_profile.draft_ms_total = 235.53609600663185
  step_profile.n_steps = 2325.0
  step_profile.sample_fraction = 0.020587097613678065
  step_profile.sample_ms_mean = 0.446362191823221
  step_profile.sample_ms_total = 1037.7920959889889
  step_profile.step_ms_mean = 21.681647418170208
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 1.6516129032258065
  step_profile.verify_fraction = 0.9557789850676253
  step_profile.verify_ms_mean = 20.722862963932815
  step_profile.verify_ms_total = 48180.6563911438
  host_syncs.by_kind = {'cuda': 102}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 100
  host_syncs.per_step = 1.02
  host_syncs.prompt_id = mt_bench/121
  host_syncs.total = 102
  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
[arm_qwen-g2]
  workload_tokens_per_second = 65.16081282682634
  tokens_per_second_median = 67.69244277078965
  tokens_per_second_p25 = 62.138363871808295
  tokens_per_second_p75 = 71.1912098786348
  tokens_per_second_mean = 65.77476600557493
  acceptance_rate_mean = 0.8321181656157729
  gpu_memory_gb = 4.039843559265137
  speedup_vs_arm_composite-g4_median = 0.8694179380069176
  speedup_vs_arm_ngram-g2_median = 0.9661466644227392
  speedup_vs_arm_ngram-g4_median = 0.9224515217861167
  speedup_vs_arm_ngram-g6_median = 0.8993722486088476
  speedup_vs_arm_qwen-g4_median = 0.9772689337355409
  speedup_vs_arm_qwen-g6_median = 0.9853868104113217
  speedup_vs_bandit_thompson_median = 0.9066683445330217
  speedup_vs_bandit_ucb1_legacy_median = 0.9510639046786769
  speedup_vs_bandit_ucb1_median = 0.9156774441831979
  speedup_vs_flashspec_compiled_legacy_median = 0.9843777440035795
  forward_counts.emitted_tokens_per_step = 2.650103519668737
  forward_counts.n_generations = 30
  forward_counts.n_steps = 1449
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1479
  forward_counts.target_forwards_per_emitted_token = 0.38515625
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 3840
  step_profile.bookkeeping_fraction = 0.007641159186839026
  step_profile.bookkeeping_ms_mean = 0.31040002226706126
  step_profile.bookkeeping_ms_total = 449.76963226497173
  step_profile.draft_fraction = 0.5073407027792057
  step_profile.draft_ms_mean = 20.609250715636154
  step_profile.draft_ms_total = 29862.804286956787
  step_profile.n_steps = 1449.0
  step_profile.sample_fraction = 0.0038261973979359957
  step_profile.sample_ms_mean = 0.15542821821630043
  step_profile.sample_ms_total = 225.2154881954193
  step_profile.step_ms_mean = 40.622111734262496
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 2.650103519668737
  step_profile.verify_fraction = 0.4811919406360192
  step_profile.verify_ms_mean = 19.547032778142977
  step_profile.verify_ms_total = 28323.650495529175
  host_syncs.by_kind = {'cuda': 51}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 48
  host_syncs.per_step = 1.0625
  host_syncs.prompt_id = mt_bench/121
  host_syncs.total = 51
  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_qwen-g4]
  workload_tokens_per_second = 64.74130386751415
  tokens_per_second_median = 69.88257997073731
  tokens_per_second_p25 = 56.313250088992
  tokens_per_second_p75 = 77.02643948975683
  tokens_per_second_mean = 66.89429783060397
  acceptance_rate_mean = 0.7286482952790188
  gpu_memory_gb = 4.039843559265137
  speedup_vs_arm_composite-g4_median = 0.8682155775870054
  speedup_vs_arm_ngram-g2_median = 0.997977160580819
  speedup_vs_arm_ngram-g4_median = 0.9435848189283118
  speedup_vs_arm_ngram-g6_median = 0.9195160601713572
  speedup_vs_arm_qwen-g2_median = 1.0232638227991622
  speedup_vs_arm_qwen-g6_median = 1.030517500943087
  speedup_vs_bandit_thompson_median = 0.873315277353574
  speedup_vs_bandit_ucb1_legacy_median = 0.9293725198699183
  speedup_vs_bandit_ucb1_median = 0.9200505914468673
  speedup_vs_flashspec_compiled_legacy_median = 1.00104216101917
  forward_counts.emitted_tokens_per_step = 3.8984771573604062
  forward_counts.n_generations = 30
  forward_counts.n_steps = 985
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1015
  forward_counts.target_forwards_per_emitted_token = 0.2643229166666667
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 3840
  step_profile.bookkeeping_fraction = 0.00679703404660272
  step_profile.bookkeeping_ms_mean = 0.408934432873266
  step_profile.bookkeeping_ms_total = 402.800416380167
  step_profile.draft_fraction = 0.656497199545015
  step_profile.draft_ms_mean = 39.497273095610176
  step_profile.draft_ms_total = 38904.813999176025
  step_profile.n_steps = 985.0
  step_profile.sample_fraction = 0.0027170574846107766
  step_profile.sample_ms_mean = 0.16346811770182576
  step_profile.sample_ms_total = 161.01609593629837
  step_profile.step_ms_mean = 60.16365815876099
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 3.8984771573604062
  step_profile.verify_fraction = 0.3339887089237716
  step_profile.verify_ms_mean = 20.09398251257572
  step_profile.verify_ms_total = 19792.572774887085
  host_syncs.by_kind = {'cuda': 37}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 34
  host_syncs.per_step = 1.088235294117647
  host_syncs.prompt_id = mt_bench/121
  host_syncs.total = 37
  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_qwen-g6]
  workload_tokens_per_second = 60.82585585764901
  tokens_per_second_median = 68.3914017640563
  tokens_per_second_p25 = 51.50863625833867
  tokens_per_second_p75 = 78.43882976876453
  tokens_per_second_mean = 64.60915569986759
  acceptance_rate_mean = 0.6324407358953257
  gpu_memory_gb = 4.039843559265137
  speedup_vs_arm_composite-g4_median = 0.7958534482749944
  speedup_vs_arm_ngram-g2_median = 0.8758413060038004
  speedup_vs_arm_ngram-g4_median = 0.8628761895657081
  speedup_vs_arm_ngram-g6_median = 0.8675059789349397
  speedup_vs_arm_qwen-g2_median = 1.0148666196293665
  speedup_vs_arm_qwen-g4_median = 0.970403113139046
  speedup_vs_bandit_thompson_median = 0.8403532143834513
  speedup_vs_bandit_ucb1_legacy_median = 0.8546753894854232
  speedup_vs_bandit_ucb1_median = 0.8634098488096809
  speedup_vs_flashspec_compiled_legacy_median = 0.9710261197107427
  forward_counts.emitted_tokens_per_step = 4.830188679245283
  forward_counts.n_generations = 30
  forward_counts.n_steps = 795
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 825
  forward_counts.target_forwards_per_emitted_token = 0.21484375
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 3840
  step_profile.bookkeeping_fraction = 0.006082863721397092
  step_profile.bookkeeping_ms_mean = 0.48270172609848044
  step_profile.bookkeeping_ms_total = 383.74787224829197
  step_profile.draft_fraction = 0.7335652529012936
  step_profile.draft_ms_mean = 58.21159736585317
  step_profile.draft_ms_total = 46278.21990585327
  step_profile.n_steps = 795.0
  step_profile.sample_fraction = 0.0022072223349477384
  step_profile.sample_ms_mean = 0.1751527043446055
  step_profile.sample_ms_total = 139.24639995396137
  step_profile.step_ms_mean = 79.35435482477241
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 4.830188679245283
  step_profile.verify_fraction = 0.25814466104236156
  step_profile.verify_ms_mean = 20.484903028476165
  step_profile.verify_ms_total = 16285.49790763855
  host_syncs.by_kind = {'cuda': 31}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 28
  host_syncs.per_step = 1.1071428571428572
  host_syncs.prompt_id = mt_bench/121
  host_syncs.total = 31
  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
[bandit_thompson]
  workload_tokens_per_second = 75.33373795383227
  tokens_per_second_median = 67.91255961316804
  tokens_per_second_p25 = 60.01384892378891
  tokens_per_second_p75 = 128.951453330462
  tokens_per_second_mean = 91.27232526290959
  acceptance_rate_mean = 0.429918679458752
  gpu_memory_gb = 4.039843559265137
  speedup_vs_arm_composite-g4_median = 0.9747565516812156
  speedup_vs_arm_ngram-g2_median = 1.158856677472004
  speedup_vs_arm_ngram-g4_median = 0.9885378333001533
  speedup_vs_arm_ngram-g6_median = 0.977776643359053
  speedup_vs_arm_qwen-g2_median = 1.1029517108473423
  speedup_vs_arm_qwen-g4_median = 1.1452488573597561
  speedup_vs_arm_qwen-g6_median = 1.1899878373506638
  speedup_vs_bandit_ucb1_legacy_median = 1.0167911617211822
  speedup_vs_bandit_ucb1_median = 1.008885799144973
  speedup_vs_flashspec_compiled_legacy_median = 1.1399106884641643
  forward_counts.emitted_tokens_per_step = 2.3202416918429005
  forward_counts.n_generations = 30
  forward_counts.n_steps = 1655
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1685
  forward_counts.target_forwards_per_emitted_token = 0.4388020833333333
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 3840
  step_profile.bookkeeping_fraction = 0.013254976288340685
  step_profile.bookkeeping_ms_mean = 0.40728844000853803
  step_profile.bookkeeping_ms_total = 674.0623682141304
  step_profile.draft_fraction = 0.2957582536795583
  step_profile.draft_ms_mean = 9.087825971198045
  step_profile.draft_ms_total = 15040.351982332766
  step_profile.n_steps = 1655.0
  step_profile.sample_fraction = 0.014372382081662102
  step_profile.sample_ms_mean = 0.4416232024794979
  step_profile.sample_ms_total = 730.886400103569
  step_profile.step_ms_mean = 30.727210003898403
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 2.3202416918429005
  step_profile.verify_fraction = 0.6766143879504389
  step_profile.verify_ms_mean = 20.790472390212322
  step_profile.verify_ms_total = 34408.23180580139
  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 = mt_bench/121
  host_syncs.total = 54
  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
[bandit_ucb1]
  workload_tokens_per_second = 75.01966105459854
  tokens_per_second_median = 74.41632442656362
  tokens_per_second_p25 = 57.80442393699757
  tokens_per_second_p75 = 115.5627393977952
  tokens_per_second_mean = 90.23560444557413
  acceptance_rate_mean = 0.5182552049826765
  gpu_memory_gb = 4.039843559265137
  speedup_vs_arm_composite-g4_median = 0.9330610595072647
  speedup_vs_arm_ngram-g2_median = 1.087124396064299
  speedup_vs_arm_ngram-g4_median = 0.9985233612608594
  speedup_vs_arm_ngram-g6_median = 0.9602252481264328
  speedup_vs_arm_qwen-g2_median = 1.0922293760398643
  speedup_vs_arm_qwen-g4_median = 1.0871964068763054
  speedup_vs_arm_qwen-g6_median = 1.1582194596017337
  speedup_vs_bandit_thompson_median = 0.9911965805836027
  speedup_vs_bandit_ucb1_legacy_median = 1.0249459671229628
  speedup_vs_flashspec_compiled_legacy_median = 1.071633391670719
  forward_counts.emitted_tokens_per_step = 2.694736842105263
