{ "calibration": { "boltzmann_gated_mobs_v11": { "plan": { "calibration_work_reduction_fraction": 0.5, "full_grid_prompt_config_evaluations": 72, "stage1_prompts": 1, "successive_halving_prompt_config_evaluations": 36, "survivors": 4, "total_configs": 24, "training_steps": 0 }, "selected": { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.16, "mobs_budget": 3, "pair_weight": 1.0, "temperature": 0.04, "top_k": 8 }, "config_key": "T=0.04|gate=0.16|budget=3|k=8|pair=1|early=0.15", "mean_acceptance_rate": 0.0872762714867978, "mean_gated_positions": 1.3333333333333333, "mean_guidance_scores": 38.333333333333336, "mean_pair_scores": 18.666666666666668, "mean_target_forward_passes": 6.333333333333333, "tokens_per_second": 6.549208943783108 }, "stage1": [ { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.08, "mobs_budget": 1, "pair_weight": 1.0, "temperature": 0.04, "top_k": 4 }, "config_key": "T=0.04|gate=0.08|budget=1|k=4|pair=1|early=0.15", "mean_acceptance_rate": 0.045454545454545456, "mean_gated_positions": 2.0, "mean_guidance_scores": 34.0, "mean_pair_scores": 12.0, "mean_target_forward_passes": 7.0, "tokens_per_second": 5.9301536642543855 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.08, "mobs_budget": 1, "pair_weight": 1.0, "temperature": 0.04, "top_k": 8 }, "config_key": "T=0.04|gate=0.08|budget=1|k=8|pair=1|early=0.15", "mean_acceptance_rate": 0.045454545454545456, "mean_gated_positions": 2.0, "mean_guidance_scores": 46.0, "mean_pair_scores": 24.0, "mean_target_forward_passes": 7.0, "tokens_per_second": 5.898194902866635 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.08, "mobs_budget": 2, "pair_weight": 1.0, "temperature": 0.04, "top_k": 4 }, "config_key": "T=0.04|gate=0.08|budget=2|k=4|pair=1|early=0.15", "mean_acceptance_rate": 0.045454545454545456, "mean_gated_positions": 2.0, "mean_guidance_scores": 34.0, "mean_pair_scores": 12.0, "mean_target_forward_passes": 7.0, "tokens_per_second": 5.817033717640731 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.08, "mobs_budget": 2, "pair_weight": 1.0, "temperature": 0.04, "top_k": 8 }, "config_key": "T=0.04|gate=0.08|budget=2|k=8|pair=1|early=0.15", "mean_acceptance_rate": 0.045454545454545456, "mean_gated_positions": 2.0, "mean_guidance_scores": 46.0, "mean_pair_scores": 24.0, "mean_target_forward_passes": 7.0, "tokens_per_second": 5.816098170345032 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.08, "mobs_budget": 3, "pair_weight": 1.0, "temperature": 0.04, "top_k": 4 }, "config_key": "T=0.04|gate=0.08|budget=3|k=4|pair=1|early=0.15", "mean_acceptance_rate": 0.045454545454545456, "mean_gated_positions": 2.0, "mean_guidance_scores": 34.0, "mean_pair_scores": 12.0, "mean_target_forward_passes": 7.0, "tokens_per_second": 5.784051685620768 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.08, "mobs_budget": 3, "pair_weight": 1.0, "temperature": 0.04, "top_k": 8 }, "config_key": "T=0.04|gate=0.08|budget=3|k=8|pair=1|early=0.15", "mean_acceptance_rate": 0.045454545454545456, "mean_gated_positions": 2.0, "mean_guidance_scores": 46.0, "mean_pair_scores": 24.0, "mean_target_forward_passes": 7.0, "tokens_per_second": 5.905684712149763 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.16, "mobs_budget": 1, "pair_weight": 1.0, "temperature": 0.04, "top_k": 4 }, "config_key": "T=0.04|gate=0.16|budget=1|k=4|pair=1|early=0.15", "mean_acceptance_rate": 0.045454545454545456, "mean_gated_positions": 1.0, "mean_guidance_scores": 30.0, "mean_pair_scores": 8.0, "mean_target_forward_passes": 7.0, "tokens_per_second": 5.898000849167787 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.16, "mobs_budget": 1, "pair_weight": 1.0, "temperature": 0.04, "top_k": 8 }, "config_key": "T=0.04|gate=0.16|budget=1|k=8|pair=1|early=0.15", "mean_acceptance_rate": 0.045454545454545456, "mean_gated_positions": 1.0, "mean_guidance_scores": 38.0, "mean_pair_scores": 16.0, "mean_target_forward_passes": 7.0, "tokens_per_second": 5.615753125911015 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.16, "mobs_budget": 2, "pair_weight": 1.0, "temperature": 0.04, "top_k": 4 }, "config_key": "T=0.04|gate=0.16|budget=2|k=4|pair=1|early=0.15", "mean_acceptance_rate": 0.045454545454545456, "mean_gated_positions": 1.0, "mean_guidance_scores": 30.0, "mean_pair_scores": 8.0, "mean_target_forward_passes": 7.0, "tokens_per_second": 5.917159286264204 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.16, "mobs_budget": 2, "pair_weight": 1.0, "temperature": 0.04, "top_k": 8 }, "config_key": "T=0.04|gate=0.16|budget=2|k=8|pair=1|early=0.15", "mean_acceptance_rate": 0.045454545454545456, "mean_gated_positions": 1.0, "mean_guidance_scores": 38.0, "mean_pair_scores": 16.0, "mean_target_forward_passes": 7.0, "tokens_per_second": 5.884649533829743 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.16, "mobs_budget": 3, "pair_weight": 1.0, "temperature": 0.04, "top_k": 4 }, "config_key": "T=0.04|gate=0.16|budget=3|k=4|pair=1|early=0.15", "mean_acceptance_rate": 0.045454545454545456, "mean_gated_positions": 1.0, "mean_guidance_scores": 30.0, "mean_pair_scores": 8.0, "mean_target_forward_passes": 7.0, "tokens_per_second": 5.944698318854493 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.16, "mobs_budget": 3, "pair_weight": 1.0, "temperature": 0.04, "top_k": 8 }, "config_key": "T=0.04|gate=0.16|budget=3|k=8|pair=1|early=0.15", "mean_acceptance_rate": 0.045454545454545456, "mean_gated_positions": 1.0, "mean_guidance_scores": 38.0, "mean_pair_scores": 16.0, "mean_target_forward_passes": 7.0, "tokens_per_second": 5.964950305427247 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.08, "mobs_budget": 1, "pair_weight": 1.0, "temperature": 0.08, "top_k": 4 }, "config_key": "T=0.08|gate=0.08|budget=1|k=4|pair=1|early=0.15", "mean_acceptance_rate": 0.045454545454545456, "mean_gated_positions": 2.0, "mean_guidance_scores": 34.0, "mean_pair_scores": 12.0, "mean_target_forward_passes": 7.0, "tokens_per_second": 5.8742121124891735 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.08, "mobs_budget": 1, "pair_weight": 1.0, "temperature": 0.08, "top_k": 8 }, "config_key": "T=0.08|gate=0.08|budget=1|k=8|pair=1|early=0.15", "mean_acceptance_rate": 0.045454545454545456, "mean_gated_positions": 2.0, "mean_guidance_scores": 46.0, "mean_pair_scores": 24.0, "mean_target_forward_passes": 7.0, "tokens_per_second": 5.7325823194747025 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.08, "mobs_budget": 2, "pair_weight": 1.0, "temperature": 0.08, "top_k": 4 }, "config_key": "T=0.08|gate=0.08|budget=2|k=4|pair=1|early=0.15", "mean_acceptance_rate": 0.045454545454545456, "mean_gated_positions": 3.0, "mean_guidance_scores": 42.0, "mean_pair_scores": 20.0, "mean_target_forward_passes": 7.0, "tokens_per_second": 5.870593361366573 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.08, "mobs_budget": 2, "pair_weight": 1.0, "temperature": 0.08, "top_k": 8 }, "config_key": "T=0.08|gate=0.08|budget=2|k=8|pair=1|early=0.15", "mean_acceptance_rate": 0.045454545454545456, "mean_gated_positions": 3.0, "mean_guidance_scores": 62.0, "mean_pair_scores": 40.0, "mean_target_forward_passes": 7.0, "tokens_per_second": 5.97721132891495 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.08, "mobs_budget": 3, "pair_weight": 1.0, "temperature": 0.08, "top_k": 4 }, "config_key": "T=0.08|gate=0.08|budget=3|k=4|pair=1|early=0.15", "mean_acceptance_rate": 0.045454545454545456, "mean_gated_positions": 3.0, "mean_guidance_scores": 42.0, "mean_pair_scores": 20.0, "mean_target_forward_passes": 7.0, "tokens_per_second": 5.907505893708322 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.08, "mobs_budget": 3, "pair_weight": 1.0, "temperature": 0.08, "top_k": 8 }, "config_key": "T=0.08|gate=0.08|budget=3|k=8|pair=1|early=0.15", "mean_acceptance_rate": 0.045454545454545456, "mean_gated_positions": 3.0, "mean_guidance_scores": 62.0, "mean_pair_scores": 40.0, "mean_target_forward_passes": 7.0, "tokens_per_second": 5.850268156365895 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.16, "mobs_budget": 1, "pair_weight": 1.0, "temperature": 0.08, "top_k": 4 }, "config_key": "T=0.08|gate=0.16|budget=1|k=4|pair=1|early=0.15", "mean_acceptance_rate": 0.045454545454545456, "mean_gated_positions": 2.0, "mean_guidance_scores": 34.0, "mean_pair_scores": 12.0, "mean_target_forward_passes": 7.0, "tokens_per_second": 5.843571575696665 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.16, "mobs_budget": 1, "pair_weight": 1.0, "temperature": 0.08, "top_k": 8 }, "config_key": "T=0.08|gate=0.16|budget=1|k=8|pair=1|early=0.15", "mean_acceptance_rate": 0.045454545454545456, "mean_gated_positions": 2.0, "mean_guidance_scores": 46.0, "mean_pair_scores": 24.0, "mean_target_forward_passes": 7.0, "tokens_per_second": 5.866376795690763 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.16, "mobs_budget": 2, "pair_weight": 1.0, "temperature": 0.08, "top_k": 4 }, "config_key": "T=0.08|gate=0.16|budget=2|k=4|pair=1|early=0.15", "mean_acceptance_rate": 0.045454545454545456, "mean_gated_positions": 2.0, "mean_guidance_scores": 34.0, "mean_pair_scores": 12.0, "mean_target_forward_passes": 7.0, "tokens_per_second": 5.836324863840083 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.16, "mobs_budget": 2, "pair_weight": 1.0, "temperature": 0.08, "top_k": 8 }, "config_key": "T=0.08|gate=0.16|budget=2|k=8|pair=1|early=0.15", "mean_acceptance_rate": 0.045454545454545456, "mean_gated_positions": 2.0, "mean_guidance_scores": 46.0, "mean_pair_scores": 24.0, "mean_target_forward_passes": 7.0, "tokens_per_second": 5.939855300194306 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.16, "mobs_budget": 3, "pair_weight": 1.0, "temperature": 0.08, "top_k": 4 }, "config_key": "T=0.08|gate=0.16|budget=3|k=4|pair=1|early=0.15", "mean_acceptance_rate": 0.045454545454545456, "mean_gated_positions": 2.0, "mean_guidance_scores": 34.0, "mean_pair_scores": 12.0, "mean_target_forward_passes": 7.0, "tokens_per_second": 5.974721007694674 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.16, "mobs_budget": 3, "pair_weight": 1.0, "temperature": 0.08, "top_k": 8 }, "config_key": "T=0.08|gate=0.16|budget=3|k=8|pair=1|early=0.15", "mean_acceptance_rate": 0.045454545454545456, "mean_gated_positions": 2.0, "mean_guidance_scores": 46.0, "mean_pair_scores": 24.0, "mean_target_forward_passes": 7.0, "tokens_per_second": 5.960267272926077 } ], "stage2": [ { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.16, "mobs_budget": 3, "pair_weight": 1.0, "temperature": 0.04, "top_k": 8 }, "config_key": "T=0.04|gate=0.16|budget=3|k=8|pair=1|early=0.15", "mean_acceptance_rate": 0.0872762714867978, "mean_gated_positions": 1.3333333333333333, "mean_guidance_scores": 38.333333333333336, "mean_pair_scores": 18.666666666666668, "mean_target_forward_passes": 6.333333333333333, "tokens_per_second": 6.549208943783108 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.08, "mobs_budget": 2, "pair_weight": 1.0, "temperature": 0.08, "top_k": 8 }, "config_key": "T=0.08|gate=0.08|budget=2|k=8|pair=1|early=0.15", "mean_acceptance_rate": 0.0872762714867978, "mean_gated_positions": 5.666666666666667, "mean_guidance_scores": 97.0, "mean_pair_scores": 77.33333333333333, "mean_target_forward_passes": 6.333333333333333, "tokens_per_second": 6.345626880128733 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.16, "mobs_budget": 3, "pair_weight": 1.0, "temperature": 0.08, "top_k": 4 }, "config_key": "T=0.08|gate=0.16|budget=3|k=4|pair=1|early=0.15", "mean_acceptance_rate": 0.0872762714867978, "mean_gated_positions": 4.0, "mean_guidance_scores": 46.333333333333336, "mean_pair_scores": 26.666666666666668, "mean_target_forward_passes": 6.333333333333333, "tokens_per_second": 6.353458561024625 }, { "all_exact": true, "config": { "early_position_bias": 0.15, "gate_floor": 0.16, "mobs_budget": 3, "pair_weight": 1.0, "temperature": 0.08, "top_k": 8 }, "config_key": "T=0.08|gate=0.16|budget=3|k=8|pair=1|early=0.15", "mean_acceptance_rate": 0.0872762714867978, "mean_gated_positions": 4.0, "mean_guidance_scores": 73.0, "mean_pair_scores": 53.333333333333336, "mean_target_forward_passes": 6.333333333333333, "tokens_per_second": 6.414861393209996 } ] }, "boltzmann_v10": { "plan": { "calibration_work_reduction_fraction": 0.16666666666666663, "full_grid_prompt_config_evaluations": 72, "stage1_prompts": 2, "successive_halving_prompt_config_evaluations": 60, "survivors": 4, "total_configs": 24, "training_steps": 0 }, "selected": { "all_exact": true, "config": { "margin_cutoff": 0.35, "margin_slope": 1.5, "stochastic_budget": 2, "temperature": 0.16 }, "config_key": "T=0.16|cut=0.35|budget=2|slope=1.5", "mean_acceptance_rate": 0.0872762714867978, "mean_candidate_scores": 14.666666666666666, "mean_target_forward_passes": 6.333333333333333, "tokens_per_second": 6.5722068118653985 }, "stage1": [ { "all_exact": true, "config": { "margin_cutoff": 0.05, "margin_slope": 1.5, "stochastic_budget": 1, "temperature": 0.03 }, "config_key": "T=0.03|cut=0.05|budget=1|slope=1.5", "mean_acceptance_rate": 0.07535885167464115, "mean_candidate_scores": 2.0, "mean_target_forward_passes": 6.5, "tokens_per_second": 6.1647276093909085 }, { "all_exact": true, "config": { "margin_cutoff": 0.05, "margin_slope": 1.5, "stochastic_budget": 2, "temperature": 0.03 }, "config_key": "T=0.03|cut=0.05|budget=2|slope=1.5", "mean_acceptance_rate": 0.07535885167464115, "mean_candidate_scores": 2.0, "mean_target_forward_passes": 6.5, "tokens_per_second": 6.172082061690334 }, { "all_exact": true, "config": { "margin_cutoff": 0.15, "margin_slope": 1.5, "stochastic_budget": 1, "temperature": 0.03 }, "config_key": "T=0.03|cut=0.15|budget=1|slope=1.5", "mean_acceptance_rate": 0.07535885167464115, "mean_candidate_scores": 7.0, "mean_target_forward_passes": 6.5, "tokens_per_second": 6.205053272448845 }, { "all_exact": true, "config": { "margin_cutoff": 0.15, "margin_slope": 1.5, "stochastic_budget": 2, "temperature": 0.03 }, "config_key": "T=0.03|cut=0.15|budget=2|slope=1.5", "mean_acceptance_rate": 0.07535885167464115, "mean_candidate_scores": 8.0, "mean_target_forward_passes": 6.5, "tokens_per_second": 6.092381970324525 }, { "all_exact": true, "config": { "margin_cutoff": 0.35, "margin_slope": 1.5, "stochastic_budget": 1, "temperature": 0.03 }, "config_key": "T=0.03|cut=0.35|budget=1|slope=1.5", "mean_acceptance_rate": 0.07535885167464115, "mean_candidate_scores": 11.0, "mean_target_forward_passes": 6.5, "tokens_per_second": 6.038064372164027 }, { "all_exact": true, "config": { "margin_cutoff": 0.35, "margin_slope": 1.5, "stochastic_budget": 2, "temperature": 0.03 }, "config_key": "T=0.03|cut=0.35|budget=2|slope=1.5", "mean_acceptance_rate": 0.07535885167464115, "mean_candidate_scores": 14.0, "mean_target_forward_passes": 6.5, "tokens_per_second": 6.300111735985931 }, { "all_exact": true, "config": { "margin_cutoff": 0.05, "margin_slope": 1.5, "stochastic_budget": 1, "temperature": 0.06 }, "config_key": "T=0.06|cut=0.05|budget=1|slope=1.5", "mean_acceptance_rate": 0.07535885167464115, "mean_candidate_scores": 2.0, "mean_target_forward_passes": 6.5, "tokens_per_second": 6.158668926945642 }, { "all_exact": true, "config": { "margin_cutoff": 0.05, "margin_slope": 1.5, "stochastic_budget": 2, "temperature": 0.06 }, "config_key": "T=0.06|cut=0.05|budget=2|slope=1.5", "mean_acceptance_rate": 0.07535885167464115, "mean_candidate_scores": 2.0, "mean_target_forward_passes": 6.5, "tokens_per_second": 6.232103065137487 }, { "all_exact": true, "config": { "margin_cutoff": 0.15, "margin_slope": 1.5, "stochastic_budget": 1, "temperature": 0.06 }, "config_key": "T=0.06|cut=0.15|budget=1|slope=1.5", "mean_acceptance_rate": 0.07535885167464115, "mean_candidate_scores": 7.0, "mean_target_forward_passes": 6.5, "tokens_per_second": 6.205502248658734 }, { "all_exact": true, "config": { "margin_cutoff": 0.15, "margin_slope": 1.5, "stochastic_budget": 2, "temperature": 0.06 }, "config_key": "T=0.06|cut=0.15|budget=2|slope=1.5", "mean_acceptance_rate": 0.07535885167464115, "mean_candidate_scores": 8.0, "mean_target_forward_passes": 6.5, "tokens_per_second": 6.165223991729381 }, { "all_exact": true, "config": { "margin_cutoff": 0.35, "margin_slope": 1.5, "stochastic_budget": 1, "temperature": 0.06 }, "config_key": "T=0.06|cut=0.35|budget=1|slope=1.5", "mean_acceptance_rate": 0.07535885167464115, "mean_candidate_scores": 11.0, "mean_target_forward_passes": 6.5, "tokens_per_second": 6.015169993303459 }, { "all_exact": true, "config": { "margin_cutoff": 0.35, "margin_slope": 1.5, "stochastic_budget": 2, "temperature": 0.06 }, "config_key": "T=0.06|cut=0.35|budget=2|slope=1.5", "mean_acceptance_rate": 0.07535885167464115, "mean_candidate_scores": 14.0, "mean_target_forward_passes": 6.5, "tokens_per_second": 6.058920822270032 }, { "all_exact": true, "config": { "margin_cutoff": 0.05, "margin_slope": 1.5, "stochastic_budget": 1, "temperature": 0.1 }, "config_key": "T=0.1|cut=0.05|budget=1|slope=1.5", "mean_acceptance_rate": 0.07535885167464115, "mean_candidate_scores": 2.0, "mean_target_forward_passes": 6.5, "tokens_per_second": 6.046594982329176 }, { "all_exact": true, "config": { "margin_cutoff": 0.05, "margin_slope": 1.5, "stochastic_budget": 2, "temperature": 0.1 }, "config_key": "T=0.1|cut=0.05|budget=2|slope=1.5", "mean_acceptance_rate": 0.07535885167464115, "mean_candidate_scores": 2.0, "mean_target_forward_passes": 6.5, "tokens_per_second": 6.081020435485939 }, { "all_exact": true, "config": { "margin_cutoff": 0.15, "margin_slope": 1.5, "stochastic_budget": 1, "temperature": 0.1 }, "config_key": "T=0.1|cut=0.15|budget=1|slope=1.5", "mean_acceptance_rate": 0.07535885167464115, "mean_candidate_scores": 7.0, "mean_target_forward_passes": 6.5, "tokens_per_second": 6.088803547389854 }, { "all_exact": true, "config": { "margin_cutoff": 0.15, "margin_slope": 1.5, "stochastic_budget": 2, "temperature": 0.1 }, "config_key": "T=0.1|cut=0.15|budget=2|slope=1.5", "mean_acceptance_rate": 0.07535885167464115, "mean_candidate_scores": 8.0, "mean_target_forward_passes": 6.5, "tokens_per_second": 6.162694026100379 }, { "all_exact": true, "config": { "margin_cutoff": 0.35, "margin_slope": 1.5, "stochastic_budget": 1, "temperature": 0.1 }, "config_key": "T=0.1|cut=0.35|budget=1|slope=1.5", "mean_acceptance_rate": 0.07535885167464115, "mean_candidate_scores": 11.0, "mean_target_forward_passes": 6.5, "tokens_per_second": 6.14600398041445 }, { "all_exact": true, "config": { "margin_cutoff": 0.35, "margin_slope": 1.5, "stochastic_budget": 2, "temperature": 0.1 }, "config_key": "T=0.1|cut=0.35|budget=2|slope=1.5", "mean_acceptance_rate": 0.07535885167464115, "mean_candidate_scores": 14.0, "mean_target_forward_passes": 6.5, "tokens_per_second": 6.203360357435485 }, { "all_exact": true, "config": { "margin_cutoff": 0.05, "margin_slope": 1.5, "stochastic_budget": 1, "temperature": 0.16 }, "config_key": "T=0.16|cut=0.05|budget=1|slope=1.5", "mean_acceptance_rate": 0.07535885167464115, "mean_candidate_scores": 2.0, "mean_target_forward_passes": 6.5, "tokens_per_second": 6.225112438175082 }, { "all_exact": true, "config": { "margin_cutoff": 0.05, "margin_slope": 1.5, "stochastic_budget": 2, "temperature": 0.16 }, "config_key": "T=0.16|cut=0.05|budget=2|slope=1.5", "mean_acceptance_rate": 0.07535885167464115, "mean_candidate_scores": 2.0, "mean_target_forward_passes": 6.5, "tokens_per_second": 6.212595192304136 }, { "all_exact": true, "config": { "margin_cutoff": 0.15, "margin_slope": 1.5, "stochastic_budget": 1, "temperature": 0.16 }, "config_key": "T=0.16|cut=0.15|budget=1|slope=1.5", "mean_acceptance_rate": 0.07535885167464115, "mean_candidate_scores": 7.0, "mean_target_forward_passes": 6.5, "tokens_per_second": 6.266297405875774 }, { "all_exact": true, "config": { "margin_cutoff": 0.15, "margin_slope": 1.5, "stochastic_budget": 2, "temperature": 0.16 }, "config_key": "T=0.16|cut=0.15|budget=2|slope=1.5", "mean_acceptance_rate": 0.07535885167464115, "mean_candidate_scores": 8.0, "mean_target_forward_passes": 6.5, "tokens_per_second": 6.221350005195755 }, { "all_exact": true, "config": { "margin_cutoff": 0.35, "margin_slope": 1.5, "stochastic_budget": 1, "temperature": 0.16 }, "config_key": "T=0.16|cut=0.35|budget=1|slope=1.5", "mean_acceptance_rate": 0.07535885167464115, "mean_candidate_scores": 11.0, "mean_target_forward_passes": 6.5, "tokens_per_second": 6.172580765017732 }, { "all_exact": true, "config": { "margin_cutoff": 0.35, "margin_slope": 1.5, "stochastic_budget": 2, "temperature": 0.16 }, "config_key": "T=0.16|cut=0.35|budget=2|slope=1.5", "mean_acceptance_rate": 0.07535885167464115, "mean_candidate_scores": 14.0, "mean_target_forward_passes": 6.5, "tokens_per_second": 6.284191374551125 } ], "stage2": [ { "all_exact": true, "config": { "margin_cutoff": 0.35, "margin_slope": 1.5, "stochastic_budget": 2, "temperature": 0.03 }, "config_key": "T=0.03|cut=0.35|budget=2|slope=1.5", "mean_acceptance_rate": 0.0872762714867978, "mean_candidate_scores": 14.666666666666666, "mean_target_forward_passes": 6.333333333333333, "tokens_per_second": 6.503546528218914 }, { "all_exact": true, "config": { "margin_cutoff": 0.05, "margin_slope": 1.5, "stochastic_budget": 2, "temperature": 0.06 }, "config_key": "T=0.06|cut=0.05|budget=2|slope=1.5", "mean_acceptance_rate": 0.0872762714867978, "mean_candidate_scores": 2.0, "mean_target_forward_passes": 6.333333333333333, "tokens_per_second": 6.500557625146053 }, { "all_exact": true, "config": { "margin_cutoff": 0.15, "margin_slope": 