  forward_counts.n_generations = 30
  forward_counts.n_steps = 1425
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1455
  forward_counts.target_forwards_per_emitted_token = 0.37890625
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 3840
  step_profile.bookkeeping_fraction = 0.010885327191225166
  step_profile.bookkeeping_ms_mean = 0.3901519045076872
  step_profile.bookkeeping_ms_total = 555.9664639234543
  step_profile.draft_fraction = 0.3978737487162451
  step_profile.draft_ms_mean = 14.260591169036271
  step_profile.draft_ms_total = 20321.342415876687
  step_profile.n_steps = 1425.0
  step_profile.sample_fraction = 0.012376995165733313
  step_profile.sample_ms_mean = 0.44361626905307433
  step_profile.sample_ms_total = 632.153183400631
  step_profile.step_ms_mean = 35.842000672446005
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 2.694736842105263
  step_profile.verify_fraction = 0.5788639289267964
  step_profile.verify_ms_mean = 20.747641329848975
  step_profile.verify_ms_total = 29565.38889503479
  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
  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
[bandit_ucb1_legacy]
  workload_tokens_per_second = 23.26460410753764
  tokens_per_second_median = 70.001800108375
  tokens_per_second_p25 = 56.099147482578
  tokens_per_second_p75 = 113.69131136712147
  tokens_per_second_mean = 85.42549938666663
  acceptance_rate_mean = 0.5431340530481747
  gpu_memory_gb = 4.039843559265137
  speedup_vs_arm_composite-g4_median = 0.9416540354597551
  speedup_vs_arm_ngram-g2_median = 1.1001070051158646
  speedup_vs_arm_ngram-g4_median = 1.0105729375517365
  speedup_vs_arm_ngram-g6_median = 0.9946228571636812
  speedup_vs_arm_qwen-g2_median = 1.051461863761331
  speedup_vs_arm_qwen-g4_median = 1.0759958533954066
  speedup_vs_arm_qwen-g6_median = 1.1703950187993246
  speedup_vs_bandit_thompson_median = 0.9835120451502921
  speedup_vs_bandit_ucb1_median = 0.9757146472799527
  speedup_vs_flashspec_compiled_legacy_median = 1.0759192812530265
  forward_counts.emitted_tokens_per_step = 2.689075630252101
  forward_counts.n_generations = 30
  forward_counts.n_steps = 1428
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1458
  forward_counts.target_forwards_per_emitted_token = 0.3796875
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 3840
  step_profile.bookkeeping_fraction = 0.0033639496352671407
  step_profile.bookkeeping_ms_mean = 0.3885679776031597
  step_profile.bookkeeping_ms_total = 554.875072017312
  step_profile.draft_fraction = 0.1278469481656526
  step_profile.draft_ms_mean = 14.767530872238702
  step_profile.draft_ms_total = 21088.034085556865
  step_profile.n_steps = 1428.0
  step_profile.sample_fraction = 0.003899293800410411
  step_profile.sample_ms_mean = 0.45040528853984774
  step_profile.sample_ms_total = 643.1787520349026
  step_profile.step_ms_mean = 115.5094516069657
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 2.689075630252101
  step_profile.verify_fraction = 0.8648898083986698
  step_profile.verify_ms_mean = 99.90294746858399
  step_profile.verify_ms_total = 142661.40898513794
  host_syncs.by_kind = {'cuda': 55}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 50
  host_syncs.per_step = 1.1
  host_syncs.prompt_id = mt_bench/121
  host_syncs.total = 55
  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, 97, None, None, None, 94, 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, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, 97, None, None, None, None, None]
  greedy_token_match_compiled.n_prompts = 30
  greedy_token_match_compiled.prompts_matched = 29
  greedy_token_match_compiled.reference = hf_greedy_compiled
  greedy_token_match_compiled.source = per-set blocks, by prompt id
[flashspec_compiled_legacy]
  workload_tokens_per_second = 64.77049385924838
  tokens_per_second_median = 69.8607916846708
  tokens_per_second_p25 = 56.60595315651337
  tokens_per_second_p75 = 76.71887150968617
  tokens_per_second_mean = 66.87708099831714
  acceptance_rate_mean = 0.7286482952790188
  gpu_memory_gb = 4.039843559265137
  speedup_vs_arm_composite-g4_median = 0.8723786499980005
  speedup_vs_arm_ngram-g2_median = 1.000549361888451
  speedup_vs_arm_ngram-g4_median = 0.9572820176572617
  speedup_vs_arm_ngram-g6_median = 0.9326499981176062
  speedup_vs_arm_qwen-g2_median = 1.0158725360388707
  speedup_vs_arm_qwen-g4_median = 0.9989589241168761
  speedup_vs_arm_qwen-g6_median = 1.0298396545518083
  speedup_vs_bandit_thompson_median = 0.8774861245875237
  speedup_vs_bandit_ucb1_legacy_median = 0.9295099199687387
  speedup_vs_bandit_ucb1_median = 0.9331685627317419
  forward_counts.emitted_tokens_per_step = 3.8984771573604062
  forward_counts.n_generations = 30
  forward_counts.n_steps = 985
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1015
  forward_counts.target_forwards_per_emitted_token = 0.2643229166666667
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 3840
  step_profile.bookkeeping_fraction = 0.006837288211663942
  step_profile.bookkeeping_ms_mean = 0.41117868824356096
  step_profile.bookkeeping_ms_total = 405.01100791990757
  step_profile.draft_fraction = 0.6587365829964462
  step_profile.draft_ms_mean = 39.614893464993706
  step_profile.draft_ms_total = 39020.6700630188
  step_profile.n_steps = 985.0
  step_profile.sample_fraction = 0.002744967829878545
  step_profile.sample_ms_mean = 0.16507601210005998
  step_profile.sample_ms_total = 162.59987191855907
  step_profile.step_ms_mean = 60.137685514283056
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 3.8984771573604062
  step_profile.verify_fraction = 0.3316811609620113
  step_profile.verify_ms_mean = 19.946537348945736
  step_profile.verify_ms_total = 19647.339288711548
  host_syncs.by_kind = {'cuda': 37}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 34
  host_syncs.per_step = 1.088235294117647
  host_syncs.prompt_id = mt_bench/121
  host_syncs.total = 37
  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

==============================================================================
target=Qwen/Qwen2.5-1.5B-Instruct  drafter=Qwen/Qwen2.5-0.5B-Instruct  dataset=mixed  mode=sample T=0.7
commit=3e38251  n_samples=30
------------------------------------------------------------------------------
[arm_composite-g4]
  workload_tokens_per_second = 73.51670787656015
  tokens_per_second_median = 63.83315490805465
  tokens_per_second_p25 = 58.91178221646625
  tokens_per_second_p75 = 133.2383704947346
  tokens_per_second_mean = 89.57480353780062
  acceptance_rate_mean = 0.4550193832379414
  gpu_memory_gb = 4.043139457702637
  speedup_vs_arm_ngram-g2_median = 1.1208424484741157
  speedup_vs_arm_ngram-g4_median = 0.9845389853280666
  speedup_vs_arm_ngram-g6_median = 0.9392016019584624
  speedup_vs_arm_qwen-g2_median = 1.0207857417698276
  speedup_vs_arm_qwen-g4_median = 1.0911465303659145
  speedup_vs_arm_qwen-g6_median = 1.2618864493314328
  speedup_vs_bandit_thompson_median = 1.0138175271406937
  speedup_vs_bandit_ucb1_legacy_median = 0.9978213751472333
  speedup_vs_bandit_ucb1_median = 1.0131264500612418
  speedup_vs_flashspec_compiled_legacy_median = 1.0933752401624315
  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 = 0.010539991339202645
  step_profile.bookkeeping_ms_mean = 0.40972938388586044
  step_profile.bookkeeping_ms_total = 549.4471037909389
  step_profile.draft_fraction = 0.44372309114929076
  step_profile.draft_ms_mean = 17.2491971673937
  step_profile.draft_ms_total = 23131.173401474953
  step_profile.n_steps = 1341.0
  step_profile.sample_fraction = 0.017639156474557677
  step_profile.sample_ms_mean = 0.6857008209964573
  step_profile.sample_ms_total = 919.5248009562492
  step_profile.step_ms_mean = 38.87378752977767
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 2.8635346756152127
  step_profile.verify_fraction = 0.5280977610369489
  step_profile.verify_ms_mean = 20.529160157501654
  step_profile.verify_ms_total = 27529.603771209717
  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_ngram-g2]
  workload_tokens_per_second = 66.61968010815546
  tokens_per_second_median = 60.67094047143456
  tokens_per_second_p25 = 54.040484691372974
  tokens_per_second_p75 = 95.97660079887714
  tokens_per_second_mean = 74.18246712085728
  acceptance_rate_mean = 0.3466734538890745
  gpu_memory_gb = 4.043139457702637
  speedup_vs_arm_composite-g4_median = 0.8923432194568508
  speedup_vs_arm_ngram-g4_median = 0.9258287957705373
  speedup_vs_arm_ngram-g6_median = 0.9243905358853473
  speedup_vs_arm_qwen-g2_median = 0.9541107067262504
  speedup_vs_arm_qwen-g4_median = 1.0515954106873
  speedup_vs_arm_qwen-g6_median = 1.2039602414190358
  speedup_vs_bandit_thompson_median = 0.9075105411199849
  speedup_vs_bandit_ucb1_legacy_median = 0.8924766406711762
  speedup_vs_bandit_ucb1_median = 0.899289447323558
  speedup_vs_flashspec_compiled_legacy_median = 1.049160104027267
  forward_counts.emitted_tokens_per_step = 1.4312336936265375
  forward_counts.n_generations = 30
  forward_counts.n_steps = 2683
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 2713
  forward_counts.target_forwards_per_emitted_token = 0.7065104166666667
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 3840
  step_profile.bookkeeping_fraction = 0.017588541559752206
  step_profile.bookkeeping_ms_mean = 0.3768521443126009
  step_profile.bookkeeping_ms_total = 1011.0943031907082
  step_profile.draft_fraction = 0.004718274921936216
  step_profile.draft_ms_mean = 0.10109377265576325
  step_profile.draft_ms_total = 271.2345920354128
  step_profile.n_steps = 2683.0
  step_profile.sample_fraction = 0.03160107681571255
  step_profile.sample_ms_mean = 0.6770847668143026
  step_profile.sample_ms_total = 1816.618429362774
  step_profile.step_ms_mean = 21.42600300498765
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 1.4312336936265375
  step_profile.verify_fraction = 0.946092106702599
  step_profile.verify_ms_mean = 20.270972321204983
  step_profile.verify_ms_total = 54387.01873779297
  host_syncs.by_kind = {'cuda': 104}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 102
  host_syncs.per_step = 1.0196078431372548
  host_syncs.prompt_id = mt_bench/121
  host_syncs.total = 104
  nonfinite_logits.prompt_id = mt_bench/121
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[arm_ngram-g4]
  workload_tokens_per_second = 73.04645366961192
  tokens_per_second_median = 66.2839820565969
  tokens_per_second_p25 = 53.27180389613784
  tokens_per_second_p75 = 142.64675524412434
  tokens_per_second_mean = 92.52285086068942
  acceptance_rate_mean = 0.24320480292691596
  gpu_memory_gb = 4.043139457702637
  speedup_vs_arm_composite-g4_median = 1.0157305757387103
  speedup_vs_arm_ngram-g2_median = 1.080514748931313
  speedup_vs_arm_ngram-g6_median = 0.9915231464203035
  speedup_vs_arm_qwen-g2_median = 0.9875249423804126
  speedup_vs_arm_qwen-g4_median = 1.0953370763917007
  speedup_vs_arm_qwen-g6_median = 1.338278750133858
  speedup_vs_bandit_thompson_median = 1.0188931726459631
  speedup_vs_bandit_ucb1_legacy_median = 0.9733489964087789
  speedup_vs_bandit_ucb1_median = 0.9736118907314266
  speedup_vs_flashspec_compiled_legacy_median = 1.0966369708429338
  forward_counts.emitted_tokens_per_step = 1.5927001244296972
  forward_counts.n_generations = 30
  forward_counts.n_steps = 2411
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 2441
  forward_counts.target_forwards_per_emitted_token = 0.6356770833333333
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 3840
  step_profile.bookkeeping_fraction = 0.018246134116944375
  step_profile.bookkeeping_ms_mean = 0.39677576458043745
  step_profile.bookkeeping_ms_total = 956.6263684034348
  step_profile.draft_fraction = 0.004629564596450917
  step_profile.draft_ms_mean = 0.10067332732830736
  step_profile.draft_ms_total = 242.72339218854904
  step_profile.n_steps = 2411.0
  step_profile.sample_fraction = 0.03167436811530909
  step_profile.sample_ms_mean = 0.6887827057504199
  step_profile.sample_ms_total = 1660.6551035642624
  step_profile.step_ms_mean = 21.74574416900561
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 1.5927001244296972
  step_profile.verify_fraction = 0.9454499331712957
  step_profile.verify_ms_mean = 20.559512371346447
  step_profile.verify_ms_total = 49568.984327316284
  host_syncs.by_kind = {'cuda': 112}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 110
  host_syncs.per_step = 1.018181818181818
  host_syncs.prompt_id = mt_bench/121
  host_syncs.total = 112
  nonfinite_logits.prompt_id = mt_bench/121
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[arm_ngram-g6]
  workload_tokens_per_second = 74.69249788795351
  tokens_per_second_median = 67.28383888121273
  tokens_per_second_p25 = 53.41949639360453
  tokens_per_second_p75 = 156.7963332129638
  tokens_per_second_mean = 104.50854974609898
  acceptance_rate_mean = 0.1833123952992227
  gpu_memory_gb = 4.043139457702637
  speedup_vs_arm_composite-g4_median = 1.0647382926977293
  speedup_vs_arm_ngram-g2_median = 1.0822409449766952
  speedup_vs_arm_ngram-g4_median = 1.0085844851271661
  speedup_vs_arm_qwen-g2_median = 1.0104188718958023
  speedup_vs_arm_qwen-g4_median = 1.1307097511353723
  speedup_vs_arm_qwen-g6_median = 1.2634273994229948
  speedup_vs_bandit_thompson_median = 1.0705984500572336
  speedup_vs_bandit_ucb1_legacy_median = 1.018475442812833
  speedup_vs_bandit_ucb1_median = 1.015156012737466
  speedup_vs_flashspec_compiled_legacy_median = 1.1305434091705273
  forward_counts.emitted_tokens_per_step = 1.6473616473616473
  forward_counts.n_generations = 30
  forward_counts.n_steps = 2331
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 2361
  forward_counts.target_forwards_per_emitted_token = 0.61484375
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 3840
  step_profile.bookkeeping_fraction = 0.018629062787996906
  step_profile.bookkeeping_ms_mean = 0.4097874009528676
  step_profile.bookkeeping_ms_total = 955.2144316211343
  step_profile.draft_fraction = 0.004644213851770672
  step_profile.draft_ms_mean = 0.10215974606154866
  step_profile.draft_ms_total = 238.13436806946993
  step_profile.n_steps = 2331.0
  step_profile.sample_fraction = 0.03177782643703546
  step_profile.sample_ms_mean = 0.6990235124417816
  step_profile.sample_ms_total = 1629.423807501793
  step_profile.step_ms_mean = 21.99720971561179
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 1.6473616473616473
  step_profile.verify_fraction = 0.944948896923197
  step_profile.verify_ms_mean = 20.78623905615559
  step_profile.verify_ms_total = 48452.72323989868
  host_syncs.by_kind = {'cuda': 86}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 84
  host_syncs.per_step = 1.0238095238095237
  host_syncs.prompt_id = mt_bench/121
  host_syncs.total = 86
  nonfinite_logits.prompt_id = mt_bench/121
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[arm_qwen-g2]
  workload_tokens_per_second = 64.25245644372939
  tokens_per_second_median = 66.61885402104305
  tokens_per_second_p25 = 59.633410635909705
  tokens_per_second_p75 = 69.12814178034925
  tokens_per_second_mean = 64.83382382000615
  acceptance_rate_mean = 0.8193242910626626
  gpu_memory_gb = 4.043139457702637
  speedup_vs_arm_composite-g4_median = 0.9797648735726006
  speedup_vs_arm_ngram-g2_median = 1.0480977561702265
  speedup_vs_arm_ngram-g4_median = 1.0126456600509548
  speedup_vs_arm_ngram-g6_median = 0.989790494729949
  speedup_vs_arm_qwen-g4_median = 1.0382262737378338
  speedup_vs_arm_qwen-g6_median = 1.0307140834018647
  speedup_vs_bandit_thompson_median = 0.9950384428951482
  speedup_vs_bandit_ucb1_legacy_median = 1.0024922563412288
  speedup_vs_bandit_ucb1_median = 0.9473061162598932
  speedup_vs_flashspec_compiled_legacy_median = 1.033934572604185
  forward_counts.emitted_tokens_per_step = 2.640990371389271
  forward_counts.n_generations = 30
  forward_counts.n_steps = 1454
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1484
  forward_counts.target_forwards_per_emitted_token = 0.38645833333333335
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 3840
  step_profile.bookkeeping_fraction = 0.007521325725560252
  step_profile.bookkeeping_ms_mean = 0.3088004841088757
  step_profile.bookkeeping_ms_total = 448.99590389430523
  step_profile.draft_fraction = 0.5072789919945356
  step_profile.draft_ms_mean = 20.827179146600066
  step_profile.draft_ms_total = 30282.718479156494
  step_profile.n_steps = 1454.0
  step_profile.sample_fraction = 0.009700463370133826
  step_profile.sample_ms_mean = 0.39826858908634893
  step_profile.sample_ms_total = 579.0825285315514
  step_profile.step_ms_mean = 41.05665615031898
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 2.640990371389271
  step_profile.verify_fraction = 0.47549921890977026
  step_profile.verify_ms_mean = 19.52240793052369
  step_profile.verify_ms_total = 28385.581130981445
  host_syncs.by_kind = {'cuda': 51}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 48
  host_syncs.per_step = 1.0625
  host_syncs.prompt_id = mt_bench/121
  host_syncs.total = 51
  nonfinite_logits.prompt_id = mt_bench/121
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[arm_qwen-g4]
  workload_tokens_per_second = 61.973555181507464
  tokens_per_second_median = 63.81691652058262
  tokens_per_second_p25 = 55.63274142062995
  tokens_per_second_p75 = 69.58629332125827
  tokens_per_second_mean = 63.54235962219434
  acceptance_rate_mean = 0.7025707380410877
  gpu_memory_gb = 4.043139457702637
  speedup_vs_arm_composite-g4_median = 0.9167224359953561
  speedup_vs_arm_ngram-g2_median = 0.9509517620033048
  speedup_vs_arm_ngram-g4_median = 0.9130303564939579
  speedup_vs_arm_ngram-g6_median = 0.8846581955601529
  speedup_vs_arm_qwen-g2_median = 0.9632638731133925
  speedup_vs_arm_qwen-g6_median = 1.0228671273363839
  speedup_vs_bandit_thompson_median = 0.8992296861337137
  speedup_vs_bandit_ucb1_legacy_median = 0.9219335833598703
  speedup_vs_bandit_ucb1_median = 0.8960767313884073
  speedup_vs_flashspec_compiled_legacy_median = 1.003041475562351
  forward_counts.emitted_tokens_per_step = 3.7721021611001966
  forward_counts.n_generations = 30
  forward_counts.n_steps = 1018
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1048
  forward_counts.target_forwards_per_emitted_token = 0.27291666666666664
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 3840
  step_profile.bookkeeping_fraction = 0.007065651467401701
  step_profile.bookkeeping_ms_mean = 0.42969722649144754
  step_profile.bookkeeping_ms_total = 437.43177656829357
  step_profile.draft_fraction = 0.656044108118714
  step_profile.draft_ms_mean = 39.897288313080615
  step_profile.draft_ms_total = 40615.439502716064
  step_profile.n_steps = 1018.0
  step_profile.sample_fraction = 0.006770987581514669
  step_profile.sample_ms_mean = 0.4117772575972825
  step_profile.sample_ms_total = 419.1892482340336
  step_profile.step_ms_mean = 60.81494798801093
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 3.7721021611001966
  step_profile.verify_fraction = 0.3301192528323696
  step_profile.verify_ms_mean = 20.076185190841585
  step_profile.verify_ms_total = 20437.556524276733
  host_syncs.by_kind = {'cuda': 38}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 35
  host_syncs.per_step = 1.0857142857142856
  host_syncs.prompt_id = mt_bench/121
  host_syncs.total = 38
  nonfinite_logits.prompt_id = mt_bench/121
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[arm_qwen-g6]
  workload_tokens_per_second = 57.185096449718024
  tokens_per_second_median = 63.52185717469397
  tokens_per_second_p25 = 50.45415313460596
  tokens_per_second_p75 = 74.83929552788464
  tokens_per_second_mean = 61.73321157059031
  acceptance_rate_mean = 0.6018387627120985
  gpu_memory_gb = 4.043139457702637
  speedup_vs_arm_composite-g4_median = 0.7926076381246835
  speedup_vs_arm_ngram-g2_median = 0.8306323802901865
  speedup_vs_arm_ngram-g4_median = 0.7479189072739216
  speedup_vs_arm_ngram-g6_median = 0.7915064895667813
  speedup_vs_arm_qwen-g2_median = 0.9702011697115142
  speedup_vs_arm_qwen-g4_median = 0.9776475307319572
  speedup_vs_bandit_thompson_median = 0.8032591063215418
  speedup_vs_bandit_ucb1_legacy_median = 0.8041339616174445
  speedup_vs_bandit_ucb1_median = 0.7492334786610114
  speedup_vs_flashspec_compiled_legacy_median = 0.9761366222676526
  forward_counts.emitted_tokens_per_step = 4.5933014354066986
  forward_counts.n_generations = 30
  forward_counts.n_steps = 836
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 866
  forward_counts.target_forwards_per_emitted_token = 0.22552083333333334
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 3840
  step_profile.bookkeeping_fraction = 0.006201616379817446
  step_profile.bookkeeping_ms_mean = 0.4978042105881221
  step_profile.bookkeeping_ms_total = 416.1643200516701
  step_profile.draft_fraction = 0.7334844853902132
  step_profile.draft_ms_mean = 58.87685449499833
  step_profile.draft_ms_total = 49221.0503578186
  step_profile.n_steps = 836.0
  step_profile.sample_fraction = 0.005248120009519816