1.5, "stochastic_budget": 1, "temperature": 0.16 }, "config_key": "T=0.16|cut=0.15|budget=1|slope=1.5", "mean_acceptance_rate": 0.0872762714867978, "mean_candidate_scores": 7.333333333333333, "mean_target_forward_passes": 6.333333333333333, "tokens_per_second": 6.397925036894238 }, { "all_exact": true, "config": { "margin_cutoff": 0.35, "margin_slope": 1.5, "stochastic_budget": 2, "temperature": 0.16 }, "config_key": "T=0.16|cut=0.35|budget=2|slope=1.5", "mean_acceptance_rate": 0.0872762714867978, "mean_candidate_scores": 14.666666666666666, "mean_target_forward_passes": 6.333333333333333, "tokens_per_second": 6.5722068118653985 } ] }, "dflash7_act": [ { "all_exact": true, "mean_acceptance_rate": 0.0872762714867978, "mean_verified_drafts_per_block": 3.1031746031746033, "threshold": 0.0, "tokens_per_second": 6.47785334278853 }, { "all_exact": true, "mean_acceptance_rate": 0.1716269841269841, "mean_verified_drafts_per_block": 1.8174603174603174, "threshold": 0.25, "tokens_per_second": 6.759485398495714 }, { "all_exact": true, "mean_acceptance_rate": 0.1970899470899471, "mean_verified_drafts_per_block": 1.4603174603174605, "threshold": 0.5, "tokens_per_second": 6.715956662052035 }, { "all_exact": true, "mean_acceptance_rate": 0.2698412698412698, "mean_verified_drafts_per_block": 1.0, "threshold": 1.0, "tokens_per_second": 6.890091232074371 }, { "all_exact": true, "mean_acceptance_rate": 0.2698412698412698, "mean_verified_drafts_per_block": 1.0, "threshold": 2.0, "tokens_per_second": 6.8076862999093235 }, { "all_exact": true, "mean_acceptance_rate": 0.2698412698412698, "mean_verified_drafts_per_block": 1.0, "threshold": 4.0, "tokens_per_second": 6.861062758724484 }, { "all_exact": true, "mean_acceptance_rate": 0.2698412698412698, "mean_verified_drafts_per_block": 1.0, "threshold": 8.0, "tokens_per_second": 6.805828209440527 } ], "dspark_v9": [ { "all_exact": true, "mean_acceptance_rate": 0.06734006734006734, "mean_target_forward_passes": 6.666666666666667, "survival_floor": 0.0, "tokens_per_second": 5.973447874860754 }, { "all_exact": true, "mean_acceptance_rate": 0.09433760683760684, "mean_target_forward_passes": 6.666666666666667, "survival_floor": 0.05, "tokens_per_second": 6.209839602420841 }, { "all_exact": true, "mean_acceptance_rate": 0.10073260073260075, "mean_target_forward_passes": 6.666666666666667, "survival_floor": 0.1, "tokens_per_second": 6.225443039567187 }, { "all_exact": true, "mean_acceptance_rate": 0.1537037037037037, "mean_target_forward_passes": 6.666666666666667, "survival_floor": 0.2, "tokens_per_second": 6.419748673964916 }, { "all_exact": true, "mean_acceptance_rate": 0.20039682539682538, "mean_target_forward_passes": 6.666666666666667, "survival_floor": 0.35, "tokens_per_second": 6.50903892538947 }, { "all_exact": true, "mean_acceptance_rate": 0.20634920634920637, "mean_target_forward_passes": 6.666666666666667, "survival_floor": 0.5, "tokens_per_second": 6.37197612715654 }, { "all_exact": true, "mean_acceptance_rate": 0.20634920634920637, "mean_target_forward_passes": 6.666666666666667, "survival_floor": 0.7, "tokens_per_second": 6.396342157071125 } ], "minop_v13": { "reuses_v10_config": true, "routing_budget": 1 }, "parareal_linear_v12": { "contraction_ratio_by_round": [ 0.9538399916842966, 0.9553103903397087 ], "examples": 120, "fine_top1_agreement_by_round": [ 0.5479166666666667, 0.5291666666666667, 0.5020833333333333 ], "log_teacher_mse_by_round": [ 3.2050065597068795, 3.1577472145044867, 3.112028239259323 ], "teacher_mse_by_round": [ 24.655661749839783, 23.517556198438008, 22.46656579176585 ] }, "simple_parareal_v14": { "coarse_sign_accuracy": 0.69375, "corrected_sign_accuracy": 0.69375, "fine_gap_mae": 7.1506517525257465, "fine_gap_mse": 76.69805961652828, "mean_coarse_gap": 0.9496589293082555, "mean_estimated_fine_gap": 3.658739121657603, "mean_fine_gap": 3.9155469074845315, "predicted_switch_rate": 0.0, "rows": 480, "switch_precision": 0.0, "switch_recall": 0.0, "teacher_prefers_top2_rate": 0.30625 }, "verify_widths": { "enabled": false } }, "config": { "bmobs_temperature": 0.35, "boltzmann_temperature": 0.15, "bonus_token": true, "calibration_prompt_count": 3, "calibration_tokens": 8, "cpu_threads": 2, "dtype": "float32", "fused_weight": 1.0, "jump_weight": 0.5, "max_new_tokens": 24, "prompt_count": 6, "repeats": 3, "target_argmax_on_device": true, "target_cache": false, "target_suffix_projection": true, "top_k": 8, "verify_width_tuning": false }, "examples": [ { "outputs": { "boltzmann_gated_mobs_v11": "The capital of Austria is which city?\n\nA) Vienna\nB) Munich\nC) Salzburg\nD) Graz\n\nCorrect Answer: A", "boltzmann_v10": "The capital of Austria is which city?\n\nA) Vienna\nB) Munich\nC) Salzburg\nD) Graz\n\nCorrect Answer: A", "dflash": "The capital of Austria is which city?\n\nA) Vienna\nB) Munich\nC) Salzburg\nD) Graz\n\nCorrect Answer: A", "dflash2": "The capital of Austria is which city?\n\nA) Vienna\nB) Munich\nC) Salzburg\nD) Graz\n\nCorrect Answer: A", "dflash3_mobs": "The capital of Austria is which city?\n\nA) Vienna\nB) Munich\nC) Salzburg\nD) Graz\n\nCorrect Answer: A", "dflash4_jump_mobs": "The capital of Austria is which city?\n\nA) Vienna\nB) Munich\nC) Salzburg\nD) Graz\n\nCorrect Answer: A", "dflash5_fused_jump_mobs": "The capital of Austria is which city?\n\nA) Vienna\nB) Munich\nC) Salzburg\nD) Graz\n\nCorrect Answer: A", "dflash6_bmobs": "The capital of Austria is which city?\n\nA) Vienna\nB) Munich\nC) Salzburg\nD) Graz\n\nCorrect Answer: A", "dflash6_boltzmann": "The capital of Austria is which city?\n\nA) Vienna\nB) Munich\nC) Salzburg\nD) Graz\n\nCorrect Answer: A", "dflash7_act": "The capital of Austria is which city?\n\nA) Vienna\nB) Munich\nC) Salzburg\nD) Graz\n\nCorrect Answer: A", "dspark_v9": "The capital of Austria is which city?\n\nA) Vienna\nB) Munich\nC) Salzburg\nD) Graz\n\nCorrect Answer: A", "minop_v13": "The capital of Austria is which city?\n\nA) Vienna\nB) Munich\nC) Salzburg\nD) Graz\n\nCorrect Answer: A", "normal": "The capital of Austria is which city?\n\nA) Vienna\nB) Munich\nC) Salzburg\nD) Graz\n\nCorrect Answer: A", "parareal_linear_v12": "The capital of Austria is which city?\n\nA) Vienna\nB) Munich\nC) Salzburg\nD) Graz\n\nCorrect Answer: A", "simple_parareal_v14": "The capital of Austria is which city?\n\nA) Vienna\nB) Munich\nC) Salzburg\nD) Graz\n\nCorrect Answer: A" }, "prompt": "The capital of Austria is", "reference": "The capital of Austria is which city?\n\nA) Vienna\nB) Munich\nC) Salzburg\nD) Graz\n\nCorrect Answer: A" }, { "outputs": { "boltzmann_gated_mobs_v11": "Speculative decoding can improve language model inference by providing a more accurate representation of the input data. This is because the model can learn to predict the next word in a", "boltzmann_v10": "Speculative decoding can improve language model inference by providing a more accurate representation of the input data. This is because the model can learn to predict the next word in a", "dflash": "Speculative decoding can improve language model inference by providing a more accurate representation of the input data. This is because the model can learn to predict the next word in a", "dflash2": "Speculative decoding can improve language model inference by providing a more accurate representation of the input data. This is because the model can learn to predict the next word in a", "dflash3_mobs": "Speculative decoding can improve language model inference by providing a more accurate representation of the input data. This is because the model can learn to predict the next word in a", "dflash4_jump_mobs": "Speculative decoding can improve language model inference by providing a more accurate representation of the input data. This is because the model can learn to predict the next word in a", "dflash5_fused_jump_mobs": "Speculative decoding can improve language model inference by providing a more accurate representation of the input data. This is because the model can learn to predict the next word in a", "dflash6_bmobs": "Speculative decoding can improve language model inference by providing a more accurate representation of the input data. This is because the model can learn to predict the next word in a", "dflash6_boltzmann": "Speculative decoding can improve language model inference by providing a more accurate representation of the input data. This is because the model can learn to predict the next word in a", "dflash7_act": "Speculative decoding can improve language model inference by providing a more accurate representation of the input data. This is because the model can learn to predict the next word in a", "dspark_v9": "Speculative decoding can improve language model inference by providing a more accurate representation of the input data. This is because the model can learn to predict the next word in a", "minop_v13": "Speculative decoding can improve language model inference by providing a more accurate representation of the input data. This is because the model can learn to predict the next word in a", "normal": "Speculative decoding can improve language model inference by providing a more accurate representation of the input data. This is because the model can learn to predict the next word in a", "parareal_linear_v12": "Speculative decoding can improve language model inference by providing a more accurate representation of the input data. This is because the model can learn to predict the next word in a", "simple_parareal_v14": "Speculative decoding can improve language model inference by providing a more accurate representation of the input data. This is because the model can learn to predict the next word in a" }, "prompt": "Speculative decoding can improve language model inference by", "reference": "Speculative decoding can improve language model inference by providing a more accurate representation of the input data. This is because the model can learn to predict the next word in a" }, { "outputs": { "boltzmann_gated_mobs_v11": "A digital twin represents a city's infrastructure. If the city's population is 500,000 and the digital twin's data processing speed is", "boltzmann_v10": "A digital twin represents a city's infrastructure. If the city's population is 500,000 and the digital twin's data processing speed is", "dflash": "A digital twin represents a city's infrastructure. If the city's population is 500,000 and the digital twin's data processing speed is", "dflash2": "A digital twin represents a city's infrastructure. If the city's population is 500,000 and the digital twin's data processing speed is", "dflash3_mobs": "A digital twin represents a city's infrastructure. If the city's population is 500,000 and the digital twin's data processing speed is", "dflash4_jump_mobs": "A digital twin represents a city's infrastructure. If the city's population is 500,000 and the digital twin's data processing speed is", "dflash5_fused_jump_mobs": "A digital twin represents a city's infrastructure. If the city's population is 500,000 and the digital twin's data processing speed is", "dflash6_bmobs": "A digital twin represents a city's infrastructure. If the city's population is 500,000 and the digital twin's data processing speed is", "dflash6_boltzmann": "A digital twin represents a city's infrastructure. If the city's population is 500,000 and the digital twin's data processing speed is", "dflash7_act": "A digital twin represents a city's infrastructure. If the city's population is 500,000 and the digital twin's data processing speed is", "dspark_v9": "A digital twin represents a city's infrastructure. If the city's population is 500,000 and the digital twin's data processing speed is", "minop_v13": "A digital twin represents a city's infrastructure. If the city's population is 500,000 and the digital twin's data processing speed is", "normal": "A digital twin represents a city's infrastructure. If the city's population is 500,000 and the digital twin's data processing speed is", "parareal_linear_v12": "A digital twin represents a city's infrastructure. If the city's population is 500,000 and the digital twin's data processing speed is", "simple_parareal_v14": "A digital twin represents a city's infrastructure. If the city's population is 500,000 and the digital twin's data processing speed is" }, "prompt": "A digital twin represents", "reference": "A digital twin represents a city's infrastructure. If the city's population is 500,000 and the digital twin's data processing speed is" }, { "outputs": { "boltzmann_gated_mobs_v11": "Reliable software systems should be designed to be resilient to failures and to recover from them. The term \"fault tolerance\" refers to the ability", "boltzmann_v10": "Reliable software systems should be designed to be resilient to failures and to recover from them. The term \"fault tolerance\" refers to the ability", "dflash": "Reliable software systems should be designed to be resilient to failures and to recover from them. The term \"fault tolerance\" refers to the ability", "dflash2": "Reliable software systems should be designed to be resilient to failures and to recover from them. The term \"fault tolerance\" refers to the ability", "dflash3_mobs": "Reliable software systems should be designed to be resilient to failures and to recover from them. The term \"fault tolerance\" refers to the ability", "dflash4_jump_mobs": "Reliable software systems should be designed to be resilient to failures and to recover from them. The term \"fault tolerance\" refers to the ability", "dflash5_fused_jump_mobs": "Reliable software systems should be designed to be resilient to failures and to recover from them. The term \"fault tolerance\" refers to the ability", "dflash6_bmobs": "Reliable software systems should be designed to be resilient to failures and to recover from them. The term \"fault tolerance\" refers to the ability", "dflash6_boltzmann": "Reliable software systems should be designed to be resilient to failures and to recover from them. The term \"fault tolerance\" refers to the ability", "dflash7_act": "Reliable software systems should be designed to be resilient to failures and to recover from them. The term \"fault tolerance\" refers to the ability", "dspark_v9": "Reliable software systems should be designed to be resilient to failures and to recover from them. The term \"fault tolerance\" refers to the ability", "minop_v13": "Reliable software systems should be designed to be resilient to failures and to recover from them. The term \"fault tolerance\" refers to the ability", "normal": "Reliable software systems should be designed to be resilient to failures and to recover from them. The term \"fault tolerance\" refers to the ability", "parareal_linear_v12": "Reliable software systems should be designed to be resilient to failures and to recover from them. The term \"fault tolerance\" refers to the ability", "simple_parareal_v14": "Reliable software systems should be designed to be resilient to failures and to recover from them. The term \"fault tolerance\" refers to the ability" }, "prompt": "Reliable software systems should", "reference": "Reliable software systems should be designed to be resilient to failures and to recover from them. The term \"fault tolerance\" refers to the ability" }, { "outputs": { "boltzmann_gated_mobs_v11": "Water distribution networks must be able to handle a large volume of water. The water supply network must be able to handle a large volume of water", "boltzmann_v10": "Water distribution networks must be able to handle a large volume of water. The water supply network must be able to handle a large volume of water", "dflash": "Water distribution networks must be able to handle a large volume of water. The water supply network must be able to handle a large volume of water", "dflash2": "Water distribution networks must be able to handle a large volume of water. The water supply network must be able to handle a large volume of water", "dflash3_mobs": "Water distribution networks must be able to handle a large volume of water. The water supply network must be able to handle a large volume of water", "dflash4_jump_mobs": "Water distribution networks must be able to handle a large volume of water. The water supply network must be able to handle a large volume of water", "dflash5_fused_jump_mobs": "Water distribution networks must be able to handle a large volume of water. The water supply network must be able to handle a large volume of water", "dflash6_bmobs": "Water distribution networks must be able to handle a large volume of water. The water supply network must be able to handle a large volume of water", "dflash6_boltzmann": "Water distribution networks must be able to handle a large volume of water. The water supply network must be able to handle a large volume of water", "dflash7_act": "Water distribution networks must be able to handle a large volume of water. The water supply network must be able to handle a large volume of water", "dspark_v9": "Water distribution networks must be able to handle a large volume of water. The water supply network must be able to handle a large volume of water", "minop_v13": "Water distribution networks must be able to handle a large volume of water. The water supply network must be able to handle a large volume of water", "normal": "Water distribution networks must be able to handle a large volume of water. The water supply network must be able to handle a large volume of water", "parareal_linear_v12": "Water distribution networks must be able to handle a large volume of water. The water supply network must be able to handle a large volume of water", "simple_parareal_v14": "Water distribution networks must be able to handle a large volume of water. The water supply network must be able to handle a large volume of water" }, "prompt": "Water distribution networks must", "reference": "Water distribution networks must be able to handle a large volume of water. The water supply network must be able to handle a large volume of water" }, { "outputs": { "boltzmann_gated_mobs_v11": "Artificial intelligence can support scientific research by providing tools for data analysis, hypothesis generation, and experimental design. AI can also assist in the interpretation of results, identify", "boltzmann_v10": "Artificial intelligence can support scientific research by providing tools for data analysis, hypothesis generation, and experimental design. AI can also assist in the interpretation of results, identify", "dflash": "Artificial intelligence can support scientific research by providing tools for data analysis, hypothesis generation, and experimental design. AI can also assist in the interpretation of results, identify", "dflash2": "Artificial intelligence can support scientific research by providing tools for data analysis, hypothesis generation, and experimental design. AI can also assist in the interpretation of results, identify", "dflash3_mobs": "Artificial intelligence can support scientific research by providing tools for data analysis, hypothesis generation, and experimental design. AI can also assist in the interpretation of results, identify", "dflash4_jump_mobs": "Artificial intelligence can support scientific research by providing tools for data analysis, hypothesis generation, and experimental design. AI can also assist in the interpretation of results, identify", "dflash5_fused_jump_mobs": "Artificial intelligence can support scientific research by providing tools for data analysis, hypothesis generation, and experimental design. AI can also assist in the interpretation of results, identify", "dflash6_bmobs": "Artificial intelligence can support scientific research by providing tools for data analysis, hypothesis generation, and experimental design. AI can also assist in the interpretation of results, identify", "dflash6_boltzmann": "Artificial intelligence can support scientific research by providing tools for data analysis, hypothesis generation, and experimental design. AI can also assist in the interpretation of results, identify", "dflash7_act": "Artificial intelligence can support scientific research by providing tools for data analysis, hypothesis generation, and experimental design. AI can also assist in the interpretation of results, identify", "dspark_v9": "Artificial intelligence can support scientific research by providing tools for data analysis, hypothesis generation, and experimental design. AI can also assist in the interpretation of results, identify", "minop_v13": "Artificial intelligence can support scientific research by providing tools for data analysis, hypothesis generation, and experimental design. AI can also assist in the interpretation of results, identify", "normal": "Artificial intelligence can support scientific research by providing tools for data analysis, hypothesis generation, and experimental design. AI can also assist in the interpretation of results, identify", "parareal_linear_v12": "Artificial intelligence can support scientific research by providing tools for data analysis, hypothesis generation, and experimental design. AI can also assist in the interpretation of results, identify", "simple_parareal_v14": "Artificial intelligence can support scientific research by providing tools for data analysis, hypothesis generation, and experimental design. AI can also assist in the interpretation of results, identify" }, "prompt": "Artificial intelligence can support