  step_profile.sample_ms_mean = 0.4212669856382899
  step_profile.sample_ms_total = 352.1791999936104
  step_profile.step_ms_mean = 80.27007478375754
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 4.5933014354066986
  step_profile.verify_fraction = 0.25506577822044957
  step_profile.verify_ms_mean = 20.4741490925328
  step_profile.verify_ms_total = 17116.388641357422
  host_syncs.by_kind = {'cuda': 30}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 27
  host_syncs.per_step = 1.1111111111111112
  host_syncs.prompt_id = mt_bench/121
  host_syncs.total = 30
  nonfinite_logits.prompt_id = mt_bench/121
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[bandit_thompson]
  workload_tokens_per_second = 73.15578604574415
  tokens_per_second_median = 65.19713489299605
  tokens_per_second_p25 = 56.58062512791939
  tokens_per_second_p75 = 131.66638462943354
  tokens_per_second_mean = 88.49924995664449
  acceptance_rate_mean = 0.4515649508568427
  gpu_memory_gb = 4.043139457702637
  speedup_vs_arm_composite-g4_median = 0.9863712807601308
  speedup_vs_arm_ngram-g2_median = 1.101915656400194
  speedup_vs_arm_ngram-g4_median = 0.9814633469394548
  speedup_vs_arm_ngram-g6_median = 0.9341259792359708
  speedup_vs_arm_qwen-g2_median = 1.0053026864709398
  speedup_vs_arm_qwen-g4_median = 1.1122110874196207
  speedup_vs_arm_qwen-g6_median = 1.2450473978467476
  speedup_vs_bandit_ucb1_legacy_median = 0.996993750065753
  speedup_vs_bandit_ucb1_median = 0.9962990261500457
  speedup_vs_flashspec_compiled_legacy_median = 1.1112289199198957
  forward_counts.emitted_tokens_per_step = 2.7214741318214033
  forward_counts.n_generations = 30
  forward_counts.n_steps = 1411
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1441
  forward_counts.target_forwards_per_emitted_token = 0.37526041666666665
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 3840
  step_profile.bookkeeping_fraction = 0.010861676233826197
  step_profile.bookkeeping_ms_mean = 0.4032236542433902
  step_profile.bookkeeping_ms_total = 568.9485761374235
  step_profile.draft_fraction = 0.4109500199186756
  step_profile.draft_ms_mean = 15.255911258609688
  step_profile.draft_ms_total = 21526.09078589827
  step_profile.n_steps = 1411.0
  step_profile.sample_fraction = 0.01848376623143081
  step_profile.sample_ms_mean = 0.6861824642505145
  step_profile.sample_ms_total = 968.203457057476
  step_profile.step_ms_mean = 37.12351993955065
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 2.7214741318214033
  step_profile.verify_fraction = 0.5597045376160674
  step_profile.verify_ms_mean = 20.778202562447053
  step_profile.verify_ms_total = 29318.043815612793
  host_syncs.by_kind = {'cuda': 51}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 47
  host_syncs.per_step = 1.0851063829787233
  host_syncs.prompt_id = mt_bench/121
  host_syncs.total = 51
  nonfinite_logits.prompt_id = mt_bench/121
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[bandit_ucb1]
  workload_tokens_per_second = 72.4274254670285
  tokens_per_second_median = 70.64922827407952
  tokens_per_second_p25 = 57.77335446935691
  tokens_per_second_p75 = 101.77403840736075
  tokens_per_second_mean = 88.93683616391426
  acceptance_rate_mean = 0.5149190648974282
  gpu_memory_gb = 4.043139457702637
  speedup_vs_arm_composite-g4_median = 0.9870487537312368
  speedup_vs_arm_ngram-g2_median = 1.1119925115965064
  speedup_vs_arm_ngram-g4_median = 1.0271753803365773
  speedup_vs_arm_ngram-g6_median = 0.9851773860273003
  speedup_vs_arm_qwen-g2_median = 1.0559777614483705
  speedup_vs_arm_qwen-g4_median = 1.1160934235100828
  speedup_vs_arm_qwen-g6_median = 1.3347035875896074
  speedup_vs_bandit_thompson_median = 1.0037259693465628
  speedup_vs_bandit_ucb1_legacy_median = 1.0102997976299153
  speedup_vs_flashspec_compiled_legacy_median = 1.1089554130137644
  forward_counts.emitted_tokens_per_step = 2.6928471248246844
  forward_counts.n_generations = 30
  forward_counts.n_steps = 1426
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1456
  forward_counts.target_forwards_per_emitted_token = 0.37916666666666665
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 3840
  step_profile.bookkeeping_fraction = 0.01074126161487204
  step_profile.bookkeeping_ms_mean = 0.39855771094853215
  step_profile.bookkeeping_ms_total = 568.3432958126068
  step_profile.draft_fraction = 0.4124388811076316
  step_profile.draft_ms_mean = 15.303667507067761
  step_profile.draft_ms_total = 21823.029865078628
  step_profile.n_steps = 1426.0
  step_profile.sample_fraction = 0.018666677837394792
  step_profile.sample_ms_mean = 0.692632640060168
  step_profile.sample_ms_total = 987.6941447257996
  step_profile.step_ms_mean = 37.105297798230765
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 2.6928471248246844
  step_profile.verify_fraction = 0.5581531794401016
  step_profile.verify_ms_mean = 20.710439940154302
  step_profile.verify_ms_total = 29533.087354660034
  host_syncs.by_kind = {'cuda': 53}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 49
  host_syncs.per_step = 1.0816326530612246
  host_syncs.prompt_id = mt_bench/121
  host_syncs.total = 53
  nonfinite_logits.prompt_id = mt_bench/121
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[bandit_ucb1_legacy]
  workload_tokens_per_second = 72.07069363483879
  tokens_per_second_median = 66.23037531851578
  tokens_per_second_p25 = 54.37141168226234
  tokens_per_second_p75 = 120.20213483035229
  tokens_per_second_mean = 92.11336354405358
  acceptance_rate_mean = 0.47460558823160376
  gpu_memory_gb = 4.043139457702637
  speedup_vs_arm_composite-g4_median = 1.0022774232996121
  speedup_vs_arm_ngram-g2_median = 1.1204789120777061
  speedup_vs_arm_ngram-g4_median = 1.0273912055696761
  speedup_vs_arm_ngram-g6_median = 0.9819698234802401
  speedup_vs_arm_qwen-g2_median = 0.9975252924959808
  speedup_vs_arm_qwen-g4_median = 1.0853372414661224
  speedup_vs_arm_qwen-g6_median = 1.2440516704702185
  speedup_vs_bandit_thompson_median = 1.003188242089129
  speedup_vs_bandit_ucb1_median = 0.9898210483302352
  speedup_vs_flashspec_compiled_legacy_median = 1.088793084872885
  forward_counts.emitted_tokens_per_step = 2.518032786885246
  forward_counts.n_generations = 30
  forward_counts.n_steps = 1525
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1555
  forward_counts.target_forwards_per_emitted_token = 0.4049479166666667
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 3840
  step_profile.bookkeeping_fraction = 0.01147554078282889
  step_profile.bookkeeping_ms_mean = 0.40006358014267
  step_profile.bookkeeping_ms_total = 610.0969597175717
  step_profile.draft_fraction = 0.38451321790082565
  step_profile.draft_ms_mean = 13.405009618000905
  step_profile.draft_ms_total = 20442.63966745138
  step_profile.n_steps = 1525.0
  step_profile.sample_fraction = 0.019817830211349623
  step_profile.sample_ms_mean = 0.6908948567265369
  step_profile.sample_ms_total = 1053.614656507969
  step_profile.step_ms_mean = 34.86228559627396
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 2.518032786885246
  step_profile.verify_fraction = 0.5841934111049958
  step_profile.verify_ms_mean = 20.36631754140385
  step_profile.verify_ms_total = 31058.63425064087
  host_syncs.by_kind = {'cuda': 51}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 47
  host_syncs.per_step = 1.0851063829787233
  host_syncs.prompt_id = mt_bench/121
  host_syncs.total = 51
  nonfinite_logits.prompt_id = mt_bench/121
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[flashspec_compiled_legacy]
  workload_tokens_per_second = 61.927292857811615
  tokens_per_second_median = 63.421710243622655
  tokens_per_second_p25 = 56.08554655436377
  tokens_per_second_p75 = 68.64315297490492
  tokens_per_second_mean = 63.44138963935148
  acceptance_rate_mean = 0.7025707380410877
  gpu_memory_gb = 4.043139457702637
  speedup_vs_arm_composite-g4_median = 0.9149025964677968
  speedup_vs_arm_ngram-g2_median = 0.9531449007009378
  speedup_vs_arm_ngram-g4_median = 0.9118839980812384
  speedup_vs_arm_ngram-g6_median = 0.8846098875188488
  speedup_vs_arm_qwen-g2_median = 0.9672006493980851
  speedup_vs_arm_qwen-g4_median = 0.9969677907522974
  speedup_vs_arm_qwen-g6_median = 1.0244649399590813
  speedup_vs_bandit_thompson_median = 0.8999506934320832
  speedup_vs_bandit_ucb1_legacy_median = 0.9187588809852187
  speedup_vs_bandit_ucb1_median = 0.90181057417253
  forward_counts.emitted_tokens_per_step = 3.7721021611001966
  forward_counts.n_generations = 30
  forward_counts.n_steps = 1018
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1048
  forward_counts.target_forwards_per_emitted_token = 0.27291666666666664
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 3840
  step_profile.bookkeeping_fraction = 0.0069731443004236926
  step_profile.bookkeeping_ms_mean = 0.4243905067385297
  step_profile.bookkeeping_ms_total = 432.0295358598232
  step_profile.draft_fraction = 0.6588700959261553
  step_profile.draft_ms_mean = 40.09930123889376
  step_profile.draft_ms_total = 40821.08866119385
  step_profile.n_steps = 1018.0
  step_profile.sample_fraction = 0.006795530840843305
  step_profile.sample_ms_mean = 0.41358082564382986
  step_profile.sample_ms_total = 421.0252805054188
  step_profile.step_ms_mean = 60.860709093994146
  step_profile.tokens_emitted = 3840.0
  step_profile.tokens_per_step_mean = 3.7721021611001966
  step_profile.verify_fraction = 0.32736122893257763
  step_profile.verify_ms_mean = 19.923436522718028
  step_profile.verify_ms_total = 20282.058380126953
  host_syncs.by_kind = {'cuda': 38}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 35
  host_syncs.per_step = 1.0857142857142856
  host_syncs.prompt_id = mt_bench/121
  host_syncs.total = 38
  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=3e38251  n_samples=10
------------------------------------------------------------------------------
[arm_composite-g4]
  workload_tokens_per_second = 56.82215431935964
  tokens_per_second_median = 55.0591984541738
  tokens_per_second_p25 = 53.218150770832224
  tokens_per_second_p75 = 62.514588785472625
  tokens_per_second_mean = 57.91880540911844
  acceptance_rate_mean = 0.36448637869169703
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_ngram-g2_median = 1.0862177330066818
  speedup_vs_arm_ngram-g4_median = 1.0859653124997495
  speedup_vs_arm_ngram-g6_median = 1.0927299879126529
  speedup_vs_arm_qwen-g2_median = 0.9655741338749915
  speedup_vs_arm_qwen-g4_median = 1.043894286037364
  speedup_vs_arm_qwen-g6_median = 1.200485430056768
  speedup_vs_bandit_thompson_median = 1.0148021799132794
  speedup_vs_bandit_ucb1_median = 1.0982139278722034
  speedup_vs_hf_assisted_median = 3.2040425778258026
  speedup_vs_hf_greedy_compiled_median = 1.1579411273420512
  speedup_vs_hf_greedy_median = 2.2592748635113535
  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 = 0.010997559400739614
  step_profile.bookkeeping_ms_mean = 0.4648608105971401
  step_profile.bookkeeping_ms_total = 247.30595123767853
  step_profile.draft_fraction = 0.5020091857982137
  step_profile.draft_ms_mean = 21.219653246126004
  step_profile.draft_ms_total = 11288.855526939034
  step_profile.n_steps = 532.0
  step_profile.sample_fraction = 0.010332326350318423
  step_profile.sample_ms_mean = 0.4367417740194421
  step_profile.sample_ms_total = 232.3466237783432
  step_profile.step_ms_mean = 42.269452126430615
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.406015037593985
  step_profile.verify_fraction = 0.47666092845072827
  step_profile.verify_ms_mean = 20.148196295688027
  step_profile.verify_ms_total = 10718.84042930603
  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_ngram-g2]
  workload_tokens_per_second = 54.23677455470492
  tokens_per_second_median = 53.66227723372641
  tokens_per_second_p25 = 51.6283286702775
  tokens_per_second_p75 = 58.1427520850623
  tokens_per_second_mean = 54.523160602104426
  acceptance_rate_mean = 0.14265616767345787
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.9206729491798036
  speedup_vs_arm_ngram-g4_median = 0.9993727746213037
  speedup_vs_arm_ngram-g6_median = 1.002437194139174
  speedup_vs_arm_qwen-g2_median = 0.8928985963338275
  speedup_vs_arm_qwen-g4_median = 0.971904713653138
  speedup_vs_arm_qwen-g6_median = 1.1358036630858992
  speedup_vs_bandit_thompson_median = 0.9166082142282985
  speedup_vs_bandit_ucb1_median = 1.0276678489388689
  speedup_vs_hf_assisted_median = 3.00217275014883
  speedup_vs_hf_greedy_compiled_median = 1.1025051011626283
  speedup_vs_hf_greedy_median = 2.210998575032458
  forward_counts.emitted_tokens_per_step = 1.1387900355871887
  forward_counts.n_generations = 10
  forward_counts.n_steps = 1124
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1134
  forward_counts.target_forwards_per_emitted_token = 0.8859375
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.019670260518805265
  step_profile.bookkeeping_ms_mean = 0.41195516001923654
  step_profile.bookkeeping_ms_total = 463.03759986162186
  step_profile.draft_fraction = 0.004558733602749923
  step_profile.draft_ms_mean = 0.09547376502769274
  step_profile.draft_ms_total = 107.31251189112663
  step_profile.n_steps = 1124.0
  step_profile.sample_fraction = 0.020391695187185522
  step_profile.sample_ms_mean = 0.4270641990668833
  step_profile.sample_ms_total = 480.02015975117683
  step_profile.step_ms_mean = 20.943045447994805
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 1.1387900355871887
  step_profile.verify_fraction = 0.9553793106912593
  step_profile.verify_ms_mean = 20.008552323880995
  step_profile.verify_ms_total = 22489.612812042236
  host_syncs.by_kind = {'cuda': 127}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 125
  host_syncs.per_step = 1.016
  host_syncs.prompt_id = 81
  host_syncs.total = 127
  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_ngram-g4]
  workload_tokens_per_second = 54.56594062718833
  tokens_per_second_median = 53.63983308870911
  tokens_per_second_p25 = 51.607569202230806
  tokens_per_second_p75 = 59.1209567956674
  tokens_per_second_mean = 54.958627209005556
  acceptance_rate_mean = 0.08638257328631224
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.9208406992864471
  speedup_vs_arm_ngram-g2_median = 1.0006284058855297
  speedup_vs_arm_ngram-g6_median = 1.0043894806789875
  speedup_vs_arm_qwen-g2_median = 0.8986387042654012
  speedup_vs_arm_qwen-g4_median = 0.9703115907826658
  speedup_vs_arm_qwen-g6_median = 1.146920484460796
  speedup_vs_bandit_thompson_median = 0.9255702054520187
  speedup_vs_bandit_ucb1_median = 1.0234156996176527
  speedup_vs_hf_assisted_median = 3.068019212182306
  speedup_vs_hf_greedy_compiled_median = 1.1006980729429618
  speedup_vs_hf_greedy_median = 2.2198811810240926
  forward_counts.emitted_tokens_per_step = 1.1594202898550725
  forward_counts.n_generations = 10
  forward_counts.n_steps = 1104
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1114
  forward_counts.target_forwards_per_emitted_token = 0.8703125
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.01984724105388465
  step_profile.bookkeeping_ms_mean = 0.42069226113296504
  step_profile.bookkeeping_ms_total = 464.4442562907934
  step_profile.draft_fraction = 0.004474298037006947
  step_profile.draft_ms_mean = 0.09483950706603735
  step_profile.draft_ms_total = 104.70281580090523
  step_profile.n_steps = 1104.0
  step_profile.sample_fraction = 0.020455741507389467
  step_profile.sample_ms_mean = 0.4335903476221838
  step_profile.sample_ms_total = 478.6837437748909
  step_profile.step_ms_mean = 21.196510889891375
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 1.1594202898550725
  step_profile.verify_fraction = 0.955222719401719
  step_profile.verify_ms_mean = 20.24738877407019
  step_profile.verify_ms_total = 22353.117206573486
  host_syncs.by_kind = {'cuda': 127}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 125
  host_syncs.per_step = 1.016
  host_syncs.prompt_id = 81
  host_syncs.total = 127
  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_ngram-g6]
  workload_tokens_per_second = 54.89308401528911
  tokens_per_second_median = 53.29465454910431
  tokens_per_second_p25 = 51.29771578021103
  tokens_per_second_p75 = 59.53104774642347
  tokens_per_second_mean = 55.64277602334763
  acceptance_rate_mean = 0.06882770402530286
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.9151441621697857
  speedup_vs_arm_ngram-g2_median = 0.997570304143285
  speedup_vs_arm_ngram-g4_median = 0.995630599872162
  speedup_vs_arm_qwen-g2_median = 0.899928737696862
  speedup_vs_arm_qwen-g4_median = 0.9650877342346607
  speedup_vs_arm_qwen-g6_median = 1.1419645224972657
  speedup_vs_bandit_thompson_median = 0.9207016640146038
  speedup_vs_bandit_ucb1_median = 1.0318854248337037
  speedup_vs_hf_assisted_median = 3.0775086226545074
  speedup_vs_hf_greedy_compiled_median = 1.0944393419967786
  speedup_vs_hf_greedy_median = 2.2055823918278805
  forward_counts.emitted_tokens_per_step = 1.1764705882352942
  forward_counts.n_generations = 10
  forward_counts.n_steps = 1088
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1098
  forward_counts.target_forwards_per_emitted_token = 0.8578125
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.01958594283608823
  step_profile.bookkeeping_ms_mean = 0.41873776533997015
  step_profile.bookkeeping_ms_total = 455.5866886898875
  step_profile.draft_fraction = 0.004389125144855142
  step_profile.draft_ms_mean = 0.09383732355062795
  step_profile.draft_ms_total = 102.09500802308321
  step_profile.n_steps = 1088.0
  step_profile.sample_fraction = 0.020609816447246975
  step_profile.sample_ms_mean = 0.4406276764622068
  step_profile.sample_ms_total = 479.40291199088097
  step_profile.step_ms_mean = 21.379505129996687
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 1.1764705882352942
  step_profile.verify_fraction = 0.9554151155718097
  step_profile.verify_ms_mean = 20.42630236464388
  step_profile.verify_ms_total = 22223.816972732544
  host_syncs.by_kind = {'cuda': 129}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 127
  host_syncs.per_step = 1.015748031496063
  host_syncs.prompt_id = 81
  host_syncs.total = 129
  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
[arm_qwen-g2]
  workload_tokens_per_second = 59.12923528903212
  tokens_per_second_median = 57.937483130195304
  tokens_per_second_p25 = 55.53444053674029
  tokens_per_second_p75 = 61.96696290086642
  tokens_per_second_mean = 59.552665468236306
  acceptance_rate_mean = 0.7046860972016337
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 1.03572448976565
  speedup_vs_arm_ngram-g2_median = 1.1200848768254117
  speedup_vs_arm_ngram-g4_median = 1.112807064535978
  speedup_vs_arm_ngram-g6_median = 1.111264000884687
  speedup_vs_arm_qwen-g4_median = 1.0894033094060127
  speedup_vs_arm_qwen-g6_median = 1.252007628145663
  speedup_vs_bandit_thompson_median = 1.0424455866839957
  speedup_vs_bandit_ucb1_median = 1.1414168532881095
  speedup_vs_hf_assisted_median = 3.06292135178115
  speedup_vs_hf_greedy_compiled_median = 1.229711599542838
  speedup_vs_hf_greedy_median = 2.369629590621493
  forward_counts.emitted_tokens_per_step = 2.392523364485981
  forward_counts.n_generations = 10
  forward_counts.n_steps = 535
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 545
  forward_counts.target_forwards_per_emitted_token = 0.42578125
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.008995076768667717
  step_profile.bookkeeping_ms_mean = 0.3635619889909976
  step_profile.bookkeeping_ms_total = 194.50566411018372
  step_profile.draft_fraction = 0.508769232096773
  step_profile.draft_ms_mean = 20.563376913338065
  step_profile.draft_ms_total = 11001.406648635864
  step_profile.n_steps = 535.0
  step_profile.sample_fraction = 0.00376973845639678
  step_profile.sample_ms_mean = 0.15236485984837897
  step_profile.sample_ms_total = 81.51520001888275
  step_profile.step_ms_mean = 40.41788617717886
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.392523364485981
  step_profile.verify_fraction = 0.47846595267816244
  step_profile.verify_ms_mean = 19.338582415001415
  step_profile.verify_ms_total = 10346.141592025757
  host_syncs.by_kind = {'cuda': 59}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 56
  host_syncs.per_step = 1.0535714285714286
  host_syncs.prompt_id = 81
  host_syncs.total = 59
  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_qwen-g4]
  workload_tokens_per_second = 53.50398304775462
  tokens_per_second_median = 53.61830195857739
  tokens_per_second_p25 = 49.72825270474492
  tokens_per_second_p75 = 55.46301060479976
  tokens_per_second_mean = 54.723764943277374
  acceptance_rate_mean = 0.5578437367969221
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.9580418527482321
  speedup_vs_arm_ngram-g2_median = 1.0290626155458615
  speedup_vs_arm_ngram-g4_median = 1.0307896527085583
  speedup_vs_arm_ngram-g6_median = 1.0361759526703098
  speedup_vs_arm_qwen-g2_median = 0.9179376719381992
  speedup_vs_arm_qwen-g6_median = 1.133804001575235