scientific research by", "reference": "Artificial intelligence can support scientific research by providing tools for data analysis, hypothesis generation, and experimental design. AI can also assist in the interpretation of results, identify" } ], "method_info": { "boltzmann_gated_mobs_v11": { "complexity": "bounded sparse O(BK)", "index": 11, "label": "V11 Boltzmann-Gated MOBS", "mechanism": "Route only the most uncertain positions through deterministic Boltzmann-gated MOBS correction.", "short": "Uncertainty-gated sparse MOBS", "stages": [ "GATE", "SPARSE MOBS", "VERIFY", "COMMIT" ] }, "boltzmann_v10": { "complexity": "sparse O(BK)", "index": 10, "label": "V10 Advanced Boltzmann", "mechanism": "Keep confident slots on the argmax fast path and spend candidate work only on uncertain positions.", "short": "Sparse uncertainty routing", "stages": [ "MARGIN", "SPARSE SAMPLE", "VERIFY", "COMMIT" ] }, "dflash": { "complexity": "parallel draft + verify", "index": 1, "label": "DFlash", "mechanism": "Draft all future block positions in parallel, then verify the complete proposal with LFM2.5.", "short": "Parallel block argmax", "stages": [ "DRAFT", "BLOCK", "VERIFY", "COMMIT" ] }, "dflash2": { "complexity": "O(BK\u00b2) guidance", "index": 2, "label": "DFlash2", "mechanism": "Retain top-k candidates per slot and choose a predecessor-aware path before exact verification.", "short": "Top-k dynamic-programming path", "stages": [ "TOP-K", "DP PATH", "VERIFY", "COMMIT" ] }, "dflash3_mobs": { "complexity": "O(BK) guidance", "index": 3, "label": "DFlash3-MOBS", "mechanism": "Choose a central anchor and expand left/right using linear-in-K neighbor scoring.", "short": "Middle-out bidirectional selection", "stages": [ "TOP-K", "MIDDLE-OUT", "VERIFY", "COMMIT" ] }, "dflash4_jump_mobs": { "complexity": "O(BK + JK) + jump pass", "index": 4, "label": "DFlash4-JUMP-MOBS", "mechanism": "Predict sparse jump anchors, then fill the gaps with MOBS before target verification.", "short": "Sparse future anchors", "stages": [ "JUMP", "GAP FILL", "VERIFY", "COMMIT" ] }, "dflash5_fused_jump_mobs": { "complexity": "O(BK + JKR)", "index": 5, "label": "DFlash5-FUSED-JUMP", "mechanism": "Reuse drafter hidden states for sparse residual anchors, removing DFlash4's extra jump forward.", "short": "Fused sparse guidance", "stages": [ "FUSED", "ANCHORS", "VERIFY", "COMMIT" ] }, "dflash6_bmobs": { "complexity": "O(BK) guidance", "index": 7, "label": "DFlash6-BMOBS", "mechanism": "Explore one uncertain middle anchor, then fill the remaining block with MOBS.", "short": "Boltzmann anchor + MOBS", "stages": [ "ANCHOR", "MOBS FILL", "VERIFY", "COMMIT" ] }, "dflash6_boltzmann": { "complexity": "O(BK) guidance", "index": 6, "label": "DFlash6-Boltzmann", "mechanism": "Use confidence-adaptive deterministic Gumbel/Boltzmann scoring over retained candidates.", "short": "Deterministic Boltzmann exploration", "stages": [ "TOP-K", "BOLTZMANN", "VERIFY", "COMMIT" ] }, "dflash7_act": { "complexity": "O(B) routing", "index": 8, "label": "DFlash7-ACT", "mechanism": "Use draft margins to trim an uncertain suffix before expensive target verification.", "short": "Adaptive speculation horizon", "stages": [ "DRAFT", "TRIM", "VERIFY", "COMMIT" ] }, "dspark_v9": { "complexity": "O(BK) guidance", "index": 9, "label": "V9 DSpark-Lite", "mechanism": "Apply a low-rank previous-token correction and learned prefix-survival confidence.", "short": "Markov correction + survival gate", "stages": [ "MARKOV", "SURVIVAL", "VERIFY", "COMMIT" ] }, "minop_v13": { "complexity": "O(B) routing after fused top-2", "index": 13, "label": "V13 MinOp", "mechanism": "Preserve V10's sparse two-way decision, but select top-2 inside the Torch drafter and route at most one uncertain slot.", "short": "Fused top-2 V10 fast path", "stages": [ "FUSED TOP-2", "ONE SLOT", "VERIFY", "COMMIT" ] }, "normal": { "complexity": "N target steps", "index": 0, "label": "Normal autoregressive", "mechanism": "One authoritative LFM2.5 target step per generated token.", "short": "Target-only greedy baseline", "stages": [ "TARGET", "TOKEN", "TARGET", "TOKEN" ] }, "parareal_linear_v12": { "complexity": "O(RBKD) linear algebra", "index": 12, "label": "V12 PARAREAL", "mechanism": "Apply two vectorized affine F-G residual-correction rounds to the whole top-k score field, then verify exactly.", "short": "Parallel linear residual correction", "stages": [ "COARSE G", "F-G \u00d72", "VERIFY", "COMMIT" ] }, "simple_parareal_v14": { "complexity": "O(B) scalar estimator", "index": 14, "label": "V14 Simple PARAREAL", "mechanism": "Estimate the target fine top1-top2 gap with three affine terms, apply one Parareal residual update, and allow at most one top-2 switch.", "short": "Scalar fine-gap estimator", "stages": [ "COARSE GAP", "F-HAT", "VERIFY", "COMMIT" ] } }, "model": { "aux_parameter_count": 3646320, "candidate_size": 2048, "id": "LiquidAI/LFM2.5-350M-Base", "target_parameter_count": 354483968, "target_vocab_size": 65536 }, "rows": { "boltzmann_gated_mobs_v11": [ { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 78, "context_seconds": 0.003569547000097373, "draft_forward_passes": 21, "draft_seconds": 0.01912332000074457, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4809.039060000032, "method": "boltzmann_gated_mobs_v11", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 0, "selection_seconds": 0.0063602249997529725, "selector_pair_scores": 72, "target_forward_passes": 21, "target_seconds": 4.77835961700066, "tokens_per_second": 4.99060201020697, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 150, "v11_eligible_positions": 5, "v11_fast_argmax_positions": 79, "v11_gate_evaluations": 78, "v11_gated_positions": 5, "v11_mean_uncertainty": 0.024174495517104596, "v11_pair_scores": 72, "wall_seconds": 4.809039060000032 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 78, "context_seconds": 0.0035699039985956915, "draft_forward_passes": 21, "draft_seconds": 0.018817553000644693, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4624.0194169999995, "method": "boltzmann_gated_mobs_v11", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 1, "selection_seconds": 0.006347069000185002, "selector_pair_scores": 72, "target_forward_passes": 21, "target_seconds": 4.593726084998707, "tokens_per_second": 5.1902896237340785, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 150, "v11_eligible_positions": 5, "v11_fast_argmax_positions": 79, "v11_gate_evaluations": 78, "v11_gated_positions": 5, "v11_mean_uncertainty": 0.024174495517104596, "v11_pair_scores": 72, "wall_seconds": 4.6240194169999995 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 78, "context_seconds": 0.0036057599982086685, "draft_forward_passes": 21, "draft_seconds": 0.018962772001486883, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4638.277302000006, "method": "boltzmann_gated_mobs_v11", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 2, "selection_seconds": 0.0062298300013026164, "selector_pair_scores": 72, "target_forward_passes": 21, "target_seconds": 4.6078577589992165, "tokens_per_second": 5.174334874211014, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 150, "v11_eligible_positions": 5, "v11_fast_argmax_positions": 79, "v11_gate_evaluations": 78, "v11_gated_positions": 5, "v11_mean_uncertainty": 0.024174495517104596, "v11_pair_scores": 72, "wall_seconds": 4.638277302000006 }, { "acceptance_rate": 0.10606060606060606, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 66, "context_seconds": 0.0031061880008564913, "draft_forward_passes": 18, "draft_seconds": 0.016410982001161756, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4386.861838000186, "method": "boltzmann_gated_mobs_v11", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 66, "repeat": 0, "selection_seconds": 0.005150130000401987, "selector_pair_scores": 32, "target_forward_passes": 18, "target_seconds": 4.3608646490001775, "tokens_per_second": 5.4708812099130855, "tokens_per_target_pass": 1.3333333333333333, "total_guidance_scores": 98, "v11_eligible_positions": 2, "v11_fast_argmax_positions": 70, "v11_gate_evaluations": 66, "v11_gated_positions": 2, "v11_mean_uncertainty": 0.014111192589585507, "v11_pair_scores": 32, "wall_seconds": 4.386861838000186 }, { "acceptance_rate": 0.10606060606060606, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 66, "context_seconds": 0.0031559100002596097, "draft_forward_passes": 18, "draft_seconds": 0.016072575999260152, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4391.303095000239, "method": "boltzmann_gated_mobs_v11", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 66, "repeat": 1, "selection_seconds": 0.005247633999715617, "selector_pair_scores": 32, "target_forward_passes": 18, "target_seconds": 4.365472743998907, "tokens_per_second": 5.465348093900745, "tokens_per_target_pass": 1.3333333333333333, "total_guidance_scores": 98, "v11_eligible_positions": 2, "v11_fast_argmax_positions": 70, "v11_gate_evaluations": 66, "v11_gated_positions": 2, "v11_mean_uncertainty": 0.014111192589585507, "v11_pair_scores": 32, "wall_seconds": 4.391303095000239 }, { "acceptance_rate": 0.10606060606060606, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 66, "context_seconds": 0.0032614089996059192, "draft_forward_passes": 18, "draft_seconds": 0.01663669200024742, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4448.410402000263, "method": "boltzmann_gated_mobs_v11", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 66, "repeat": 2, "selection_seconds": 0.005067170999154769, "selector_pair_scores": 32, "target_forward_passes": 18, "target_seconds": 4.421975171998383, "tokens_per_second": 5.395185657602142, "tokens_per_target_pass": 1.3333333333333333, "total_guidance_scores": 98, "v11_eligible_positions": 2, "v11_fast_argmax_positions": 70, "v11_gate_evaluations": 66, "v11_gated_positions": 2, "v11_mean_uncertainty": 0.014111192589585507, "v11_pair_scores": 32, "wall_seconds": 4.448410402000263 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 70, "context_seconds": 0.0031947020015650196, "draft_forward_passes": 19, "draft_seconds": 0.017294142000082502, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4259.290481999869, "method": "boltzmann_gated_mobs_v11", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 0, "selection_seconds": 0.005672979999872041, "selector_pair_scores": 72, "target_forward_passes": 19, "target_seconds": 4.231708831002834, "tokens_per_second": 5.634741302906218, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 142, "v11_eligible_positions": 5, "v11_fast_argmax_positions": 71, "v11_gate_evaluations": 70, "v11_gated_positions": 5, "v11_mean_uncertainty": 0.027768076176993606, "v11_pair_scores": 72, "wall_seconds": 4.259290481999869 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 70, "context_seconds": 0.0032173849990613235, "draft_forward_passes": 19, "draft_seconds": 0.017337239000426052, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4338.099904000046, "method": "boltzmann_gated_mobs_v11", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 1, "selection_seconds": 0.005708298000172363, "selector_pair_scores": 72, "target_forward_passes": 19, "target_seconds": 4.310280773000159, "tokens_per_second": 5.532376047372777, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 142, "v11_eligible_positions": 5, "v11_fast_argmax_positions": 71, "v11_gate_evaluations": 70, "v11_gated_positions": 5, "v11_mean_uncertainty": 0.027768076176993606, "v11_pair_scores": 72, "wall_seconds": 4.338099904000046 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 70, "context_seconds": 0.0032668729995748436, "draft_forward_passes": 19, "draft_seconds": 0.017403544998160214, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4276.407907999783, "method": "boltzmann_gated_mobs_v11", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 2, "selection_seconds": 0.005874329000562284, "selector_pair_scores": 72, "target_forward_passes": 19, "target_seconds": 4.248444226001084, "tokens_per_second": 5.61218679703209, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 142, "v11_eligible_positions": 5, "v11_fast_argmax_positions": 71, "v11_gate_evaluations": 70, "v11_gated_positions": 5, "v11_mean_uncertainty": 0.027768076176993606, "v11_pair_scores": 72, "wall_seconds": 4.276407907999783 }, { "acceptance_rate": 0.05405405405405406, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 74, "context_seconds": 0.003372615999978734, "draft_forward_passes": 20, "draft_seconds": 0.017653805000463763, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4500.897627000086, "method": "boltzmann_gated_mobs_v11", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 74, "repeat": 0, "selection_seconds": 0.006097914999372733, "selector_pair_scores": 88, "target_forward_passes": 20, "target_seconds": 4.472290007000083, "tokens_per_second": 5.332269691278526, "tokens_per_target_pass": 1.2, "total_guidance_scores": 162, "v11_eligible_positions": 6, "v11_fast_argmax_positions": 74, "v11_gate_evaluations": 74, "v11_gated_positions": 6, "v11_mean_uncertainty": 0.025400523679757522, "v11_pair_scores": 88, "wall_seconds": 4.500897627000086 }, { "acceptance_rate": 0.05405405405405406, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 74, "context_seconds": 0.003453539000474848, "draft_forward_passes": 20, "draft_seconds": 0.017794511999909446, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4498.071257999982, "method": "boltzmann_gated_mobs_v11", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 74, "repeat": 1, "selection_seconds": 0.006143299999621377, "selector_pair_scores": 88, "target_forward_passes": 20, "target_seconds": 4.469138148000184, "tokens_per_second": 5.335620229963039, "tokens_per_target_pass": 1.2, "total_guidance_scores": 162, "v11_eligible_positions": 6, "v11_fast_argmax_positions": 74, "v11_gate_evaluations": 74, "v11_gated_positions": 6, "v11_mean_uncertainty": 0.025400523679757522, "v11_pair_scores": 88, "wall_seconds": 4.498071257999982 }, { "acceptance_rate": 0.05405405405405406, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 74, "context_seconds": 0.0032985010006996163, "draft_forward_passes": 20, "draft_seconds": 0.017567595000855363, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4481.168134999734, "method": "boltzmann_gated_mobs_v11", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 74, "repeat": 2, "selection_seconds": 0.006127356001343287, "selector_pair_scores": 88, "target_forward_passes": 20, "target_seconds": 4.452677625998149, "tokens_per_second": 5.355746376162568, "tokens_per_target_pass": 1.2, "total_guidance_scores": 162, "v11_eligible_positions": 6, "v11_fast_argmax_positions": 74, "v11_gate_evaluations": 74, "v11_gated_positions": 6, "v11_mean_uncertainty": 0.025400523679757522, "v11_pair_scores": 88, "wall_seconds": 4.481168134999734 }, { "acceptance_rate": 0.1076923076923077, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 65, "context_seconds": 0.002845748000254389, "draft_forward_passes": 17, "draft_seconds": 0.015366806000656652, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3766.676974999882, "method": "boltzmann_gated_mobs_v11", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 65, "repeat": 0, "selection_seconds": 0.005298495999795705, "selector_pair_scores": 64, "target_forward_passes": 17, "target_seconds": 3.741823693000697, "tokens_per_second": 6.371663978433073, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 129, "v11_eligible_positions": 5, "v11_fast_argmax_positions": 63, "v11_gate_evaluations": 65, "v11_gated_positions": 5, "v11_mean_uncertainty": 0.03210822862775632, "v11_pair_scores": 64, "wall_seconds": 3.766676974999882 }, { "acceptance_rate": 0.1076923076923077, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 65, "context_seconds": 0.002762681000149314, "draft_forward_passes": 17, "draft_seconds": 0.015218512001411, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3809.3852079996395, "method": "boltzmann_gated_mobs_v11", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 65, "repeat": 1, "selection_seconds": 0.005225955000241811, "selector_pair_scores": 64, "target_forward_passes": 17, "target_seconds": 3.784852470998885, "tokens_per_second": 6.300229220610307, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 129, "v11_eligible_positions": 5, "v11_fast_argmax_positions": 63, "v11_gate_evaluations": 65, "v11_gated_positions": 5, "v11_mean_uncertainty": 0.03210822862775632, "v11_pair_scores": 64, "wall_seconds": 3.8093852079996395 }, { "acceptance_rate": 0.1076923076923077, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 65, "context_seconds": 0.0027899410006284597, "draft_forward_passes": 17, "draft_seconds": 0.015326430000641267, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3714.756963000127, "method": "boltzmann_gated_mobs_v11", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 65, "repeat": 2, "selection_seconds": 0.005320294000284775, "selector_pair_scores": 64, "target_forward_passes": 17, "target_seconds": 3.6900635469996814, "tokens_per_second": 6.460718760081957, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 129, "v11_eligible_positions": 5, "v11_fast_argmax_positions": 63, "v11_gate_evaluations": 65, "v11_gated_positions": 5, "v11_mean_uncertainty": 0.03210822862775632, "v11_pair_scores": 64, "wall_seconds": 3.714756963000127 }, { "acceptance_rate": 0.05333333333333334, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 75, "context_seconds": 0.0035152219998053624, "draft_forward_passes": 20, "draft_seconds": 0.018141191000722756, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4672.581500999968, "method": "boltzmann_gated_mobs_v11", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 75, "repeat": 0, "selection_seconds": 0.00585443199997826, "selector_pair_scores": 56, "target_forward_passes": 20, "target_seconds": 4.643596218998937, "tokens_per_second": 5.136347005368193, "tokens_per_target_pass": 1.2, "total_guidance_scores": 131, "v11_eligible_positions": 4, "v11_fast_argmax_positions": 76, "v11_gate_evaluations": 75, "v11_gated_positions": 4, "v11_mean_uncertainty": 0.02141236424844414, "v11_pair_scores": 56, "wall_seconds": 4.672581500999968 }, { "acceptance_rate": 0.05333333333333334, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 75, "context_seconds": 0.003421167000396963, "draft_forward_passes": 20, "draft_seconds": 0.017660188000263588, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4647.457666999799, "method": "boltzmann_gated_mobs_v11", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 75, "repeat": 1, "selection_seconds": 0.005807793999792921, "selector_pair_scores": 56, "target_forward_passes": 20, "target_seconds": 4.619083734001379, "tokens_per_second": 5.164113741243259, "tokens_per_target_pass": 1.2, "total_guidance_scores": 131, "v11_eligible_positions": 4, "v11_fast_argmax_positions": 76, "v11_gate_evaluations": 75, "v11_gated_positions": 4, "v11_mean_uncertainty": 0.02141236424844414, "v11_pair_scores": 56, "wall_seconds": 4.647457666999799 }, { "acceptance_rate": 0.05333333333333334, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 75, "context_seconds": 0.00351744899944606, "draft_forward_passes": 20, "draft_seconds": 0.01813113600064753, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4716.592277000018, "method": "boltzmann_gated_mobs_v11", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 75, "repeat": 2, "selection_seconds": 0.005975928001134889, "selector_pair_scores": 56, "target_forward_passes": 20, "target_seconds": 4.687503832001312, "tokens_per_second": 5.0884194754406815, "tokens_per_target_pass": 1.2, "total_guidance_scores": 131, "v11_eligible_positions": 4, "v11_fast_argmax_positions": 76, "v11_gate_evaluations": 75, "v11_gated_positions": 4, "v11_mean_uncertainty": 0.02141236424844414, "v11_pair_scores": 56, "wall_seconds": 4.716592277000018 } ], "boltzmann_v10": [ { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 28, "context_seconds": 0.0035797639998236264, "draft_forward_passes": 21, "draft_seconds": 0.01904764200116915, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4780.386031000035, "method": "boltzmann_v10", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 0, "selection_seconds": 0.004885568000645435, "selector_pair_scores": 0, "target_forward_passes": 21, "target_seconds": 4.751315018999321, "tokens_per_second": 5.020515047187374, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 28, "v10_candidate_scores": 28, "v10_fast_argmax_positions": 70, "v10_sampled_positions": 14, "v10_uncertain_positions": 15, "wall_seconds": 4.780386031000035 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 28, "context_seconds": 0.003556840000783268, "draft_forward_passes": 21, "draft_seconds": 0.01891476000082548, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4658.4122760000355, "method": "boltzmann_v10", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 1, "selection_seconds": 0.004854626999986067, "selector_pair_scores": 0, "target_forward_passes": 21, "target_seconds": 4.629506851000315, "tokens_per_second": 5.15196993697769, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 28, "v10_candidate_scores": 28, "v10_fast_argmax_positions": 70, "v10_sampled_positions": 14, "v10_uncertain_positions": 15, "wall_seconds": 4.6584122760000355 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 28, "context_seconds": 0.00352852700098083, "draft_forward_passes": 