  speedup_vs_bandit_thompson_median = 0.9673862920802467
  speedup_vs_bandit_ucb1_median = 1.0426838097717472
  speedup_vs_hf_assisted_median = 2.995075027981428
  speedup_vs_hf_greedy_compiled_median = 1.1073835311562061
  speedup_vs_hf_greedy_median = 2.18826574322204
  forward_counts.emitted_tokens_per_step = 3.18407960199005
  forward_counts.n_generations = 10
  forward_counts.n_steps = 402
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 412
  forward_counts.target_forwards_per_emitted_token = 0.321875
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.00845943038667968
  step_profile.bookkeeping_ms_mean = 0.5030313220931523
  step_profile.bookkeeping_ms_total = 202.21859148144722
  step_profile.draft_fraction = 0.6576327328825572
  step_profile.draft_ms_mean = 39.105453671507576
  step_profile.draft_ms_total = 15720.392375946045
  step_profile.n_steps = 402.0
  step_profile.sample_fraction = 0.002656241153599562
  step_profile.sample_ms_mean = 0.157950646582231
  step_profile.sample_ms_total = 63.49615992605686
  step_profile.step_ms_mean = 59.4639708703356
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 3.18407960199005
  step_profile.verify_fraction = 0.33125159557716355
  step_profile.verify_ms_mean = 19.697535230152642
  step_profile.verify_ms_total = 7918.409162521362
  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.prompt_id = 81
  host_syncs.total = 47
  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_qwen-g6]
  workload_tokens_per_second = 47.65504887037378
  tokens_per_second_median = 48.79914901469826
  tokens_per_second_p25 = 42.56349261384382
  tokens_per_second_p75 = 52.163381203461924
  tokens_per_second_mean = 49.93677420800062
  acceptance_rate_mean = 0.4873338799874727
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.832998337826526
  speedup_vs_arm_ngram-g2_median = 0.8804686365355211
  speedup_vs_arm_ngram-g4_median = 0.8721281058858292
  speedup_vs_arm_ngram-g6_median = 0.8758885935856171
  speedup_vs_arm_qwen-g2_median = 0.7987897210962438
  speedup_vs_arm_qwen-g4_median = 0.8821098870389994
  speedup_vs_bandit_thompson_median = 0.8205895687406595
  speedup_vs_bandit_ucb1_median = 0.9002538069724006
  speedup_vs_hf_assisted_median = 2.839644699180094
  speedup_vs_hf_greedy_compiled_median = 1.0013918678785587
  speedup_vs_hf_greedy_median = 2.0097979863139814
  forward_counts.emitted_tokens_per_step = 3.7317784256559765
  forward_counts.n_generations = 10
  forward_counts.n_steps = 343
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 353
  forward_counts.target_forwards_per_emitted_token = 0.27578125
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.007552636779698469
  step_profile.bookkeeping_ms_mean = 0.5910508224724333
  step_profile.bookkeeping_ms_total = 202.73043210804462
  step_profile.draft_fraction = 0.7349926264961008
  step_profile.draft_ms_mean = 57.518719497992066
  step_profile.draft_ms_total = 19728.92078781128
  step_profile.n_steps = 343.0
  step_profile.sample_fraction = 0.002160063373040023
  step_profile.sample_ms_mean = 0.16904125942606968
  step_profile.sample_ms_total = 57.9811519831419
  step_profile.step_ms_mean = 78.2575462997481
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 3.7317784256559765
  step_profile.verify_fraction = 0.25529467335116074
  step_profile.verify_ms_mean = 19.97873471985753
  step_profile.verify_ms_total = 6852.706008911133
  host_syncs.by_kind = {'cuda': 44}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 41
  host_syncs.per_step = 1.0731707317073171
  host_syncs.prompt_id = 81
  host_syncs.total = 44
  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
[bandit_thompson]
  workload_tokens_per_second = 56.4619612719059
  tokens_per_second_median = 58.850651201562
  tokens_per_second_p25 = 53.13243513141559
  tokens_per_second_p75 = 64.49551002675884
  tokens_per_second_mean = 57.55962443181998
  acceptance_rate_mean = 0.30803078298515385
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.9855148813408257
  speedup_vs_arm_ngram-g2_median = 1.0912090727665409
  speedup_vs_arm_ngram-g4_median = 1.080417976547051
  speedup_vs_arm_ngram-g6_median = 1.086128837868269
  speedup_vs_arm_qwen-g2_median = 0.9593831285286041
  speedup_vs_arm_qwen-g4_median = 1.0337151406621734
  speedup_vs_arm_qwen-g6_median = 1.2186833153410659
  speedup_vs_bandit_ucb1_median = 1.1021204604237673
  speedup_vs_hf_assisted_median = 3.344012516348939
  speedup_vs_hf_greedy_compiled_median = 1.207080430535157
  speedup_vs_hf_greedy_median = 2.429127720958199
  forward_counts.emitted_tokens_per_step = 2.112211221122112
  forward_counts.n_generations = 10
  forward_counts.n_steps = 606
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 616
  forward_counts.target_forwards_per_emitted_token = 0.48125
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.01215547965285104
  step_profile.bookkeeping_ms_mean = 0.4539280796651006
  step_profile.bookkeeping_ms_total = 275.080416277051
  step_profile.draft_fraction = 0.4310680083790857
  step_profile.draft_ms_mean = 16.097585520015496
  step_profile.draft_ms_total = 9755.13682512939
  step_profile.n_steps = 606.0
  step_profile.sample_fraction = 0.01173482607958197
  step_profile.sample_ms_mean = 0.4382194055385715
  step_profile.sample_ms_total = 265.56095975637436
  step_profile.step_ms_mean = 37.34349385041608
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.112211221122112
  step_profile.verify_fraction = 0.5450416858884813
  step_profile.verify_ms_mean = 20.353760845196916
  step_profile.verify_ms_total = 12334.379072189331
  host_syncs.by_kind = {'cuda': 74}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 70
  host_syncs.per_step = 1.0571428571428572
  host_syncs.prompt_id = 81
  host_syncs.total = 74
  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
[bandit_ucb1]
  workload_tokens_per_second = 53.3419607020316
  tokens_per_second_median = 51.89663091545216
  tokens_per_second_p25 = 47.81304010681093
  tokens_per_second_p75 = 58.14188797308087
  tokens_per_second_mean = 54.69722778630162
  acceptance_rate_mean = 0.4297374040188708
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.910575052153159
  speedup_vs_arm_ngram-g2_median = 0.9731483034572195
  speedup_vs_arm_ngram-g4_median = 0.9771415916998609
  speedup_vs_arm_ngram-g6_median = 0.9691392602670355
  speedup_vs_arm_qwen-g2_median = 0.87613407203823
  speedup_vs_arm_qwen-g4_median = 0.9591076503629675
  speedup_vs_arm_qwen-g6_median = 1.110806227911802
  speedup_vs_bandit_thompson_median = 0.9073428943950096
  speedup_vs_hf_assisted_median = 3.095282406608888
  speedup_vs_hf_greedy_compiled_median = 1.0662123985933682
  speedup_vs_hf_greedy_median = 2.1376165539281082
  forward_counts.emitted_tokens_per_step = 2.17687074829932
  forward_counts.n_generations = 10
  forward_counts.n_steps = 588
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 598
  forward_counts.target_forwards_per_emitted_token = 0.4671875
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.010849858530007507
  step_profile.bookkeeping_ms_mean = 0.44201844910375115
  step_profile.bookkeeping_ms_total = 259.9068480730057
  step_profile.draft_fraction = 0.4810236318873488
  step_profile.draft_ms_mean = 19.596690515461727
  step_profile.draft_ms_total = 11522.854023091495
  step_profile.n_steps = 588.0
  step_profile.sample_fraction = 0.01073070375109381
  step_profile.sample_ms_mean = 0.4371641359868504
  step_profile.sample_ms_total = 257.052511960268
  step_profile.step_ms_mean = 40.739558758416855
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.17687074829932
  step_profile.verify_fraction = 0.49739580583154985
  step_profile.verify_ms_mean = 20.263685657864524
  step_profile.verify_ms_total = 11915.04716682434
  host_syncs.by_kind = {'cuda': 73}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 69
  host_syncs.per_step = 1.0579710144927537
  host_syncs.prompt_id = 81
  host_syncs.total = 73
  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_assisted]
  workload_tokens_per_second = 16.74767289231452
  tokens_per_second_median = 19.757504021154205
  tokens_per_second_p25 = 14.873685161385204
  tokens_per_second_p75 = 21.400294324364737
  tokens_per_second_mean = 18.09610578801857
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.3126178186343996
  speedup_vs_arm_ngram-g2_median = 0.33545514962294853
  speedup_vs_arm_ngram-g4_median = 0.32782960470097455
  speedup_vs_arm_ngram-g6_median = 0.32651517538117236
  speedup_vs_arm_qwen-g2_median = 0.3267170722539938
  speedup_vs_arm_qwen-g4_median = 0.3339157087769203
  speedup_vs_arm_qwen-g6_median = 0.35215920387127686
  speedup_vs_bandit_thompson_median = 0.300438883697958
  speedup_vs_bandit_ucb1_median = 0.3235178854107241
  speedup_vs_hf_greedy_compiled_median = 0.42678121628555077
  speedup_vs_hf_greedy_median = 0.7990304101430465
  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
  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 = 24.547322177763874
  tokens_per_second_median = 24.536197301053008
  tokens_per_second_p25 = 24.35954533043947
  tokens_per_second_p75 = 24.68300820846031
  tokens_per_second_mean = 24.55091099028016
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.44265586934654233
  speedup_vs_arm_ngram-g2_median = 0.45239918780849797
  speedup_vs_arm_ngram-g4_median = 0.45053525746189793
  speedup_vs_arm_ngram-g6_median = 0.4534926990095708
  speedup_vs_arm_qwen-g2_median = 0.42265312137986677
  speedup_vs_arm_qwen-g4_median = 0.4575445396022134
  speedup_vs_arm_qwen-g6_median = 0.49760993461477787
  speedup_vs_bandit_thompson_median = 0.41224923238238986
  speedup_vs_bandit_ucb1_median = 0.4682708010606339
  speedup_vs_hf_assisted_median = 1.2562245546111008
  speedup_vs_hf_greedy_compiled_median = 0.5034144385869986
  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 = 48.141860106610245
  tokens_per_second_median = 48.67240204291895
  tokens_per_second_p25 = 48.50775500291482
  tokens_per_second_p75 = 48.84707091833104
  tokens_per_second_mean = 48.1907622217913
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.8637735057023073
  speedup_vs_arm_ngram-g2_median = 0.9071645613173123
  speedup_vs_arm_ngram-g4_median = 0.9086871632741838
  speedup_vs_arm_ngram-g6_median = 0.9139466456173739
  speedup_vs_arm_qwen-g2_median = 0.8132049030299899
  speedup_vs_arm_qwen-g4_median = 0.9035612558005638
  speedup_vs_arm_qwen-g6_median = 0.9987648470309864
  speedup_vs_bandit_thompson_median = 0.82919052170076
  speedup_vs_bandit_ucb1_median = 0.9386071314833162