21, "draft_seconds": 0.01902516400014065, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4629.671045000123, "method": "boltzmann_v10", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 2, "selection_seconds": 0.004714575999059889, "selector_pair_scores": 0, "target_forward_passes": 21, "target_seconds": 4.60079659099938, "tokens_per_second": 5.183953625802233, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 28, "v10_candidate_scores": 28, "v10_fast_argmax_positions": 70, "v10_sampled_positions": 14, "v10_uncertain_positions": 15, "wall_seconds": 4.629671045000123 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 36, "context_seconds": 0.0032999399991240352, "draft_forward_passes": 19, "draft_seconds": 0.017181255998821143, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4534.742253000331, "method": "boltzmann_v10", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 70, "repeat": 0, "selection_seconds": 0.004552402000626898, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.50828072899867, "tokens_per_second": 5.292472793601627, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 36, "v10_candidate_scores": 36, "v10_fast_argmax_positions": 58, "v10_sampled_positions": 18, "v10_uncertain_positions": 21, "wall_seconds": 4.534742253000331 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 36, "context_seconds": 0.003341195001212327, "draft_forward_passes": 19, "draft_seconds": 0.017603612000129942, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4784.048171999984, "method": "boltzmann_v10", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 70, "repeat": 1, "selection_seconds": 0.004634056000213604, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.757001201000094, "tokens_per_second": 5.016671893160878, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 36, "v10_candidate_scores": 36, "v10_fast_argmax_positions": 58, "v10_sampled_positions": 18, "v10_uncertain_positions": 21, "wall_seconds": 4.784048171999984 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 36, "context_seconds": 0.0032757000003584835, "draft_forward_passes": 19, "draft_seconds": 0.01713342100083537, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4615.139408000232, "method": "boltzmann_v10", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 70, "repeat": 2, "selection_seconds": 0.0045729000003120746, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.588715604001209, "tokens_per_second": 5.200276281664771, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 36, "v10_candidate_scores": 36, "v10_fast_argmax_positions": 58, "v10_sampled_positions": 18, "v10_uncertain_positions": 21, "wall_seconds": 4.615139408000232 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 34, "context_seconds": 0.0032101399992825463, "draft_forward_passes": 19, "draft_seconds": 0.016784167999503552, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4170.791445999839, "method": "boltzmann_v10", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 0, "selection_seconds": 0.004475641999761137, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.144885346999672, "tokens_per_second": 5.754303544239341, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 34, "v10_candidate_scores": 34, "v10_fast_argmax_positions": 59, "v10_sampled_positions": 17, "v10_uncertain_positions": 17, "wall_seconds": 4.170791445999839 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 34, "context_seconds": 0.003227470999718207, "draft_forward_passes": 19, "draft_seconds": 0.01743951799835486, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4286.106954999923, "method": "boltzmann_v10", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 1, "selection_seconds": 0.004580797000016901, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.259460776999276, "tokens_per_second": 5.599486959139691, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 34, "v10_candidate_scores": 34, "v10_fast_argmax_positions": 59, "v10_sampled_positions": 17, "v10_uncertain_positions": 17, "wall_seconds": 4.286106954999923 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 34, "context_seconds": 0.003347327000938094, "draft_forward_passes": 19, "draft_seconds": 0.018095678000008775, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4304.095272999803, "method": "boltzmann_v10", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 2, "selection_seconds": 0.0047980720005398325, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.2764118200002486, "tokens_per_second": 5.576084746672637, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 34, "v10_candidate_scores": 34, "v10_fast_argmax_positions": 59, "v10_sampled_positions": 17, "v10_uncertain_positions": 17, "wall_seconds": 4.3040952729998025 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 48, "context_seconds": 0.0034742200004984625, "draft_forward_passes": 21, "draft_seconds": 0.018744770000466815, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4700.196272999619, "method": "boltzmann_v10", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 78, "repeat": 0, "selection_seconds": 0.0049082909999924595, "selector_pair_scores": 0, "target_forward_passes": 21, "target_seconds": 4.671429233001163, "tokens_per_second": 5.10616974398889, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 48, "v10_candidate_scores": 48, "v10_fast_argmax_positions": 60, "v10_sampled_positions": 24, "v10_uncertain_positions": 26, "wall_seconds": 4.700196272999619 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 48, "context_seconds": 0.003511788000650995, "draft_forward_passes": 21, "draft_seconds": 0.0186409050011207, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4758.164086999841, "method": "boltzmann_v10", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 78, "repeat": 1, "selection_seconds": 0.005029014999308856, "selector_pair_scores": 0, "target_forward_passes": 21, "target_seconds": 4.729421033001472, "tokens_per_second": 5.043962242826453, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 48, "v10_candidate_scores": 48, "v10_fast_argmax_positions": 60, "v10_sampled_positions": 24, "v10_uncertain_positions": 26, "wall_seconds": 4.758164086999841 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 48, "context_seconds": 0.003605416999107547, "draft_forward_passes": 21, "draft_seconds": 0.01899425200008409, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4795.963545000177, "method": "boltzmann_v10", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 78, "repeat": 2, "selection_seconds": 0.005023772998811182, "selector_pair_scores": 0, "target_forward_passes": 21, "target_seconds": 4.766758735998337, "tokens_per_second": 5.004208179401229, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 48, "v10_candidate_scores": 48, "v10_fast_argmax_positions": 60, "v10_sampled_positions": 24, "v10_uncertain_positions": 26, "wall_seconds": 4.795963545000177 }, { "acceptance_rate": 0.15789473684210525, "accepted_draft_tokens": 9, "boltzmann_candidate_scores": 38, "context_seconds": 0.002535591999276221, "draft_forward_passes": 15, "draft_seconds": 0.013491562001036073, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3364.201437000247, "method": "boltzmann_v10", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 57, "repeat": 0, "selection_seconds": 0.0035296859987283824, "selector_pair_scores": 0, "target_forward_passes": 15, "target_seconds": 3.343474699998751, "tokens_per_second": 7.133936670986041, "tokens_per_target_pass": 1.6, "total_guidance_scores": 38, "v10_candidate_scores": 38, "v10_fast_argmax_positions": 41, "v10_sampled_positions": 19, "v10_uncertain_positions": 21, "wall_seconds": 3.364201437000247 }, { "acceptance_rate": 0.15789473684210525, "accepted_draft_tokens": 9, "boltzmann_candidate_scores": 38, "context_seconds": 0.002451722999467165, "draft_forward_passes": 15, "draft_seconds": 0.013342877999093616, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3314.046646999941, "method": "boltzmann_v10", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 57, "repeat": 1, "selection_seconds": 0.0034454609990461904, "selector_pair_scores": 0, "target_forward_passes": 15, "target_seconds": 3.2936952869995366, "tokens_per_second": 7.241901685881861, "tokens_per_target_pass": 1.6, "total_guidance_scores": 38, "v10_candidate_scores": 38, "v10_fast_argmax_positions": 41, "v10_sampled_positions": 19, "v10_uncertain_positions": 21, "wall_seconds": 3.314046646999941 }, { "acceptance_rate": 0.15789473684210525, "accepted_draft_tokens": 9, "boltzmann_candidate_scores": 38, "context_seconds": 0.002549111998632725, "draft_forward_passes": 15, "draft_seconds": 0.013313678999566037, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3275.4771730001266, "method": "boltzmann_v10", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 57, "repeat": 2, "selection_seconds": 0.0036024179994456063, "selector_pair_scores": 0, "target_forward_passes": 15, "target_seconds": 3.254912106998745, "tokens_per_second": 7.327176692859545, "tokens_per_target_pass": 1.6, "total_guidance_scores": 38, "v10_candidate_scores": 38, "v10_fast_argmax_positions": 41, "v10_sampled_positions": 19, "v10_uncertain_positions": 21, "wall_seconds": 3.2754771730001266 }, { "acceptance_rate": 0.05333333333333334, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 42, "context_seconds": 0.003461917000095127, "draft_forward_passes": 20, "draft_seconds": 0.018069430001105502, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4717.412800000147, "method": "boltzmann_v10", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 75, "repeat": 0, "selection_seconds": 0.0047079400001166505, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.689639259999694, "tokens_per_second": 5.0875344214098135, "tokens_per_target_pass": 1.2, "total_guidance_scores": 42, "v10_candidate_scores": 42, "v10_fast_argmax_positions": 59, "v10_sampled_positions": 21, "v10_uncertain_positions": 22, "wall_seconds": 4.717412800000147 }, { "acceptance_rate": 0.05333333333333334, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 42, "context_seconds": 0.0033701990005283733, "draft_forward_passes": 20, "draft_seconds": 0.017911074000949156, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4667.595043000347, "method": "boltzmann_v10", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 75, "repeat": 1, "selection_seconds": 0.004688099001214141, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.6401521550001235, "tokens_per_second": 5.141834237738995, "tokens_per_target_pass": 1.2, "total_guidance_scores": 42, "v10_candidate_scores": 42, "v10_fast_argmax_positions": 59, "v10_sampled_positions": 21, "v10_uncertain_positions": 22, "wall_seconds": 4.667595043000347 }, { "acceptance_rate": 0.05333333333333334, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 42, "context_seconds": 0.0035106859991174133, "draft_forward_passes": 20, "draft_seconds": 0.018001646998982324, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4718.113828999776, "method": "boltzmann_v10", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 75, "repeat": 2, "selection_seconds": 0.004748226000629074, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.690288765999867, "tokens_per_second": 5.086778502986631, "tokens_per_target_pass": 1.2, "total_guidance_scores": 42, "v10_candidate_scores": 42, "v10_fast_argmax_positions": 59, "v10_sampled_positions": 21, "v10_uncertain_positions": 22, "wall_seconds": 4.718113828999776 } ], "dflash": [ { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.003617022000526049, "draft_forward_passes": 21, "draft_seconds": 0.019459171000562492, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4767.657655999756, "method": "dflash", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 0, "selection_seconds": 0.00042798100048457854, "selector_pair_scores": 0, "target_forward_passes": 21, "target_seconds": 4.742411404999075, "tokens_per_second": 5.03391848401651, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 0, "wall_seconds": 4.767657655999756 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.0035926450000260957, "draft_forward_passes": 21, "draft_seconds": 0.01876894500128401, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4669.367974000124, "method": "dflash", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 1, "selection_seconds": 0.0004349889991317468, "selector_pair_scores": 0, "target_forward_passes": 21, "target_seconds": 4.644832280000628, "tokens_per_second": 5.139881914134052, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 0, "wall_seconds": 4.669367974000124 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.0035355000004528847, "draft_forward_passes": 21, "draft_seconds": 0.01865663500075243, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4632.001917999787, "method": "dflash", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 2, "selection_seconds": 0.000403643000026932, "selector_pair_scores": 0, "target_forward_passes": 21, "target_seconds": 4.607737709000958, "tokens_per_second": 5.181345004788727, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 0, "wall_seconds": 4.632001917999787 }, { "acceptance_rate": 0.10606060606060606, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.0033062900001823436, "draft_forward_passes": 18, "draft_seconds": 0.0169786970000132, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4394.022814999971, "method": "dflash", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 66, "repeat": 0, "selection_seconds": 0.00036786900000151945, "selector_pair_scores": 0, "target_forward_passes": 18, "target_seconds": 4.371823820000827, "tokens_per_second": 5.461965267469863, "tokens_per_target_pass": 1.3333333333333333, "total_guidance_scores": 0, "wall_seconds": 4.394022814999971 }, { "acceptance_rate": 0.10606060606060606, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.003129110999907425, "draft_forward_passes": 18, "draft_seconds": 0.016379804999360204, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4486.809223999899, "method": "dflash", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 66, "repeat": 1, "selection_seconds": 0.00037008899971624487, "selector_pair_scores": 0, "target_forward_passes": 18, "target_seconds": 4.4654281940001965, "tokens_per_second": 5.349012806611931, "tokens_per_target_pass": 1.3333333333333333, "total_guidance_scores": 0, "wall_seconds": 4.486809223999899 }, { "acceptance_rate": 0.10606060606060606, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.0031335490002675215, "draft_forward_passes": 18, "draft_seconds": 0.01615372100013701, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4382.688420999784, "method": "dflash", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 66, "repeat": 2, "selection_seconds": 0.0003615729992816341, "selector_pair_scores": 0, "target_forward_passes": 18, "target_seconds": 4.3615640800007895, "tokens_per_second": 5.476090858981276, "tokens_per_target_pass": 1.3333333333333333, "total_guidance_scores": 0, "wall_seconds": 4.382688420999784 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.0032011029984460038, "draft_forward_passes": 19, "draft_seconds": 0.01714161800009606, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4235.019190000003, "method": "dflash", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 0, "selection_seconds": 0.000418232000356511, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.212769615001434, "tokens_per_second": 5.667034533555439, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 0, "wall_seconds": 4.235019190000003 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.003261386000303901, "draft_forward_passes": 19, "draft_seconds": 0.017309060999650683, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4297.325980999631, "method": "dflash", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 1, "selection_seconds": 0.00044943999864699435, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.2747630700005175, "tokens_per_second": 5.584868382364885, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 0, "wall_seconds": 4.297325980999631 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.0033895129995471507, "draft_forward_passes": 19, "draft_seconds": 0.01737913300121363, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4208.279830000265, "method": "dflash", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 2, "selection_seconds": 0.00036871100155622116, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.185458392000328, "tokens_per_second": 5.703042803595713, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 0, "wall_seconds": 4.208279830000265 }, { "acceptance_rate": 0.05405405405405406, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 0, "context_seconds": 0.0033348649994877633, "draft_forward_passes": 20, "draft_seconds": 0.01762713499874735, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4489.9487710004, "method": "dflash", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 74, "repeat": 0, "selection_seconds": 0.00038438799947471125, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.467048608998994, "tokens_per_second": 5.345272568589372, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "wall_seconds": 4.4899487710004 }, { "acceptance_rate": 0.05405405405405406, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 0, "context_seconds": 0.003360263999638846, "draft_forward_passes": 20, "draft_seconds": 0.017647617999500653, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4486.023530999773, "method": "dflash", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 74, "repeat": 1, "selection_seconds": 0.0003902380012732465, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.463039802999901, "tokens_per_second": 5.349949645638899, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "wall_seconds": 4.486023530999773 }, { "acceptance_rate": 0.05405405405405406, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 0, "context_seconds": 0.0034845659993152367, "draft_forward_passes": 20, "draft_seconds": 0.017564609000601195, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4517.9479659996105, "method": "dflash", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 74, "repeat": 2, "selection_seconds": 0.0003960300005019235, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.494845510999312, "tokens_per_second": 5.3121461735759326, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "wall_seconds": 4.5179479659996105 }, { "acceptance_rate": 0.1076923076923077, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.002924059001088608, "draft_forward_passes": 17, "draft_seconds": 0.014773227000205225, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3671.2589409999055, "method": "dflash", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 65, "repeat": 0, "selection_seconds": 0.0003183140001965512, "selector_pair_scores": 0, "target_forward_passes": 17, "target_seconds": 3.651885157000379, "tokens_per_second": 6.5372670208501695, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 0, "wall_seconds": 3.6712589409999055 }, { "acceptance_rate": 0.1076923076923077, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.0029472389996954007, "draft_forward_passes": 17, "draft_seconds": 0.015276594998340443, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3779.3735550003476, "method": "dflash", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 65, "repeat": 1, "selection_seconds": 0.0003260159992350964, "selector_pair_scores": 0, "target_forward_passes": 17, "target_seconds": 3.7594197520006674, "tokens_per_second": 6.350258753397504, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 0, "wall_seconds": 3.7793735550003476 }, { "acceptance_rate": 0.1076923076923077, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.0028724659987346968, "draft_forward_passes": 17, "draft_seconds": 0.015472414998839668, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3821.869663000143, "method": "dflash", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 65, "repeat": 2, "selection_seconds": 0.00034182799981863354, "selector_pair_scores": 0, "target_forward_passes": 17, "target_seconds": 3.801779672000066, "tokens_per_second": 6.279648998066604, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 0, "wall_seconds": 3.821869663000143 }, { "acceptance_rate": 0.05333333333333334, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 0, "context_seconds": 0.0033964610006478324, "draft_forward_passes": 20, "draft_seconds": 0.01831294700014041, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4725.0950520001425, "method": "dflash", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 75, "repeat": 0, "selection_seconds": 0.00040009800022744457, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.701290171999517, "tokens_per_second": 5.079262900720008, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "wall_seconds": 4.7250950520001425 }, { "acceptance_rate": 0.05333333333333334, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 0, "context_seconds": 0.003418275000512949, "draft_forward_passes": 20, "draft_seconds": 0.018108597000264126, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4664.7598649997235, "method": "dflash", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 75, "repeat": 1, "selection_seconds": 0.0004147140007262351, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.641194590999476, "tokens_per_second": 5.144959375095597, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "wall_seconds": 4.6647598649997235 }, { "acceptance_rate": 0.05333333333333334, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 0, "context_seconds": 0.0033294350009782647, "draft_forward_passes": 20, "draft_seconds": 0.01786385300101756, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4583.243338000102, "method": "dflash", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 75, "repeat": 2, "selection_seconds": 0.0003845289984383271, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.560111753000001, "tokens_per_second": 5.236466456191348, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "wall_seconds": 4.583243338000102 } ], "dflash2": [ { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.003599499999836553, "draft_forward_passes": 21, "draft_seconds": 0.019924776001516875, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4788.891412999874, "method": "dflash2", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 0, "selection_seconds": 0.009215620999384555, "selector_pair_scores": 4200, "target_forward_passes": 21, "target_seconds": 4.754461625000204, "tokens_per_second": 5.011598286578362, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 4200, "wall_seconds": 4.788891412999874 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.0035661999991134508, "draft_forward_passes": 21, "draft_seconds": 0.01886298799990982, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4776.277298999958, "method": "dflash2", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 1, "selection_seconds": 0.009135477000654646, "selector_pair_scores": 4200, "target_forward_passes": 21, "target_seconds": 4.743015096000363, "tokens_per_second": 5.024833881614254, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 4200, "wall_seconds": 4.776277298999958 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.0035830479996548092, "draft_forward_passes": 21, "draft_seconds": 0.018759764001970325, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4643.639630000052, "method": "dflash2", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 2, "selection_seconds": 0.008955862000675552, "selector_pair_scores": 4200, "target_forward_passes": 21, "target_seconds": 4.6106647110004815, "tokens_per_second": 5.1683597161478545, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 4200, "wall_seconds": 4.643639630000052 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.0032545789977120876, "draft_forward_passes": 19, "draft_seconds": 0.01746657499916182, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4566.820507000102, "method": "dflash2", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 70, "repeat": 0, "selection_seconds": 0.008179894000932109, "selector_pair_scores": 3800, "target_forward_passes": 19, "target_seconds": 4.536395405999883, "tokens_per_second": 5.255297413859902, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 3800, "wall_seconds": 4.566820507000102 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.003277424000316387, "draft_forward_passes": 19, "draft_seconds": 0.017208792999554134, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4667.452643000161, "method": "dflash2", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 70, "repeat": 1, "selection_seconds": 0.008199933000014425, "selector_pair_scores": 3800, "target_forward_passes": 19, "target_seconds": 4.637278519000574, "tokens_per_second": 5.14199111071713, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 3800, "wall_seconds": 4.667452643000161 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.0035191469992241764, "draft_forward_passes": 19, "draft_seconds": 0.017233912999472523, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4637.845341999764, "method": "dflash2", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 70, "repeat": 2, "selection_seconds": 0.008174998999493255, "selector_pair_scores": 3800, "target_forward_passes": 19, "target_seconds": 4.607455444999232, "tokens_per_second": 5.17481680181504, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 3800, "wall_seconds": 4.637845341999764 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.003333401000872982, "draft_forward_passes": 19, "draft_seconds": 0.01693470700047328, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4212.682857000345, "method": "dflash2", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 0, "selection_seconds": 0.008076135999999678, "selector_pair_scores": 3800, "target_forward_passes": 19, "target_seconds": 4.1828635609999765, "tokens_per_second": 5.697082076833403, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 3800, "wall_seconds": 4.212682857000345 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.003291131000878522, "draft_forward_passes": 19, "draft_seconds": 0.01751842000066972, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4314.052203999836, "method": "dflash2", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 1, "selection_seconds": 0.008249048999459774, "selector_pair_scores": 3800, "target_forward_passes": 19, "target_seconds": 4.283455074000358, "tokens_per_second": 5.56321501574507, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 3800, "wall_seconds": 4.314052203999836 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.003206896001302084, "draft_forward_passes": 19, "draft_seconds": 0.017075324000416003, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4207.243810999898, "method": "dflash2", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 2, "selection_seconds": 0.008154596000167658, "selector_pair_scores": 3800, "target_forward_passes": 19, "target_seconds": 4.177345935000176, "tokens_per_second": 5.704447157840404, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 3800, "wall_seconds": 4.207243810999898 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.003458284001226275, "draft_forward_passes": 21, "draft_seconds": 0.018605932999889774, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4652.327265999702, "method": "dflash2", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 78, "repeat": 0, "selection_seconds": 0.008942859002672776, "selector_pair_scores": 4200, "target_forward_passes": 21, "target_seconds": 4.6196949830005, "tokens_per_second": 5.158708454454102, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 4200, "wall_seconds": 4.652327265999702 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.003505858000153239, "draft_forward_passes": 21, "draft_seconds": 0.018773441000575986, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4724.476886999582, "method": "dflash2", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 78, "repeat": 1, "selection_seconds": 0.008835133998218225, "selector_pair_scores": 4200, "target_forward_passes": 21, "target_seconds": 4.691673082999387, "tokens_per_second": 5.079927487007329, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 4200, "wall_seconds": 4.724476886999582 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.0035349009990568447, "draft_forward_passes": 21, "draft_seconds": 0.01869024299958255, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4694.671154999924, "method": "dflash2", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 78, "repeat": 2, "selection_seconds": 0.009187224000015703, "selector_pair_scores": 4200, "target_forward_passes": 21, "target_seconds": 4.661590089998754, "tokens_per_second": 5.112179151129573, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 4200, "wall_seconds": 4.694671154999924 }, { "acceptance_rate": 0.1076923076923077, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.0028176689993415494, "draft_forward_passes": 17, "draft_seconds": 0.01532364500053518, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3733.849267000096, "method": "dflash2", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 65, "repeat": 0, "selection_seconds": 0.007392045999949914, "selector_pair_scores": 3400, "target_forward_passes": 17, "target_seconds": 3.7069269170010557, "tokens_per_second": 6.427683145142716, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 3400, "wall_seconds": 3.733849267000096 }, { "acceptance_rate": 0.1076923076923077, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.002912114999617188, "draft_forward_passes": 17, "draft_seconds": 0.01541474999885395, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3843.4257829999297, "method": "dflash2", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 65, "repeat": 1, "selection_seconds": 0.007383341000149812, "selector_pair_scores": 3400, "target_forward_passes": 17, "target_seconds": 3.816317822998826, "tokens_per_second": 6.244429151242034, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 3400, "wall_seconds": 3.8434257829999297 }, { "acceptance_rate": 0.1076923076923077, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.0030049449983380327, "draft_forward_passes": 17, "draft_seconds": 0.016235316000347666, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3960.271202000058, "method": "dflash2", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 65, "repeat": 2, "selection_seconds": 0.007420537000143668, "selector_pair_scores": 3400, "target_forward_passes": 17, "target_seconds": 3.9322396599995955, "tokens_per_second": 6.0601910262810454, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 3400, "wall_seconds": 3.960271202000058 }, { "acceptance_rate": 0.07042253521126761, "accepted_draft_tokens": 5, "boltzmann_candidate_scores": 0, "context_seconds": 0.003294961000847252, "draft_forward_passes": 19, "draft_seconds": 0.017455508999319136, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4595.124258999931, "method": "dflash2", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 71, "repeat": 0, "selection_seconds": 0.008266796999578219, "selector_pair_scores": 3800, "target_forward_passes": 19, "target_seconds": 4.564579294999021, "tokens_per_second": 5.222927313226409, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 3800, "wall_seconds": 4.595124258999931 }, { "acceptance_rate": 0.07042253521126761, "accepted_draft_tokens": 5, "boltzmann_candidate_scores": 0, "context_seconds": 0.003199823000613833, "draft_forward_passes": 19, "draft_seconds": 0.017401150999376114, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4492.436810000072, "method": "dflash2", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 71, "repeat": 1, "selection_seconds": 0.008091088998753548, "selector_pair_scores": 3800, "target_forward_passes": 19, "target_seconds": 4.462219613998968, "tokens_per_second": 5.342312204943315, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 3800, "wall_seconds": 4.492436810000072 }, { "acceptance_rate": 0.07042253521126761, "accepted_draft_tokens": 5, "boltzmann_candidate_scores": 0, "context_seconds": 0.0031458230000680487, "draft_forward_passes": 19, "draft_seconds": 0.01711577300056888, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4485.95859399984, "method": "dflash2", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 71, "repeat": 2, "selection_seconds": 0.008033400000385882, "selector_pair_scores": 3800, "target_forward_passes": 19, "target_seconds": 4.456243481000456, "tokens_per_second": 5.350027089438814, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 3800, "wall_seconds": 4.48595859399984 } ], "dflash3_mobs": [ { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.003662822998194315, "draft_forward_passes": 21, "draft_seconds": 0.01970472300035908, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4809.769395999865, "method": "dflash3_mobs", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 0, "selection_seconds": 0.008557712999845535, "selector_pair_scores": 840, "target_forward_passes": 21, "target_seconds": 4.77606532100026, "tokens_per_second": 4.989844215807945, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 840, "wall_seconds": 4.809769395999865 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.003550798999640392, "draft_forward_passes": 21, "draft_seconds": 0.01853580399938437, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4623.655336999946, "method": "dflash3_mobs", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 1, "selection_seconds": 0.008388808999825415, "selector_pair_scores": 840, "target_forward_passes": 21, "target_seconds": 4.591550896001991, "tokens_per_second": 5.190698322157459, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 840, "wall_seconds": 4.623655336999946 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.00355469499982064, "draft_forward_passes": 21, "draft_seconds": 0.01884416000029887, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4645.2781729999515, "method": "dflash3_mobs", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 2, "selection_seconds": 0.008389171001454088, "selector_pair_scores": 840, "target_forward_passes": 21, "target_seconds": 4.612882934999561, "tokens_per_second": 5.1665366650154, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 840, "wall_seconds": 4.6452781729999515 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.0033429309996790835, "draft_forward_passes": 19, "draft_seconds": 0.017633241996918514, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4659.089458999915, "method": "dflash3_mobs", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 70, "repeat": 0, "selection_seconds": 0.007738281999081664, "selector_pair_scores": 760, "target_forward_passes": 19, "target_seconds": 4.628865925999435, "tokens_per_second": 5.15122111545625, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 760, "wall_seconds": 4.659089458999915 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.003337039000598452, "draft_forward_passes": 19, "draft_seconds": 0.017225598999175418, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4624.171055000261, "method": "dflash3_mobs", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 70, "repeat": 1, "selection_seconds": 0.007591570999466057, "selector_pair_scores": 760, "target_forward_passes": 19, "target_seconds": 4.594489772000088, "tokens_per_second": 5.190119421306452, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 760, "wall_seconds": 4.624171055000261 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.0032712590004848607, "draft_forward_passes": 19, "draft_seconds": 0.017014620999816543, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4589.945123000234, "method": "dflash3_mobs", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 70, "repeat": 2, "selection_seconds": 0.007415704998948058, "selector_pair_scores": 760, "target_forward_passes": 19, "target_seconds": 4.560787632000483, "tokens_per_second": 5.228820684529734, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 760, "wall_seconds": 4.589945123000234 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.003418908000639931, "draft_forward_passes": 19, "draft_seconds": 0.017680904001281306, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4273.944012999891, "method": "dflash3_mobs", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 0, "selection_seconds": 0.007715046999692277, "selector_pair_scores": 760, "target_forward_passes": 19, "target_seconds": 4.24362062200089, "tokens_per_second": 5.615422178437556, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 760, "wall_seconds": 4.273944012999891 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.003203035999831627, "draft_forward_passes": 19, "draft_seconds": 0.01726292200100943, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4257.230063999941, "method": "dflash3_mobs", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 1, "selection_seconds": 0.007453364999491896, "selector_pair_scores": 760, "target_forward_passes": 19, "target_seconds": 4.22779834199946, "tokens_per_second": 5.637468410023033, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 760, "wall_seconds": 4.257230063999941 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.0031757859992467274, "draft_forward_passes": 19, "draft_seconds": 0.01686848400049712, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4165.632274000018, "method": "dflash3_mobs", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 2, "selection_seconds": 0.007356732999141968, "selector_pair_scores": 760, "target_forward_passes": 19, "target_seconds": 4.136642709999705, "tokens_per_second": 5.761430299500291, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 760, "wall_seconds": 4.165632274000018 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.0034626420001586666, "draft_forward_passes": 21, "draft_seconds": 0.01858312800004569, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4652.944337000008, "method": "dflash3_mobs", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 78, "repeat": 0, "selection_seconds": 0.008253809999587247, "selector_pair_scores": 840, "target_forward_passes": 21, "target_seconds": 4.621032825999919, "tokens_per_second": 5.158024309285856, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 840, "wall_seconds": 4.652944337000008 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.003402075999929366, "draft_forward_passes": 21, "draft_seconds": 0.018571262999103055, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4630.809670999952, "method": "dflash3_mobs", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 78, "repeat": 1, "selection_seconds": 0.008202217999951245, "selector_pair_scores": 840, "target_forward_passes": 21, "target_seconds": 4.599015994999263, "tokens_per_second": 5.182678992466035, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 840, "wall_seconds": 4.630809670999952 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.0035892630016860494, "draft_forward_passes": 21, "draft_seconds": 0.018738508000751608, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4701.781981999829, "method": "dflash3_mobs", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 78, "repeat": 2, "selection_seconds": 0.008161047999692528, "selector_pair_scores": 840, "target_forward_passes": 21, "target_seconds": 4.669636259999152, "tokens_per_second": 5.104447652375403, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 840, "wall_seconds": 4.701781981999829 }, { "acceptance_rate": 0.1076923076923077, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.0027474070006974216, "draft_forward_passes": 17, "draft_seconds": 0.015168615999755275, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3694.6022530000846, "method": "dflash3_mobs", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 65, "repeat": 0, "selection_seconds": 0.0065612190001047566, "selector_pair_scores": 680, "target_forward_passes": 17, "target_seconds": 3.668777318001048, "tokens_per_second": 6.4959631258037485, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 680, "wall_seconds": 3.6946022530000846 }, { "acceptance_rate": 0.1076923076923077, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.00279628199950821, "draft_forward_passes": 17, "draft_seconds": 0.01534684800026298, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3773.712130000149, "method": "dflash3_mobs", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 65, "repeat": 1, "selection_seconds": 0.006704532001094776, "selector_pair_scores": 680, "target_forward_passes": 17, "target_seconds": 3.7475208769997153, "tokens_per_second": 6.359785583326689, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 680, "wall_seconds": 3.773712130000149 }, { "acceptance_rate": 0.1076923076923077, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.0031271439993361128, "draft_forward_passes": 17, "draft_seconds": 0.01628151400018396, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4054.4187460000103, "method": "dflash3_mobs", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 65, "repeat": 2, "selection_seconds": 0.006885210999371338, "selector_pair_scores": 680, "target_forward_passes": 17, "target_seconds": 4.026657120999516, "tokens_per_second": 5.919467500409969, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 680, "wall_seconds": 4.05441874600001 }, { "acceptance_rate": 0.07042253521126761, "accepted_draft_tokens": 5, "boltzmann_candidate_scores": 0, "context_seconds": 0.0033966660002988647, "draft_forward_passes": 19, "draft_seconds": 0.017206934000114416, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4477.279924999948, "method": "dflash3_mobs", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 71, "repeat": 0, "selection_seconds": 0.007537294000485417, "selector_pair_scores": 760, "target_forward_passes": 19, "target_seconds": 4.447611894000147, "tokens_per_second": 5.360397473919453, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 760, "wall_seconds": 4.477279924999948 }, { "acceptance_rate": 0.07042253521126761, "accepted_draft_tokens": 5, "boltzmann_candidate_scores": 0, "context_seconds": 0.003201967999302724, "draft_forward_passes": 19, "draft_seconds": 0.017140686001312133, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4447.732818000077, "method": "dflash3_mobs", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 71, "repeat": 1, "selection_seconds": 0.0074705409997477545, "selector_pair_scores": 760, "target_forward_passes": 19, "target_seconds": 4.418473809999796, "tokens_per_second": 5.396007580057743, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 760, "wall_seconds": 4.447732818000077 }, { "acceptance_rate": 0.07042253521126761, "accepted_draft_tokens": 5, "boltzmann_candidate_scores": 0, "context_seconds": 