  speedup_vs_hf_assisted_median = 2.343169120759474
  speedup_vs_hf_greedy_median = 1.986489026836301
  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=3e38251  n_samples=10
------------------------------------------------------------------------------
[arm_composite-g4]
  workload_tokens_per_second = 54.58859204822758
  tokens_per_second_median = 53.92420091436816
  tokens_per_second_p25 = 50.28211903654267
  tokens_per_second_p75 = 60.877036742115685
  tokens_per_second_mean = 55.56355802518901
  acceptance_rate_mean = 0.353721257472155
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_ngram-g2_median = 1.0413854936506848
  speedup_vs_arm_ngram-g4_median = 1.0124662427343494
  speedup_vs_arm_ngram-g6_median = 0.9983326712175676
  speedup_vs_arm_qwen-g2_median = 0.9425465645284161
  speedup_vs_arm_qwen-g4_median = 1.0065299601850008
  speedup_vs_arm_qwen-g6_median = 1.2369207560402051
  speedup_vs_bandit_thompson_median = 1.0368822808219664
  speedup_vs_bandit_ucb1_median = 1.0415398782733114
  speedup_vs_hf_assisted_median = 4.046556444024923
  speedup_vs_hf_sample_compiled_median = 1.185016247139596
  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 = 0.010937565686705076
  step_profile.bookkeeping_ms_mean = 0.4777076421305537
  step_profile.bookkeeping_ms_total = 256.0512961819768
  step_profile.draft_fraction = 0.5106582200635481
  step_profile.draft_ms_mean = 22.303439469869012
  step_profile.draft_ms_total = 11954.64355584979
  step_profile.n_steps = 536.0
  step_profile.sample_fraction = 0.01594742156730579
  step_profile.sample_ms_mean = 0.6965174311354979
  step_profile.sample_ms_total = 373.33334308862686
  step_profile.step_ms_mean = 43.67586497891583
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.388059701492537
  step_profile.verify_fraction = 0.46245679268244105
  step_profile.verify_ms_mean = 20.198200435780766
  step_profile.verify_ms_total = 10826.235433578491
  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_ngram-g2]
  workload_tokens_per_second = 53.67526329864018
  tokens_per_second_median = 52.99207039887425
  tokens_per_second_p25 = 49.65562342458037
  tokens_per_second_p75 = 57.84520645281031
  tokens_per_second_mean = 54.166893352410604
  acceptance_rate_mean = 0.1598060342126327
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.9602597501126218
  speedup_vs_arm_ngram-g4_median = 1.0096131091104752
  speedup_vs_arm_ngram-g6_median = 0.9974680438211598
  speedup_vs_arm_qwen-g2_median = 0.9156996304266741
  speedup_vs_arm_qwen-g4_median = 0.9707462429963097
  speedup_vs_arm_qwen-g6_median = 1.2834926123168566
  speedup_vs_bandit_thompson_median = 1.0067415418596295
  speedup_vs_bandit_ucb1_median = 1.0346744667169339
  speedup_vs_hf_assisted_median = 3.860839361858903
  speedup_vs_hf_sample_compiled_median = 1.0981752793930746
  forward_counts.emitted_tokens_per_step = 1.1459265890778871
  forward_counts.n_generations = 10
  forward_counts.n_steps = 1117
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1127
  forward_counts.target_forwards_per_emitted_token = 0.88046875
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.019217263259669163
  step_profile.bookkeeping_ms_mean = 0.40929117295874373
  step_profile.bookkeeping_ms_total = 457.1782401949167
  step_profile.draft_fraction = 0.004558218213436309
  step_profile.draft_ms_mean = 0.09708138219111714
  step_profile.draft_ms_total = 108.43990390747786
  step_profile.n_steps = 1117.0
  step_profile.sample_fraction = 0.031719698447734083
  step_profile.sample_ms_mean = 0.6755692737382122
  step_profile.sample_ms_total = 754.610878765583
  step_profile.step_ms_mean = 21.298098872263143
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 1.1459265890778871
  step_profile.verify_fraction = 0.9445048200791605
  step_profile.verify_ms_mean = 20.11615704337507
  step_profile.verify_ms_total = 22469.74741744995
  host_syncs.by_kind = {'cuda': 125}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 123
  host_syncs.per_step = 1.016260162601626
  host_syncs.prompt_id = 81
  host_syncs.total = 125
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[arm_ngram-g4]
  workload_tokens_per_second = 54.128569868604735
  tokens_per_second_median = 52.23758728221364
  tokens_per_second_p25 = 51.28229917044993
  tokens_per_second_p75 = 56.80166128668298
  tokens_per_second_mean = 54.879535128339946
  acceptance_rate_mean = 0.10118097533820895
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.98770473102118
  speedup_vs_arm_ngram-g2_median = 0.990481706736626
  speedup_vs_arm_ngram-g6_median = 1.0109378273767324
  speedup_vs_arm_qwen-g2_median = 0.9068710192170903
  speedup_vs_arm_qwen-g4_median = 0.9407340677546027
  speedup_vs_arm_qwen-g6_median = 1.3248358220438117
  speedup_vs_bandit_thompson_median = 1.0188577509271037
  speedup_vs_bandit_ucb1_median = 1.0356076915762633
  speedup_vs_hf_assisted_median = 3.892514032790659
  speedup_vs_hf_sample_compiled_median = 1.100999157660091
  forward_counts.emitted_tokens_per_step = 1.171088746569076
  forward_counts.n_generations = 10
  forward_counts.n_steps = 1093
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1103
  forward_counts.target_forwards_per_emitted_token = 0.86171875
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.019563243809902226
  step_profile.bookkeeping_ms_mean = 0.42227864607636906
  step_profile.bookkeeping_ms_total = 461.55056016147137
  step_profile.draft_fraction = 0.004431874214307862
  step_profile.draft_ms_mean = 0.09566337060377561
  step_profile.draft_ms_total = 104.56006406992674
  step_profile.n_steps = 1093.0
  step_profile.sample_fraction = 0.03196169014787225
  step_profile.sample_ms_mean = 0.6899029308792003
  step_profile.sample_ms_total = 754.0639034509659
  step_profile.step_ms_mean = 21.585308151331553
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 1.171088746569076
  step_profile.verify_fraction = 0.9440431918279176
  step_profile.verify_ms_mean = 20.37746320377221
  step_profile.verify_ms_total = 22272.567281723022
  host_syncs.by_kind = {'cuda': 125}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 123
  host_syncs.per_step = 1.016260162601626
  host_syncs.prompt_id = 81
  host_syncs.total = 125
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[arm_ngram-g6]
  workload_tokens_per_second = 54.81427827007488
  tokens_per_second_median = 52.335603119516094
  tokens_per_second_p25 = 50.59246303567149
  tokens_per_second_p75 = 62.02253119979875
  tokens_per_second_mean = 55.80791282532179
  acceptance_rate_mean = 0.08179736992779216
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 1.0021296190033426
  speedup_vs_arm_ngram-g2_median = 1.0028484627660201
  speedup_vs_arm_ngram-g4_median = 0.9891805735126563
  speedup_vs_arm_qwen-g2_median = 0.9111948858564465
  speedup_vs_arm_qwen-g4_median = 0.964701096392768
  speedup_vs_arm_qwen-g6_median = 1.3099356042783339
  speedup_vs_bandit_thompson_median = 1.0277416124733347
  speedup_vs_bandit_ucb1_median = 1.040225388833815
  speedup_vs_hf_assisted_median = 3.842852847830338
  speedup_vs_hf_sample_compiled_median = 1.103113922761208
  forward_counts.emitted_tokens_per_step = 1.2018779342723005
  forward_counts.n_generations = 10
  forward_counts.n_steps = 1065
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 1075
  forward_counts.target_forwards_per_emitted_token = 0.83984375
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.019402473576556594
  step_profile.bookkeeping_ms_mean = 0.4244326760791277
  step_profile.bookkeeping_ms_total = 452.020800024271
  step_profile.draft_fraction = 0.004386604687253762
  step_profile.draft_ms_mean = 0.09595778388596476
  step_profile.draft_ms_total = 102.19503983855247
  step_profile.n_steps = 1065.0
  step_profile.sample_fraction = 0.03225388422902899
  step_profile.sample_ms_mean = 0.7055596464676477
  step_profile.sample_ms_total = 751.4210234880447
  step_profile.step_ms_mean = 21.875183821508024
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 1.2018779342723005
  step_profile.verify_fraction = 0.9439570375071606
  step_profile.verify_ms_mean = 20.649233715075283
  step_profile.verify_ms_total = 21991.433906555176
  host_syncs.by_kind = {'cuda': 125}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 123
  host_syncs.per_step = 1.016260162601626
  host_syncs.prompt_id = 81
  host_syncs.total = 125
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[arm_qwen-g2]
  workload_tokens_per_second = 58.20314894010966
  tokens_per_second_median = 57.44825430340413
  tokens_per_second_p25 = 54.62614558957723
  tokens_per_second_p75 = 60.72331644213134
  tokens_per_second_mean = 58.689568936932304
  acceptance_rate_mean = 0.7054198887928635
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 1.0609568019336848
  speedup_vs_arm_ngram-g2_median = 1.0922576519558331
  speedup_vs_arm_ngram-g4_median = 1.1028434293454714
  speedup_vs_arm_ngram-g6_median = 1.0974798474130472
  speedup_vs_arm_qwen-g4_median = 1.066615692838658
  speedup_vs_arm_qwen-g6_median = 1.1869506298674581
  speedup_vs_bandit_thompson_median = 1.0989723274548144
  speedup_vs_bandit_ucb1_median = 1.1006020177325255
  speedup_vs_hf_assisted_median = 4.271980867353443
  speedup_vs_hf_sample_compiled_median = 1.2225465856741882
  forward_counts.emitted_tokens_per_step = 2.397003745318352
  forward_counts.n_generations = 10
  forward_counts.n_steps = 534
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  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 = 0.00874119603868696
  step_profile.bookkeeping_ms_mean = 0.35956889907184164
  step_profile.bookkeeping_ms_total = 192.00979210436344
  step_profile.draft_fraction = 0.5089586329005966
  step_profile.draft_ms_mean = 20.936001720499902
  step_profile.draft_ms_total = 11179.824918746948
  step_profile.n_steps = 534.0
  step_profile.sample_fraction = 0.009665205498204625
  step_profile.sample_ms_mean = 0.39757800704977486
  step_profile.sample_ms_total = 212.30665576457977
  step_profile.step_ms_mean = 41.13497712217578
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.397003745318352
  step_profile.verify_fraction = 0.47263496556251183
  step_profile.verify_ms_mean = 19.44182849555426
  step_profile.verify_ms_total = 10381.936416625977
  host_syncs.by_kind = {'cuda': 62}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 59
  host_syncs.per_step = 1.0508474576271187
  host_syncs.prompt_id = 81
  host_syncs.total = 62
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[arm_qwen-g4]
  workload_tokens_per_second = 54.74153915549264
  tokens_per_second_median = 54.79641307228064
  tokens_per_second_p25 = 49.709063644625886
  tokens_per_second_p75 = 59.12271390908381
  tokens_per_second_mean = 55.62746062636766
  acceptance_rate_mean = 0.5580407428308467
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.993558024060297
  speedup_vs_arm_ngram-g2_median = 1.0301381136380612
  speedup_vs_arm_ngram-g4_median = 1.0635265758404677
  speedup_vs_arm_ngram-g6_median = 1.0366992583698516
  speedup_vs_arm_qwen-g2_median = 0.937546842080305
  speedup_vs_arm_qwen-g6_median = 1.0807413622571942
  speedup_vs_bandit_thompson_median = 1.029585267546571
  speedup_vs_bandit_ucb1_median = 1.0186998501202567
  speedup_vs_hf_assisted_median = 4.022333154547398
  speedup_vs_hf_sample_compiled_median = 1.1587817380687522
  forward_counts.emitted_tokens_per_step = 3.2989690721649483
  forward_counts.n_generations = 10
  forward_counts.n_steps = 388
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 398
  forward_counts.target_forwards_per_emitted_token = 0.3109375
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.008151932682695127
  step_profile.bookkeeping_ms_mean = 0.4908812374798293
  step_profile.bookkeeping_ms_total = 190.46192014217377
  step_profile.draft_fraction = 0.6568316188905527
  step_profile.draft_ms_mean = 39.55213204609979
  step_profile.draft_ms_total = 15346.227233886719
  step_profile.n_steps = 388.0
  step_profile.sample_fraction = 0.006828806955984182
  step_profile.sample_ms_mean = 0.4112071749783054
  step_profile.sample_ms_total = 159.5483838915825
  step_profile.step_ms_mean = 60.216546994048294
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 3.2989690721649483
  step_profile.verify_fraction = 0.32818764147076795
  step_profile.verify_ms_mean = 19.76232653549037
  step_profile.verify_ms_total = 7667.782695770264
  host_syncs.by_kind = {'cuda': 49}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 46
  host_syncs.per_step = 1.065217391304348
  host_syncs.prompt_id = 81
  host_syncs.total = 49
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[arm_qwen-g6]
  workload_tokens_per_second = 44.16385268953409
  tokens_per_second_median = 47.23618860735462
  tokens_per_second_p25 = 34.65187145627606
  tokens_per_second_p75 = 56.71937602401735
  tokens_per_second_mean = 48.00689984555355
  acceptance_rate_mean = 0.4268935986441579
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.8092205491251663
  speedup_vs_arm_ngram-g2_median = 0.7846064177219301
  speedup_vs_arm_ngram-g4_median = 0.7561238665639018
  speedup_vs_arm_ngram-g6_median = 0.764683496004614
  speedup_vs_arm_qwen-g2_median = 0.8430336205728963
  speedup_vs_arm_qwen-g4_median = 0.9253151853459013
  speedup_vs_bandit_thompson_median = 0.8810714525327794
  speedup_vs_bandit_ucb1_median = 0.9125335349468771
  speedup_vs_hf_assisted_median = 3.4313818804394343
  speedup_vs_hf_sample_compiled_median = 1.0415194341765017
  forward_counts.emitted_tokens_per_step = 3.506849315068493
  forward_counts.n_generations = 10
  forward_counts.n_steps = 365
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 375
  forward_counts.target_forwards_per_emitted_token = 0.29296875
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.007499706666689244
  step_profile.bookkeeping_ms_mean = 0.5951506420765837
  step_profile.bookkeeping_ms_total = 217.22998435795307
  step_profile.draft_fraction = 0.7354799168624756
  step_profile.draft_ms_mean = 58.365128692208906
  step_profile.draft_ms_total = 21303.27197265625
  step_profile.n_steps = 365.0
  step_profile.sample_fraction = 0.005220808742972087
  step_profile.sample_ms_mean = 0.41430522734171726
  step_profile.sample_ms_total = 151.2214079797268
  step_profile.step_ms_mean = 79.35652266508096
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 3.506849315068493
  step_profile.verify_fraction = 0.25179956772786305
  step_profile.verify_ms_mean = 19.981938103453754
  step_profile.verify_ms_total = 7293.40740776062
  host_syncs.by_kind = {'cuda': 53}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 50
  host_syncs.per_step = 1.06
  host_syncs.prompt_id = 81
  host_syncs.total = 53
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[bandit_thompson]
  workload_tokens_per_second = 52.642701845919284
  tokens_per_second_median = 51.9403931241941
  tokens_per_second_p25 = 50.41533721469594
  tokens_per_second_p75 = 59.48059173696604
  tokens_per_second_mean = 53.735871758667976
  acceptance_rate_mean = 0.3417862799147732
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.9645729802297842
  speedup_vs_arm_ngram-g2_median = 0.9938659905206648
  speedup_vs_arm_ngram-g4_median = 0.981501924210229
  speedup_vs_arm_ngram-g6_median = 0.9733087604663124
  speedup_vs_arm_qwen-g2_median = 0.9099465626844444
  speedup_vs_arm_qwen-g4_median = 0.9713023929637056
  speedup_vs_arm_qwen-g6_median = 1.1358393650768144
  speedup_vs_bandit_ucb1_median = 0.9734864408732035
  speedup_vs_hf_assisted_median = 3.75421268901648
  speedup_vs_hf_sample_compiled_median = 1.1422253370879312
  forward_counts.emitted_tokens_per_step = 1.951219512195122
  forward_counts.n_generations = 10
  forward_counts.n_steps = 656
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 666
  forward_counts.target_forwards_per_emitted_token = 0.5203125
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.012361191140928887
  step_profile.bookkeeping_ms_mean = 0.45736273121452187
  step_profile.bookkeeping_ms_total = 300.02995167672634
  step_profile.draft_fraction = 0.41588482049586295
  step_profile.draft_ms_mean = 15.387693241215878
  step_profile.draft_ms_total = 10094.326766237617
  step_profile.n_steps = 656.0
  step_profile.sample_fraction = 0.018964730618785657
  step_profile.sample_ms_mean = 0.701692975512365
  step_profile.sample_ms_total = 460.31059193611145
  step_profile.step_ms_mean = 36.99989151532149
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 1.951219512195122
  step_profile.verify_fraction = 0.5527892577444224
  step_profile.verify_ms_mean = 20.45314256737872
  step_profile.verify_ms_total = 13417.26152420044
  host_syncs.by_kind = {'cuda': 66}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 62
  host_syncs.per_step = 1.064516129032258
  host_syncs.prompt_id = 81
  host_syncs.total = 66
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[bandit_ucb1]
  workload_tokens_per_second = 52.29889895462427
  tokens_per_second_median = 51.543202963385326
  tokens_per_second_p25 = 47.53668477643265
  tokens_per_second_p75 = 54.527610295946374
  tokens_per_second_mean = 53.62246756279585
  acceptance_rate_mean = 0.4160261452947848
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.9606127930574964
  speedup_vs_arm_ngram-g2_median = 0.9665765923539567
  speedup_vs_arm_ngram-g4_median = 0.9662758410094434
  speedup_vs_arm_ngram-g6_median = 0.9614314165782372
  speedup_vs_arm_qwen-g2_median = 0.9087011775143803
  speedup_vs_arm_qwen-g4_median = 0.9816435458046795
  speedup_vs_arm_qwen-g6_median = 1.0969137028684113
  speedup_vs_bandit_thompson_median = 1.027423310995641
  speedup_vs_hf_assisted_median = 3.9096484230267468
  speedup_vs_hf_sample_compiled_median = 1.0967616466857146
  forward_counts.emitted_tokens_per_step = 2.1880341880341883
  forward_counts.n_generations = 10
  forward_counts.n_steps = 585
  forward_counts.source = derived: a hook would force the compiled forward to recompile
  forward_counts.target_forwards = 595
  forward_counts.target_forwards_per_emitted_token = 0.46484375
  forward_counts.target_forwards_per_step = 1.0
  forward_counts.tokens_emitted = 1280
  step_profile.bookkeeping_fraction = 0.010929318094063308
  step_profile.bookkeeping_ms_mean = 0.45649958247812383
  step_profile.bookkeeping_ms_total = 267.05225574970245
  step_profile.draft_fraction = 0.48535892990724316
  step_profile.draft_ms_mean = 20.272641618422462
  step_profile.draft_ms_total = 11859.495346777141
  step_profile.n_steps = 585.0
  step_profile.sample_fraction = 0.01674692161731664
  step_profile.sample_ms_mean = 0.6994912821003515
  step_profile.sample_ms_total = 409.2024000287056
  step_profile.step_ms_mean = 41.768349914354644
  step_profile.tokens_emitted = 1280.0
  step_profile.tokens_per_step_mean = 2.1880341880341883
  step_profile.verify_fraction = 0.4869648303813769
  step_profile.verify_ms_mean = 20.33971743135371
  step_profile.verify_ms_total = 11898.734697341919
  host_syncs.by_kind = {'cuda': 65}
  host_syncs.counter = cuda_sync_debug_mode
  host_syncs.n_steps = 61
  host_syncs.per_step = 1.0655737704918034
  host_syncs.prompt_id = 81
  host_syncs.total = 65
  nonfinite_logits.prompt_id = 81
  nonfinite_logits.scope = not checked: a forward hook would force the compiled forward to recompile
[hf_assisted]
  workload_tokens_per_second = 10.890847798918573
  tokens_per_second_median = 13.667517193258476
  tokens_per_second_p25 = 8.531354367274075
  tokens_per_second_p75 = 18.407838971538663
  tokens_per_second_mean = 13.770204048248127
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.24719704213486934
  speedup_vs_arm_ngram-g2_median = 0.2592698086669323
  speedup_vs_arm_ngram-g4_median = 0.25700305320489936
  speedup_vs_arm_ngram-g6_median = 0.2602630014649547
  speedup_vs_arm_qwen-g2_median = 0.2352437510779194
  speedup_vs_arm_qwen-g4_median = 0.24993764893014142
  speedup_vs_arm_qwen-g6_median = 0.2914278545072734
  speedup_vs_bandit_thompson_median = 0.2668185389997238
  speedup_vs_bandit_ucb1_median = 0.2566135580638791
  speedup_vs_hf_sample_compiled_median = 0.28338512914525105
  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_sample_compiled]
  workload_tokens_per_second = 47.039167970961024
  tokens_per_second_median = 47.53665677233218
  tokens_per_second_p25 = 46.96950533316543
  tokens_per_second_p75 = 47.98696543269919
  tokens_per_second_mean = 47.11808633366134
  gpu_memory_gb = 4.030667304992676
  speedup_vs_arm_composite-g4_median = 0.8445731073182788
  speedup_vs_arm_ngram-g2_median = 0.9110187522198159
  speedup_vs_arm_ngram-g4_median = 0.9082724845968474
  speedup_vs_arm_ngram-g6_median = 0.906645193090513
  speedup_vs_arm_qwen-g2_median = 0.8179693265332029
  speedup_vs_arm_qwen-g4_median = 0.8630125592555578
  speedup_vs_arm_qwen-g6_median = 0.9602540653915208
  speedup_vs_bandit_thompson_median = 0.8776016155298161
  speedup_vs_bandit_ucb1_median = 0.9117786978910798
  speedup_vs_hf_assisted_median = 3.540946090548248
  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