0.0032208220004577015, "draft_forward_passes": 19, "draft_seconds": 0.016936252000050445, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4406.857437000326, "method": "dflash3_mobs", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 71, "repeat": 2, "selection_seconds": 0.007395109998924454, "selector_pair_scores": 760, "target_forward_passes": 19, "target_seconds": 4.377866062999601, "tokens_per_second": 5.446057727779912, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 760, "wall_seconds": 4.406857437000326 } ], "dflash4_jump_mobs": [ { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.003711468998062628, "draft_forward_passes": 21, "draft_seconds": 0.01984358800064001, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 336, "jump_forward_passes": 21, "latency_ms": 4871.631783999874, "method": "dflash4_jump_mobs", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 0, "selection_seconds": 0.014737945998604118, "selector_pair_scores": 672, "target_forward_passes": 21, "target_seconds": 4.831610546999855, "tokens_per_second": 4.926480707927129, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 1008, "wall_seconds": 4.871631783999874 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.0034953250001308334, "draft_forward_passes": 21, "draft_seconds": 0.018473877000360517, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 336, "jump_forward_passes": 21, "latency_ms": 4618.050266999944, "method": "dflash4_jump_mobs", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 1, "selection_seconds": 0.014430892998461786, "selector_pair_scores": 672, "target_forward_passes": 21, "target_seconds": 4.580018557999665, "tokens_per_second": 5.196998432758785, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 1008, "wall_seconds": 4.618050266999944 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.0036644830011027807, "draft_forward_passes": 21, "draft_seconds": 0.018762500998946052, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 336, "jump_forward_passes": 21, "latency_ms": 4633.773053999903, "method": "dflash4_jump_mobs", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 2, "selection_seconds": 0.014329417000681133, "selector_pair_scores": 672, "target_forward_passes": 21, "target_seconds": 4.595383678000871, "tokens_per_second": 5.179364574033906, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 1008, "wall_seconds": 4.633773053999903 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.0033181840003635443, "draft_forward_passes": 19, "draft_seconds": 0.017429312000786013, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 304, "jump_forward_passes": 19, "latency_ms": 4663.153352000336, "method": "dflash4_jump_mobs", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 70, "repeat": 0, "selection_seconds": 0.013177387000723684, "selector_pair_scores": 608, "target_forward_passes": 19, "target_seconds": 4.627688690000014, "tokens_per_second": 5.146731876125156, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 912, "wall_seconds": 4.663153352000336 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.0033946450007533713, "draft_forward_passes": 19, "draft_seconds": 0.01706428099896584, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 304, "jump_forward_passes": 19, "latency_ms": 4566.900073999932, "method": "dflash4_jump_mobs", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 70, "repeat": 1, "selection_seconds": 0.0130737920003412, "selector_pair_scores": 608, "target_forward_passes": 19, "target_seconds": 4.531825059001221, "tokens_per_second": 5.255205853229789, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 912, "wall_seconds": 4.566900073999932 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.0032200019991250883, "draft_forward_passes": 19, "draft_seconds": 0.01667534200123555, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 304, "jump_forward_passes": 19, "latency_ms": 4528.739760000008, "method": "dflash4_jump_mobs", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 70, "repeat": 2, "selection_seconds": 0.012652374000481359, "selector_pair_scores": 608, "target_forward_passes": 19, "target_seconds": 4.494649744999151, "tokens_per_second": 5.299487555451842, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 912, "wall_seconds": 4.528739760000008 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.0032774149995020707, "draft_forward_passes": 19, "draft_seconds": 0.01726004300007844, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 304, "jump_forward_passes": 19, "latency_ms": 4253.115918000276, "method": "dflash4_jump_mobs", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 0, "selection_seconds": 0.013088740000057442, "selector_pair_scores": 608, "target_forward_passes": 19, "target_seconds": 4.217960514000424, "tokens_per_second": 5.642921675006753, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 912, "wall_seconds": 4.253115918000276 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.0033789950011851033, "draft_forward_passes": 19, "draft_seconds": 0.01750574300012886, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 304, "jump_forward_passes": 19, "latency_ms": 4318.955789000029, "method": "dflash4_jump_mobs", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 1, "selection_seconds": 0.013216097000167792, "selector_pair_scores": 608, "target_forward_passes": 19, "target_seconds": 4.283371849000105, "tokens_per_second": 5.556898744165367, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 912, "wall_seconds": 4.318955789000029 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.0031271169991669012, "draft_forward_passes": 19, "draft_seconds": 0.016976801000055275, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 304, "jump_forward_passes": 19, "latency_ms": 4174.364743999831, "method": "dflash4_jump_mobs", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 2, "selection_seconds": 0.012840733998928044, "selector_pair_scores": 608, "target_forward_passes": 19, "target_seconds": 4.139930536999145, "tokens_per_second": 5.749377802813527, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 912, "wall_seconds": 4.174364743999831 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.003531933999965986, "draft_forward_passes": 21, "draft_seconds": 0.01836391599908893, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 336, "jump_forward_passes": 21, "latency_ms": 4644.159559999935, "method": "dflash4_jump_mobs", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 78, "repeat": 0, "selection_seconds": 0.01409790499974406, "selector_pair_scores": 672, "target_forward_passes": 21, "target_seconds": 4.606589606000853, "tokens_per_second": 5.16778110009647, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 1008, "wall_seconds": 4.644159559999935 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.003537124000104086, "draft_forward_passes": 21, "draft_seconds": 0.018486063000182185, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 336, "jump_forward_passes": 21, "latency_ms": 4657.350189000226, "method": "dflash4_jump_mobs", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 78, "repeat": 1, "selection_seconds": 0.014161320000312116, "selector_pair_scores": 672, "target_forward_passes": 21, "target_seconds": 4.619503561998954, "tokens_per_second": 5.153144819705297, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 1008, "wall_seconds": 4.657350189000226 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.0034828600005312182, "draft_forward_passes": 21, "draft_seconds": 0.018753545000436134, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 336, "jump_forward_passes": 21, "latency_ms": 4709.880899000382, "method": "dflash4_jump_mobs", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 78, "repeat": 2, "selection_seconds": 0.014276026000970887, "selector_pair_scores": 672, "target_forward_passes": 21, "target_seconds": 4.671721127999717, "tokens_per_second": 5.095670254654151, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 1008, "wall_seconds": 4.709880899000382 }, { "acceptance_rate": 0.0684931506849315, "accepted_draft_tokens": 5, "boltzmann_candidate_scores": 0, "context_seconds": 0.003230568999697425, "draft_forward_passes": 19, "draft_seconds": 0.016618174998711766, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 304, "jump_forward_passes": 19, "latency_ms": 4130.004409999856, "method": "dflash4_jump_mobs", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 73, "repeat": 0, "selection_seconds": 0.0127491119988008, "selector_pair_scores": 608, "target_forward_passes": 19, "target_seconds": 4.095933768999657, "tokens_per_second": 5.811131809421201, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 912, "wall_seconds": 4.130004409999856 }, { "acceptance_rate": 0.0684931506849315, "accepted_draft_tokens": 5, "boltzmann_candidate_scores": 0, "context_seconds": 0.0031880099991212774, "draft_forward_passes": 19, "draft_seconds": 0.017412697999589, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 304, "jump_forward_passes": 19, "latency_ms": 4295.686242000102, "method": "dflash4_jump_mobs", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 73, "repeat": 1, "selection_seconds": 0.012960130999090325, "selector_pair_scores": 608, "target_forward_passes": 19, "target_seconds": 4.2606350330006535, "tokens_per_second": 5.587000224864056, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 912, "wall_seconds": 4.295686242000102 }, { "acceptance_rate": 0.0684931506849315, "accepted_draft_tokens": 5, "boltzmann_candidate_scores": 0, "context_seconds": 0.003164369999922201, "draft_forward_passes": 19, "draft_seconds": 0.017042032000517793, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 304, "jump_forward_passes": 19, "latency_ms": 4194.796573999611, "method": "dflash4_jump_mobs", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 73, "repeat": 2, "selection_seconds": 0.012778924000485858, "selector_pair_scores": 608, "target_forward_passes": 19, "target_seconds": 4.16034112700072, "tokens_per_second": 5.721373987181631, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 912, "wall_seconds": 4.194796573999611 }, { "acceptance_rate": 0.07042253521126761, "accepted_draft_tokens": 5, "boltzmann_candidate_scores": 0, "context_seconds": 0.003192011000464845, "draft_forward_passes": 19, "draft_seconds": 0.01694742099925861, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 304, "jump_forward_passes": 19, "latency_ms": 4435.671501999877, "method": "dflash4_jump_mobs", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 71, "repeat": 0, "selection_seconds": 0.012934350999785238, "selector_pair_scores": 608, "target_forward_passes": 19, "target_seconds": 4.401125485001103, "tokens_per_second": 5.4106802068591655, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 912, "wall_seconds": 4.435671501999877 }, { "acceptance_rate": 0.07042253521126761, "accepted_draft_tokens": 5, "boltzmann_candidate_scores": 0, "context_seconds": 0.0032694850001462328, "draft_forward_passes": 19, "draft_seconds": 0.016843240000980586, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 304, "jump_forward_passes": 19, "latency_ms": 4481.38975799975, "method": "dflash4_jump_mobs", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 71, "repeat": 1, "selection_seconds": 0.012790373000825639, "selector_pair_scores": 608, "target_forward_passes": 19, "target_seconds": 4.446997674000158, "tokens_per_second": 5.355481512661889, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 912, "wall_seconds": 4.48138975799975 }, { "acceptance_rate": 0.07042253521126761, "accepted_draft_tokens": 5, "boltzmann_candidate_scores": 0, "context_seconds": 0.003248145000270597, "draft_forward_passes": 19, "draft_seconds": 0.01698046500041528, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 304, "jump_forward_passes": 19, "latency_ms": 4391.5285970001605, "method": "dflash4_jump_mobs", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 71, "repeat": 2, "selection_seconds": 0.012836963999234285, "selector_pair_scores": 608, "target_forward_passes": 19, "target_seconds": 4.357007473001431, "tokens_per_second": 5.465067452001639, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 912, "wall_seconds": 4.3915285970001605 } ], "dflash5_fused_jump_mobs": [ { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.003599599998779013, "draft_forward_passes": 21, "draft_seconds": 0.01955417699946338, "exact_match": true, "fused_anchor_uses": 42, "jump_candidate_scores": 336, "jump_forward_passes": 0, "latency_ms": 4826.176753000254, "method": "dflash5_fused_jump_mobs", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 0, "selection_seconds": 0.014639651999459602, "selector_pair_scores": 672, "target_forward_passes": 21, "target_seconds": 4.786670343998594, "tokens_per_second": 4.9728804451847095, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 1008, "wall_seconds": 4.826176753000254 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.0036531630003082682, "draft_forward_passes": 21, "draft_seconds": 0.019411714000852953, "exact_match": true, "fused_anchor_uses": 42, "jump_candidate_scores": 336, "jump_forward_passes": 0, "latency_ms": 4712.572025999634, "method": "dflash5_fused_jump_mobs", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 1, "selection_seconds": 0.014875256999403064, "selector_pair_scores": 672, "target_forward_passes": 21, "target_seconds": 4.672981862000142, "tokens_per_second": 5.092760358375446, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 1008, "wall_seconds": 4.712572025999634 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.0035053040001002955, "draft_forward_passes": 21, "draft_seconds": 0.018848293000246485, "exact_match": true, "fused_anchor_uses": 42, "jump_candidate_scores": 336, "jump_forward_passes": 0, "latency_ms": 4612.094068999795, "method": "dflash5_fused_jump_mobs", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 2, "selection_seconds": 0.014203007000105572, "selector_pair_scores": 672, "target_forward_passes": 21, "target_seconds": 4.573918395000874, "tokens_per_second": 5.20370999397347, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 1008, "wall_seconds": 4.612094068999795 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.0034368759993412823, "draft_forward_passes": 19, "draft_seconds": 0.01738043000068501, "exact_match": true, "fused_anchor_uses": 38, "jump_candidate_scores": 304, "jump_forward_passes": 0, "latency_ms": 4578.814485999828, "method": "dflash5_fused_jump_mobs", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 70, "repeat": 0, "selection_seconds": 0.013053405002210638, "selector_pair_scores": 608, "target_forward_passes": 19, "target_seconds": 4.543329369001185, "tokens_per_second": 5.241531421153301, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 912, "wall_seconds": 4.578814485999828 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.0034520729996074806, "draft_forward_passes": 19, "draft_seconds": 0.01754824299905522, "exact_match": true, "fused_anchor_uses": 38, "jump_candidate_scores": 304, "jump_forward_passes": 0, "latency_ms": 4584.062685999925, "method": "dflash5_fused_jump_mobs", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 70, "repeat": 1, "selection_seconds": 0.01327832200058765, "selector_pair_scores": 608, "target_forward_passes": 19, "target_seconds": 4.548295118999249, "tokens_per_second": 5.235530498589782, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 912, "wall_seconds": 4.584062685999925 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.003333530001782492, "draft_forward_passes": 19, "draft_seconds": 0.016791472998647805, "exact_match": true, "fused_anchor_uses": 38, "jump_candidate_scores": 304, "jump_forward_passes": 0, "latency_ms": 4511.2874650003505, "method": "dflash5_fused_jump_mobs", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 70, "repeat": 2, "selection_seconds": 0.012621179000689153, "selector_pair_scores": 608, "target_forward_passes": 19, "target_seconds": 4.477084828999523, "tokens_per_second": 5.319989068796381, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 912, "wall_seconds": 4.5112874650003505 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.003357585000230756, "draft_forward_passes": 19, "draft_seconds": 0.01718380999955116, "exact_match": true, "fused_anchor_uses": 38, "jump_candidate_scores": 304, "jump_forward_passes": 0, "latency_ms": 4261.909589999959, "method": "dflash5_fused_jump_mobs", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 0, "selection_seconds": 0.013246673999674385, "selector_pair_scores": 608, "target_forward_passes": 19, "target_seconds": 4.226653066999916, "tokens_per_second": 5.631278536811999, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 912, "wall_seconds": 4.261909589999959 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.003267177999987325, "draft_forward_passes": 19, "draft_seconds": 0.017529253000702738, "exact_match": true, "fused_anchor_uses": 38, "jump_candidate_scores": 304, "jump_forward_passes": 0, "latency_ms": 4344.387410000309, "method": "dflash5_fused_jump_mobs", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 1, "selection_seconds": 0.013305885000590933, "selector_pair_scores": 608, "target_forward_passes": 19, "target_seconds": 4.308795350998935, "tokens_per_second": 5.524369199844978, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 912, "wall_seconds": 4.344387410000309 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.0031298369999603892, "draft_forward_passes": 19, "draft_seconds": 0.016817931000787212, "exact_match": true, "fused_anchor_uses": 38, "jump_candidate_scores": 304, "jump_forward_passes": 0, "latency_ms": 4135.87792699991, "method": "dflash5_fused_jump_mobs", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 2, "selection_seconds": 0.012748141999509244, "selector_pair_scores": 608, "target_forward_passes": 19, "target_seconds": 4.101683016999232, "tokens_per_second": 5.802879200888107, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 912, "wall_seconds": 4.13587792699991 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.003502364999803831, "draft_forward_passes": 21, "draft_seconds": 0.018348866001360875, "exact_match": true, "fused_anchor_uses": 42, "jump_candidate_scores": 336, "jump_forward_passes": 0, "latency_ms": 4636.736986999949, "method": "dflash5_fused_jump_mobs", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 78, "repeat": 0, "selection_seconds": 0.01405156399914631, "selector_pair_scores": 672, "target_forward_passes": 21, "target_seconds": 4.5991873860011765, "tokens_per_second": 5.17605377818258, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 1008, "wall_seconds": 4.636736986999949 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.003501175001019874, "draft_forward_passes": 21, "draft_seconds": 0.018782161999752134, "exact_match": true, "fused_anchor_uses": 42, "jump_candidate_scores": 336, "jump_forward_passes": 0, "latency_ms": 4790.287486000125, "method": "dflash5_fused_jump_mobs", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 78, "repeat": 1, "selection_seconds": 0.014127752999684162, "selector_pair_scores": 672, "target_forward_passes": 21, "target_seconds": 4.752215822998551, "tokens_per_second": 5.010137715145762, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 1008, "wall_seconds": 4.790287486000125 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.0034503900001254806, "draft_forward_passes": 21, "draft_seconds": 0.01873935999947207, "exact_match": true, "fused_anchor_uses": 42, "jump_candidate_scores": 336, "jump_forward_passes": 0, "latency_ms": 4685.192116000053, "method": "dflash5_fused_jump_mobs", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 78, "repeat": 2, "selection_seconds": 0.014157477998651302, "selector_pair_scores": 672, "target_forward_passes": 21, "target_seconds": 4.6472154649986805, "tokens_per_second": 5.122522066499552, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 1008, "wall_seconds": 4.685192116000053 }, { "acceptance_rate": 0.1076923076923077, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.0027891320009985066, "draft_forward_passes": 17, "draft_seconds": 0.015192413000022498, "exact_match": true, "fused_anchor_uses": 34, "jump_candidate_scores": 272, "jump_forward_passes": 0, "latency_ms": 3694.986993999919, "method": "dflash5_fused_jump_mobs", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 65, "repeat": 0, "selection_seconds": 0.011430224000378075, "selector_pair_scores": 544, "target_forward_passes": 17, "target_seconds": 3.6642148139999335, "tokens_per_second": 6.495286732801021, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 816, "wall_seconds": 3.694986993999919 }, { "acceptance_rate": 0.1076923076923077, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.0027951649994975014, "draft_forward_passes": 17, "draft_seconds": 0.015191482000318501, "exact_match": true, "fused_anchor_uses": 34, "jump_candidate_scores": 272, "jump_forward_passes": 0, "latency_ms": 3819.635309000205, "method": "dflash5_fused_jump_mobs", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 65, "repeat": 1, "selection_seconds": 0.011557189000996004, "selector_pair_scores": 544, "target_forward_passes": 17, "target_seconds": 3.7887587839991284, "tokens_per_second": 6.28332237463844, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 816, "wall_seconds": 3.819635309000205 }, { "acceptance_rate": 0.1076923076923077, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.002804850999837072, "draft_forward_passes": 17, "draft_seconds": 0.015407691999826056, "exact_match": true, "fused_anchor_uses": 34, "jump_candidate_scores": 272, "jump_forward_passes": 0, "latency_ms": 3810.656703999939, "method": "dflash5_fused_jump_mobs", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 65, "repeat": 2, "selection_seconds": 0.01146162999975786, "selector_pair_scores": 544, "target_forward_passes": 17, "target_seconds": 3.7796607320001385, "tokens_per_second": 6.298127032752091, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 816, "wall_seconds": 3.810656703999939 }, { "acceptance_rate": 0.07042253521126761, "accepted_draft_tokens": 5, "boltzmann_candidate_scores": 0, "context_seconds": 0.003259677000642114, "draft_forward_passes": 19, "draft_seconds": 0.016995198000131495, "exact_match": true, "fused_anchor_uses": 38, "jump_candidate_scores": 304, "jump_forward_passes": 0, "latency_ms": 4459.756576000018, "method": "dflash5_fused_jump_mobs", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 71, "repeat": 0, "selection_seconds": 0.01285423699846433, "selector_pair_scores": 608, "target_forward_passes": 19, "target_seconds": 4.425125676999869, "tokens_per_second": 5.381459635970926, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 912, "wall_seconds": 4.459756576000018 }, { "acceptance_rate": 0.07042253521126761, "accepted_draft_tokens": 5, "boltzmann_candidate_scores": 0, "context_seconds": 0.003227155999866227, "draft_forward_passes": 19, "draft_seconds": 0.01733073999912449, "exact_match": true, "fused_anchor_uses": 38, "jump_candidate_scores": 304, "jump_forward_passes": 0, "latency_ms": 4499.969126999986, "method": "dflash5_fused_jump_mobs", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 71, "repeat": 1, "selection_seconds": 0.012943962000463216, "selector_pair_scores": 608, "target_forward_passes": 19, "target_seconds": 4.464953059999971, "tokens_per_second": 5.3333699238066075, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 912, "wall_seconds": 4.499969126999986 }, { "acceptance_rate": 0.07042253521126761, "accepted_draft_tokens": 5, "boltzmann_candidate_scores": 0, "context_seconds": 0.0033064199997170363, "draft_forward_passes": 19, "draft_seconds": 0.016959837000285916, "exact_match": true, "fused_anchor_uses": 38, "jump_candidate_scores": 304, "jump_forward_passes": 0, "latency_ms": 4452.43925099976, "method": "dflash5_fused_jump_mobs", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 71, "repeat": 2, "selection_seconds": 0.012788494998858368, "selector_pair_scores": 608, "target_forward_passes": 19, "target_seconds": 4.417951636997714, "tokens_per_second": 5.39030375195147, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 912, "wall_seconds": 4.45243925099976 } ], "dflash6_bmobs": [ { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 168, "context_seconds": 0.0036115770008109394, "draft_forward_passes": 21, "draft_seconds": 0.019446715001322445, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4793.154626999694, "method": "dflash6_bmobs", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 0, "selection_seconds": 0.011631787000624172, "selector_pair_scores": 616, "target_forward_passes": 21, "target_seconds": 4.756792017998578, "tokens_per_second": 5.007140780480715, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 784, "wall_seconds": 4.793154626999694 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 168, "context_seconds": 0.0036562679988492164, "draft_forward_passes": 21, "draft_seconds": 0.01899842600187185, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4686.256483999841, "method": "dflash6_bmobs", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 1, "selection_seconds": 0.01136750500108974, "selector_pair_scores": 616, "target_forward_passes": 21, "target_seconds": 4.650578699998732, "tokens_per_second": 5.121358611493535, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 784, "wall_seconds": 4.686256483999841 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 168, "context_seconds": 0.0035319429998708074, "draft_forward_passes": 21, "draft_seconds": 0.019066332000420516, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4677.661791999981, "method": "dflash6_bmobs", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 2, "selection_seconds": 0.011347157000727748, "selector_pair_scores": 616, "target_forward_passes": 21, "target_seconds": 4.642055379999874, "tokens_per_second": 5.130768547877114, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 784, "wall_seconds": 4.677661791999981 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 152, "context_seconds": 0.003279780999491777, "draft_forward_passes": 19, "draft_seconds": 0.01712134799981868, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4564.751458000046, "method": "dflash6_bmobs", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 70, "repeat": 0, "selection_seconds": 0.010298300000158633, "selector_pair_scores": 544, "target_forward_passes": 19, "target_seconds": 4.532609459000469, "tokens_per_second": 5.257679464221063, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 696, "wall_seconds": 4.564751458000046 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 152, "context_seconds": 0.0035294969998176384, "draft_forward_passes": 19, "draft_seconds": 0.017440269998132862, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4604.259846999867, "method": "dflash6_bmobs", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 70, "repeat": 1, "selection_seconds": 0.010563714001364133, "selector_pair_scores": 544, "target_forward_passes": 19, "target_seconds": 4.571238997999899, "tokens_per_second": 5.212564190015989, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 696, "wall_seconds": 4.604259846999867 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 152, "context_seconds": 0.0032489030004398955, "draft_forward_passes": 19, "draft_seconds": 0.01686215700055982, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4528.488336000009, "method": "dflash6_bmobs", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 70, "repeat": 2, "selection_seconds": 0.010381736999534041, "selector_pair_scores": 544, "target_forward_passes": 19, "target_seconds": 4.496504387001096, "tokens_per_second": 5.299781785724787, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 696, "wall_seconds": 4.528488336000009 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 152, "context_seconds": 0.0031783190011083207, "draft_forward_passes": 19, "draft_seconds": 0.016671452000082354, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4153.6644759999035, "method": "dflash6_bmobs", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 0, "selection_seconds": 0.01009890200020891, "selector_pair_scores": 512, "target_forward_passes": 19, "target_seconds": 4.122266726999442, "tokens_per_second": 5.778030492995592, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 664, "wall_seconds": 4.1536644759999035 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 152, "context_seconds": 0.0032357279997086152, "draft_forward_passes": 19, "draft_seconds": 0.017267385998820828, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4288.547992000076, "method": "dflash6_bmobs", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 1, "selection_seconds": 0.010237923999738996, "selector_pair_scores": 512, "target_forward_passes": 19, "target_seconds": 4.256270766999933, "tokens_per_second": 5.596299737060183, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 664, "wall_seconds": 4.288547992000076 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 152, "context_seconds": 0.0032644150001033267, "draft_forward_passes": 19, "draft_seconds": 0.017299036001077184, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4208.932310999899, "method": "dflash6_bmobs", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 2, "selection_seconds": 0.010234855000817333, "selector_pair_scores": 512, "target_forward_passes": 19, "target_seconds": 4.176692565999474, "tokens_per_second": 5.702158701216656, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 664, "wall_seconds": 4.208932310999899 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 168, "context_seconds": 0.0035054300005867844, "draft_forward_passes": 21, "draft_seconds": 0.018787844000598852, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4728.228176999892, "method": "dflash6_bmobs", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 78, "repeat": 0, "selection_seconds": 0.011284301000159758, "selector_pair_scores": 576, "target_forward_passes": 21, "target_seconds": 4.693040329000723, "tokens_per_second": 5.075897165188893, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 744, "wall_seconds": 4.728228176999892 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 168, "context_seconds": 0.003550947999883647, "draft_forward_passes": 21, "draft_seconds": 0.018957106998641393, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4783.3443839999745, "method": "dflash6_bmobs", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 78, "repeat": 1, "selection_seconds": 0.011252261999743496, "selector_pair_scores": 576, "target_forward_passes": 21, "target_seconds": 4.747893628000384, "tokens_per_second": 5.017410011346598, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 744, "wall_seconds": 4.7833443839999745 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 168, "context_seconds": 0.0035689050000655698, "draft_forward_passes": 21, "draft_seconds": 0.01859255399858739, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4731.017194999822, "method": "dflash6_bmobs", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 78, "repeat": 2, "selection_seconds": 0.011266093998983706, "selector_pair_scores": 576, "target_forward_passes": 21, "target_seconds": 4.695955847001187, "tokens_per_second": 5.072904834369536, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 744, "wall_seconds": 4.731017194999822 }, { "acceptance_rate": 0.08695652173913043, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 144, "context_seconds": 0.0029654159989149775, "draft_forward_passes": 18, "draft_seconds": 0.015959506999479345, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3939.581743999952, "method": "dflash6_bmobs", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 69, "repeat": 0, "selection_seconds": 0.009421721998933208, "selector_pair_scores": 504, "target_forward_passes": 18, "target_seconds": 3.9098260870005106, "tokens_per_second": 6.092017264663284, "tokens_per_target_pass": 1.3333333333333333, "total_guidance_scores": 648, "wall_seconds": 3.939581743999952 }, { "acceptance_rate": 0.08695652173913043, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 144, "context_seconds": 0.003003151000029902, "draft_forward_passes": 18, "draft_seconds": 0.016208902999096608, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4031.181980999918, "method": "dflash6_bmobs", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 69, "repeat": 1, "selection_seconds": 0.009647288000451226, "selector_pair_scores": 504, "target_forward_passes": 18, "target_seconds": 4.0009474540020165, "tokens_per_second": 5.953588826582049, "tokens_per_target_pass": 1.3333333333333333, "total_guidance_scores": 648, "wall_seconds": 4.031181980999918 }, { "acceptance_rate": 0.08695652173913043, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 144, "context_seconds": 0.0029019290004725917, "draft_forward_passes": 18, "draft_seconds": 0.016044634000081714, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3974.46941599992, "method": "dflash6_bmobs", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 69, "repeat": 2, "selection_seconds": 0.009431824000785127, "selector_pair_scores": 504, "target_forward_passes": 18, "target_seconds": 3.944650003999868, "tokens_per_second": 6.03854187514535, "tokens_per_target_pass": 1.3333333333333333, "total_guidance_scores": 648, "wall_seconds": 3.97446941599992 }, { "acceptance_rate": 0.05333333333333334, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 160, "context_seconds": 0.0033946560001822945, "draft_forward_passes": 20, "draft_seconds": 0.01813429399908273, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4714.6775680002975, "method": "dflash6_bmobs", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 75, "repeat": 0, "selection_seconds": 0.01041744199983441, "selector_pair_scores": 536, "target_forward_passes": 20, "target_seconds": 4.681167908000134, "tokens_per_second": 5.090485967247058, "tokens_per_target_pass": 1.2, "total_guidance_scores": 696, "wall_seconds": 4.7146775680002975 }, { "acceptance_rate": 0.05333333333333334, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 160, "context_seconds": 0.0034741010008474404, "draft_forward_passes": 20, "draft_seconds": 0.018053089000204636, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4847.5867849997485, "method": "dflash6_bmobs", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 75, "repeat": 1, "selection_seconds": 0.01069806699979381, "selector_pair_scores": 536, "target_forward_passes": 20, "target_seconds": 4.813789776999329, "tokens_per_second": 4.950917036547958, "tokens_per_target_pass": 1.2, "total_guidance_scores": 696, "wall_seconds": 4.8475867849997485 }, { "acceptance_rate": 0.05333333333333334, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 160, "context_seconds": 0.003394407000087085, "draft_forward_passes": 20, "draft_seconds": 0.01828589599881525, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4741.281897000135, "method": "dflash6_bmobs", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 75, "repeat": 2, "selection_seconds": 0.010754848000487982, "selector_pair_scores": 536, "target_forward_passes": 20, "target_seconds": 4.707293167000444, "tokens_per_second": 5.061922180831535, "tokens_per_target_pass": 1.2, "total_guidance_scores": 696, "wall_seconds": 4.741281897000135 } ], "dflash6_boltzmann": [ { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 672, "context_seconds": 0.0036384940012794686, "draft_forward_passes": 21, "draft_seconds": 0.019668179997552215, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4852.541634999852, "method": "dflash6_boltzmann", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 0, "selection_seconds": 0.008896901999833062, "selector_pair_scores": 0, "target_forward_passes": 21, "target_seconds": 4.818611774000146, "tokens_per_second": 4.94586173705251, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 672, "wall_seconds": 4.852541634999852 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 672, "context_seconds": 0.003531031000420626, "draft_forward_passes": 21, "draft_seconds": 0.01892252000152439, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4678.252828000041, "method": "dflash6_boltzmann", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 1, "selection_seconds": 0.008336320001490094, "selector_pair_scores": 0, "target_forward_passes": 21, "target_seconds": 4.645815867999318, "tokens_per_second": 5.1301203424399, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 672, "wall_seconds": 4.678252828000041 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 672, "context_seconds": 0.0035774339994532056, "draft_forward_passes": 21, "draft_seconds": 0.018719180002335634, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4608.139368000138, "method": "dflash6_boltzmann", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 2, "selection_seconds": 0.008090358000117703, "selector_pair_scores": 0, "target_forward_passes": 21, "target_seconds": 4.57611583500011, "tokens_per_second": 5.208175813140745, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 672, "wall_seconds": 4.608139368000138 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 608, "context_seconds": 0.003371805002188921, "draft_forward_passes": 19, "draft_seconds": 0.01740264099998967, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4673.043802000393, "method": "dflash6_boltzmann", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 70, "repeat": 0, "selection_seconds": 0.007706193000103667, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.643015440000454, "tokens_per_second": 5.13583887010139, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 608, "wall_seconds": 4.673043802000393 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 608, "context_seconds": 0.0033424299999751383, "draft_forward_passes": 19, "draft_seconds": 0.017411661998266936, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4586.102383999787, "method": "dflash6_boltzmann", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 70, "repeat": 1, "selection_seconds": 0.0075689530012823525, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.556302590999621, "tokens_per_second": 5.233201963334344, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 608, "wall_seconds": 4.586102383999787 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 608, "context_seconds": 0.0032526719996894826, "draft_forward_passes": 19, "draft_seconds": 0.017021414000737423, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4516.388534999805, "method": "dflash6_boltzmann", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 70, "repeat": 2, "selection_seconds": 0.007548164999661822, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.487112837998666, "tokens_per_second": 5.31398036595207, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 608, "wall_seconds": 4.516388534999805 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 608, "context_seconds": 0.003190579999682086, "draft_forward_passes": 19, "draft_seconds": 0.01656412200145496, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4095.694683000147, "method": "dflash6_boltzmann", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 0, "selection_seconds": 0.007328758998482954, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.067209048999757, "tokens_per_second": 5.859811792030285, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 608, "wall_seconds": 4.095694683000147 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 608, "context_seconds": 0.003214503999970475, "draft_forward_passes": 19, "draft_seconds": 0.017650893999871187, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4345.368437999696, "method": "dflash6_boltzmann", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 1, "selection_seconds": 0.007596998000281019, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.3152878030000466, "tokens_per_second": 5.523121995852652, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 608, "wall_seconds": 4.345368437999696 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 608, "context_seconds": 0.003126894998786156, "draft_forward_passes": 19, "draft_seconds": 0.01683706099902338, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4162.716377999914, "method": "dflash6_boltzmann", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 2, "selection_seconds": 0.007441275999553909, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.1338089510004465, "tokens_per_second": 5.765466061257681, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 608, "wall_seconds": 4.162716377999914 }, { "acceptance_rate": 0.05405405405405406, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 640, "context_seconds": 0.0032906350011217, "draft_forward_passes": 20, "draft_seconds": 0.017794140000660263, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4460.664372999872, "method": "dflash6_boltzmann", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 74, "repeat": 0, "selection_seconds": 0.007789987000251131, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.430206036000072, "tokens_per_second": 5.380364446442222, "tokens_per_target_pass": 1.2, "total_guidance_scores": 640, "wall_seconds": 4.460664372999872 }, { "acceptance_rate": 0.05405405405405406, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 640, "context_seconds": 0.0034008190004897187, "draft_forward_passes": 20, "draft_seconds": 0.01777032700192649, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4579.230268000174, "method": "dflash6_boltzmann", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 74, "repeat": 1, "selection_seconds": 0.007786872999531624, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.54869871399751, "tokens_per_second": 5.241055503959446, "tokens_per_target_pass": 1.2, "total_guidance_scores": 640, "wall_seconds": 4.579230268000174 }, { "acceptance_rate": 0.05405405405405406, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 640, "context_seconds": 0.0033474709994152363, "draft_forward_passes": 20, "draft_seconds": 0.017722198998853855, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4516.302253000049, "method": "dflash6_boltzmann", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 74, "repeat": 2, "selection_seconds": 0.007756238000183657, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.485898227000234, "tokens_per_second": 5.314081887246917, "tokens_per_target_pass": 1.2, "total_guidance_scores": 640, "wall_seconds": 4.516302253000049 }, { "acceptance_rate": 0.15789473684210525, "accepted_draft_tokens": 9, "boltzmann_candidate_scores": 480, "context_seconds": 0.002471590999448381, "draft_forward_passes": 15, "draft_seconds": 0.013107182999647193, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3274.0075500000785, "method": "dflash6_boltzmann", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 57, "repeat": 0, "selection_seconds": 0.005770811998900172, "selector_pair_scores": 0, "target_forward_passes": 15, "target_seconds": 3.2514731619994564, "tokens_per_second": 7.330465685700519, "tokens_per_target_pass": 1.6, "total_guidance_scores": 480, "wall_seconds": 3.2740075500000785 }, { "acceptance_rate": 0.15789473684210525, "accepted_draft_tokens": 9, "boltzmann_candidate_scores": 480, "context_seconds": 0.002481598000485974, "draft_forward_passes": 15, "draft_seconds": 0.013336670000626327, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3351.066074999835, "method": "dflash6_boltzmann", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 57, "repeat": 1, "selection_seconds": 0.0058332190001237905, "selector_pair_scores": 0, "target_forward_passes": 15, "target_seconds": 3.328238400000828, "tokens_per_second": 7.1618999634321385, "tokens_per_target_pass": 1.6, "total_guidance_scores": 480, "wall_seconds": 3.351066074999835 }, { "acceptance_rate": 0.15789473684210525, "accepted_draft_tokens": 9, "boltzmann_candidate_scores": 480, "context_seconds": 0.002599970000119356, "draft_forward_passes": 15, "draft_seconds": 0.013605289001134224, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3317.6240439997855, "method": "dflash6_boltzmann", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 57, "repeat": 2, "selection_seconds": 0.005906517000312306, "selector_pair_scores": 0, "target_forward_passes": 15, "target_seconds": 3.2943059990011534, "tokens_per_second": 7.234092736760245, "tokens_per_target_pass": 1.6, "total_guidance_scores": 480, "wall_seconds": 3.3176240439997855 }, { "acceptance_rate": 0.05333333333333334, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 640, "context_seconds": 0.0034179429990217614, "draft_forward_passes": 20, "draft_seconds": 0.018070189999889408, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4695.919505999882, "method": "dflash6_boltzmann", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 75, "repeat": 0, "selection_seconds": 0.007810636000158411, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.665044559999842, "tokens_per_second": 5.110820142750676, "tokens_per_target_pass": 1.2, "total_guidance_scores": 640, "wall_seconds": 4.695919505999882 }, { "acceptance_rate": 0.05333333333333334, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 640, "context_seconds": 0.00330818000020372, "draft_forward_passes": 20, "draft_seconds": 0.017640373000176623, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4627.270667000175, "method": "dflash6_boltzmann", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 75, "repeat": 1, "selection_seconds": 0.007751370999812934, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.597000276999097, "tokens_per_second": 5.1866427808423445, "tokens_per_target_pass": 1.2, "total_guidance_scores": 640, "wall_seconds": 4.627270667000175 }, { "acceptance_rate": 0.05333333333333334, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 640, "context_seconds": 0.0034134410002479854, "draft_forward_passes": 20, "draft_seconds": 0.018051165999622754, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4758.297206000407, "method": "dflash6_boltzmann", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 75, "repeat": 2, "selection_seconds": 0.007825616001809976, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.727433657998972, "tokens_per_second": 5.0438211320081106, "tokens_per_target_pass": 1.2, "total_guidance_scores": 640, "wall_seconds": 4.758297206000407 } ], "dflash7_act": [ { "acceptance_rate": 0.12, "accepted_draft_tokens": 3, "act_margin_threshold": 1.0, "act_mean_margin": 1.021249844477727, "act_mean_verified_drafts_per_block": 1.1904761904761905, "act_raw_draft_tokens": 78, "act_trimmed_draft_tokens": 53, "act_verified_draft_tokens": 25, "boltzmann_candidate_scores": 0, "context_seconds": 0.003545577999375382, "draft_forward_passes": 21, "draft_seconds": 0.019284287999653316, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4304.748320000272, "method": "dflash7_act", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 25, "repeat": 0, "selection_seconds": 0.00255196800117119, "selector_pair_scores": 0, "target_forward_passes": 21, "target_seconds": 4.277710139000192, "tokens_per_second": 5.575238833009983, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 0, "wall_seconds": 4.304748320000272 }, { "acceptance_rate": 0.12, "accepted_draft_tokens": 3, "act_margin_threshold": 1.0, "act_mean_margin": 1.021249844477727, "act_mean_verified_drafts_per_block": 1.1904761904761905, "act_raw_draft_tokens": 78, "act_trimmed_draft_tokens": 53, "act_verified_draft_tokens": 25, "boltzmann_candidate_scores": 0, "context_seconds": 0.0036073150013180566, "draft_forward_passes": 21, "draft_seconds": 0.019343585000569874, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4205.0470560002395, "method": "dflash7_act", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 25, "repeat": 1, "selection_seconds": 0.0025206880000041565, "selector_pair_scores": 0, "target_forward_passes": 21, "target_seconds": 4.1779857749993425, "tokens_per_second": 5.707427213151888, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 0, "wall_seconds": 4.2050470560002395 }, { "acceptance_rate": 0.12, "accepted_draft_tokens": 3, "act_margin_threshold": 1.0, "act_mean_margin": 1.021249844477727, "act_mean_verified_drafts_per_block": 1.1904761904761905, "act_raw_draft_tokens": 78, "act_trimmed_draft_tokens": 53, "act_verified_draft_tokens": 25, "boltzmann_candidate_scores": 0, "context_seconds": 0.0038510539998242166, "draft_forward_passes": 21, "draft_seconds": 0.020555853999667306, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4235.2366329996585, "method": "dflash7_act", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 25, "repeat": 2, "selection_seconds": 0.00268107700139808, "selector_pair_scores": 0, "target_forward_passes": 21, "target_seconds": 4.206505060001291, "tokens_per_second": 5.666743580039754, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 0, "wall_seconds": 4.2352366329996585 }, { "acceptance_rate": 0.2916666666666667, "accepted_draft_tokens": 7, "act_margin_threshold": 1.0, "act_mean_margin": 1.0464007926709724, "act_mean_verified_drafts_per_block": 1.3333333333333333, "act_raw_draft_tokens": 66, "act_trimmed_draft_tokens": 42, "act_verified_draft_tokens": 24, "boltzmann_candidate_scores": 0, "context_seconds": 0.0032323859991265635, "draft_forward_passes": 18, "draft_seconds": 0.017027494000103616, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3949.4586370001343, "method": "dflash7_act", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 24, "repeat": 0, "selection_seconds": 0.00225698500071303, "selector_pair_scores": 0, "target_forward_passes": 18, "target_seconds": 3.92552853000052, "tokens_per_second": 6.076782213936422, "tokens_per_target_pass": 1.3333333333333333, "total_guidance_scores": 0, "wall_seconds": 3.9494586370001343 }, { "acceptance_rate": 0.2916666666666667, "accepted_draft_tokens": 7, "act_margin_threshold": 1.0, "act_mean_margin": 1.0464007926709724, "act_mean_verified_drafts_per_block": 1.3333333333333333, "act_raw_draft_tokens": 66, "act_trimmed_draft_tokens": 42, "act_verified_draft_tokens": 24, "boltzmann_candidate_scores": 0, "context_seconds": 0.003183782001997315, "draft_forward_passes": 18, "draft_seconds": 0.017513557000711444, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4058.117591000155, "method": "dflash7_act", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 24, "repeat": 1, "selection_seconds": 0.0023238519997903495, "selector_pair_scores": 0, "target_forward_passes": 18, "target_seconds": 4.0337460749992715, "tokens_per_second": 5.914072094220663, "tokens_per_target_pass": 1.3333333333333333, "total_guidance_scores": 0, "wall_seconds": 4.058117591000155 }, { "acceptance_rate": 0.2916666666666667, "accepted_draft_tokens": 7, "act_margin_threshold": 1.0, "act_mean_margin": 1.0464007926709724, "act_mean_verified_drafts_per_block": 1.3333333333333333, "act_raw_draft_tokens": 66, "act_trimmed_draft_tokens": 42, "act_verified_draft_tokens": 24, "boltzmann_candidate_scores": 0, "context_seconds": 0.003289951000624569, "draft_forward_passes": 18, "draft_seconds": 0.01713372799940771, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3996.447298000021, "method": "dflash7_act", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 24, "repeat": 2, "selection_seconds": 0.0022453549991041655, "selector_pair_scores": 0, "target_forward_passes": 18, "target_seconds": 3.972368223000103, "tokens_per_second": 6.005333790341872, "tokens_per_target_pass": 1.3333333333333333, "total_guidance_scores": 0, "wall_seconds": 3.996447298000021 }, { "acceptance_rate": 0.1724137931034483, "accepted_draft_tokens": 5, "act_margin_threshold": 1.0, "act_mean_margin": 1.3108126794969714, "act_mean_verified_drafts_per_block": 1.45, "act_raw_draft_tokens": 74, "act_trimmed_draft_tokens": 45, "act_verified_draft_tokens": 29, "boltzmann_candidate_scores": 0, "context_seconds": 0.003387179001038021, "draft_forward_passes": 20, "draft_seconds": 0.019076939998740272, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3956.9573599997057, "method": "dflash7_act", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 29, "repeat": 0, "selection_seconds": 0.00247021900031541, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 3.9304679619995113, "tokens_per_second": 6.065266268121167, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "wall_seconds": 3.9569573599997057 }, { "acceptance_rate": 0.1724137931034483, "accepted_draft_tokens": 5, "act_margin_threshold": 1.0, "act_mean_margin": 1.3108126794969714, "act_mean_verified_drafts_per_block": 1.45, "act_raw_draft_tokens": 74, "act_trimmed_draft_tokens": 45, "act_verified_draft_tokens": 29, "boltzmann_candidate_scores": 0, "context_seconds": 0.0035323209995112848, "draft_forward_passes": 20, "draft_seconds": 0.019593087999965064, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4048.7962649999645, "method": "dflash7_act", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 29, "repeat": 1, "selection_seconds": 0.002525754000998859, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.021606549999888, "tokens_per_second": 5.927687744495637, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "wall_seconds": 4.0487962649999645 }, { "acceptance_rate": 0.1724137931034483, "accepted_draft_tokens": 5, "act_margin_threshold": 1.0, "act_mean_margin": 1.3108126794969714, "act_mean_verified_drafts_per_block": 1.45, "act_raw_draft_tokens": 74, "act_trimmed_draft_tokens": 45, "act_verified_draft_tokens": 29, "boltzmann_candidate_scores": 0, "context_seconds": 0.003488340001695178, "draft_forward_passes": 20, "draft_seconds": 0.019625359001111065, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4080.5285619999268, "method": "dflash7_act", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 29, "repeat": 2, "selection_seconds": 0.002584834999197483, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.053295491999961, "tokens_per_second": 5.8815909839477385, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "wall_seconds": 4.080528561999927 }, { "acceptance_rate": 0.15384615384615385, "accepted_draft_tokens": 4, "act_margin_threshold": 1.0, "act_mean_margin": 1.1626239918373726, "act_mean_verified_drafts_per_block": 1.3, "act_raw_draft_tokens": 74, "act_trimmed_draft_tokens": 48, "act_verified_draft_tokens": 26, "boltzmann_candidate_scores": 0, "context_seconds": 0.0036462569987634197, "draft_forward_passes": 20, "draft_seconds": 0.01901365399999122, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4051.0069130000375, "method": "dflash7_act", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 26, "repeat": 0, "selection_seconds": 0.002527958999053226, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.02424859499979, "tokens_per_second": 5.92445298549896, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "wall_seconds": 4.0510069130000375 }, { "acceptance_rate": 0.15384615384615385, "accepted_draft_tokens": 4, "act_margin_threshold": 1.0, "act_mean_margin": 1.1626239918373726, "act_mean_verified_drafts_per_block": 1.3, "act_raw_draft_tokens": 74, "act_trimmed_draft_tokens": 48, "act_verified_draft_tokens": 26, "boltzmann_candidate_scores": 0, "context_seconds": 0.003463237001142261, "draft_forward_passes": 20, "draft_seconds": 0.019174351002220646, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4095.778041000358, "method": "dflash7_act", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 26, "repeat": 1, "selection_seconds": 0.002679605000139418, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.068943525999202, "tokens_per_second": 5.859692532102694, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "wall_seconds": 4.095778041000358 }, { "acceptance_rate": 0.15384615384615385, "accepted_draft_tokens": 4, "act_margin_threshold": 1.0, "act_mean_margin": 1.1626239918373726, "act_mean_verified_drafts_per_block": 1.3, "act_raw_draft_tokens": 74, "act_trimmed_draft_tokens": 48, "act_verified_draft_tokens": 26, "boltzmann_candidate_scores": 0, "context_seconds": 0.0034855429994422593, "draft_forward_passes": 20, "draft_seconds": 0.019062046999351878, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4165.828993999639, "method": "dflash7_act", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 26, "repeat": 2, "selection_seconds": 0.0025571510000190756, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.139086297001995, "tokens_per_second": 5.761158231547437, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "wall_seconds": 4.165828993999639 }, { "acceptance_rate": 0.3333333333333333, "accepted_draft_tokens": 7, "act_margin_threshold": 1.0, "act_mean_margin": 0.8446970352759728, "act_mean_verified_drafts_per_block": 1.2352941176470589, "act_raw_draft_tokens": 65, "act_trimmed_draft_tokens": 44, "act_verified_draft_tokens": 21, "boltzmann_candidate_scores": 0, "context_seconds": 0.0029781249991174263, "draft_forward_passes": 17, "draft_seconds": 0.016352753999399283, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3350.9070790000806, "method": "dflash7_act", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 21, "repeat": 0, "selection_seconds": 0.0021945699995740142, "selector_pair_scores": 0, "target_forward_passes": 17, "target_seconds": 3.32808726400026, "tokens_per_second": 7.162239785879608, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 0, "wall_seconds": 3.3509070790000806 }, { "acceptance_rate": 0.3333333333333333, "accepted_draft_tokens": 7, "act_margin_threshold": 1.0, "act_mean_margin": 0.8446970352759728, "act_mean_verified_drafts_per_block": 1.2352941176470589, "act_raw_draft_tokens": 65, "act_trimmed_draft_tokens": 44, "act_verified_draft_tokens": 21, "boltzmann_candidate_scores": 0, "context_seconds": 0.002951535999272892, "draft_forward_passes": 17, "draft_seconds": 0.016047155998876406, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3280.740614000024, "method": "dflash7_act", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 21, "repeat": 1, "selection_seconds": 0.002050582000265422, "selector_pair_scores": 0, "target_forward_passes": 17, "target_seconds": 3.2583934199988107, "tokens_per_second": 7.315421370889221, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 0, "wall_seconds": 3.280740614000024 }, { "acceptance_rate": 0.3333333333333333, "accepted_draft_tokens": 7, "act_margin_threshold": 1.0, "act_mean_margin": 0.8446970352759728, "act_mean_verified_drafts_per_block": 1.2352941176470589, "act_raw_draft_tokens": 65, "act_trimmed_draft_tokens": 44, "act_verified_draft_tokens": 21, "boltzmann_candidate_scores": 0, "context_seconds": 0.002890078001200891, "draft_forward_passes": 17, "draft_seconds": 0.01576202399974136, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3220.1932010002565, "method": "dflash7_act", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 21, "repeat": 2, "selection_seconds": 0.002091239998662786, "selector_pair_scores": 0, "target_forward_passes": 17, "target_seconds": 3.1981801039996753, "tokens_per_second": 7.452968968615026, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 0, "wall_seconds": 3.2201932010002565 }, { "acceptance_rate": 0.13333333333333333, "accepted_draft_tokens": 4, "act_margin_threshold": 1.0, "act_mean_margin": 1.147683900197347, "act_mean_verified_drafts_per_block": 1.5, "act_raw_draft_tokens": 75, "act_trimmed_draft_tokens": 45, "act_verified_draft_tokens": 30, "boltzmann_candidate_scores": 0, "context_seconds": 0.003533589998824027, "draft_forward_passes": 20, "draft_seconds": 0.019811613001365913, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4318.1275669999195, "method": "dflash7_act", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 30, "repeat": 0, "selection_seconds": 0.0025197799991474312, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.290691719997994, "tokens_per_second": 5.557964563949726, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "wall_seconds": 4.3181275669999195 }, { "acceptance_rate": 0.13333333333333333, "accepted_draft_tokens": 4, "act_margin_threshold": 1.0, "act_mean_margin": 1.147683900197347, "act_mean_verified_drafts_per_block": 1.5, "act_raw_draft_tokens": 75, "act_trimmed_draft_tokens": 45, "act_verified_draft_tokens": 30, "boltzmann_candidate_scores": 0, "context_seconds": 0.003537382000558864, "draft_forward_passes": 20, "draft_seconds": 0.01853002600137188, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4269.758693999847, "method": "dflash7_act", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 30, "repeat": 1, "selection_seconds": 0.002500331000192091, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.243710604000171, "tokens_per_second": 5.620926548782819, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "wall_seconds": 4.269758693999847 }, { "acceptance_rate": 0.13333333333333333, "accepted_draft_tokens": 4, "act_margin_threshold": 1.0, "act_mean_margin": 1.147683900197347, "act_mean_verified_drafts_per_block": 1.5, "act_raw_draft_tokens": 75, "act_trimmed_draft_tokens": 45, "act_verified_draft_tokens": 30, "boltzmann_candidate_scores": 0, "context_seconds": 0.0034992929995496524, "draft_forward_passes": 20, "draft_seconds": 0.019248937998781912, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4297.325780999927, "method": "dflash7_act", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 30, "repeat": 2, "selection_seconds": 0.002537682998990931, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.270540744999835, "tokens_per_second": 5.58486864228747, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "wall_seconds": 4.297325780999927 } ], "dspark_v9": [ { "acceptance_rate": 0.075, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.003535416000886471, "draft_forward_passes": 21, "draft_seconds": 0.019333885999913036, "dspark_confidence_evaluations": 78, "dspark_markov_candidate_scores": 672, "dspark_mean_confidence": 0.5608076073038273, "dspark_survival_floor": 0.35, "dspark_trimmed_tokens": 38, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4479.005426000185, "method": "dspark_v9", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 40, "repeat": 0, "selection_seconds": 0.0069866280005044246, "selector_pair_scores": 672, "target_forward_passes": 21, "target_seconds": 4.447590340999341, "tokens_per_second": 5.358332423685483, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 672, "wall_seconds": 4.479005426000185 }, { "acceptance_rate": 0.075, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.0035269230006633734, "draft_forward_passes": 21, "draft_seconds": 0.019242205999944417, "dspark_confidence_evaluations": 78, "dspark_markov_candidate_scores": 672, "dspark_mean_confidence": 0.5608076073038273, "dspark_survival_floor": 0.35, "dspark_trimmed_tokens": 38, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4307.446116999927, "method": "dspark_v9", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 40, "repeat": 1, "selection_seconds": 0.006974171999445389, "selector_pair_scores": 672, "target_forward_passes": 21, "target_seconds": 4.27614790600046, "tokens_per_second": 5.571747004630124, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 672, "wall_seconds": 4.307446116999927 }, { "acceptance_rate": 0.075, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.0036959060003027844, "draft_forward_passes": 21, "draft_seconds": 0.020354298999336606, "dspark_confidence_evaluations": 78, "dspark_markov_candidate_scores": 672, "dspark_mean_confidence": 0.5608076073038273, "dspark_survival_floor": 0.35, "dspark_trimmed_tokens": 38, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4375.856764000218, "method": "dspark_v9", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 40, "repeat": 2, "selection_seconds": 0.007056379999994533, "selector_pair_scores": 672, "target_forward_passes": 21, "target_seconds": 4.343208920998222, "tokens_per_second": 5.484640218904295, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 672, "wall_seconds": 4.375856764000218 }, { "acceptance_rate": 0.30434782608695654, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.0032437040022159636, "draft_forward_passes": 18, "draft_seconds": 0.01737620500034609, "dspark_confidence_evaluations": 66, "dspark_markov_candidate_scores": 576, "dspark_mean_confidence": 0.41720192822994606, "dspark_survival_floor": 0.35, "dspark_trimmed_tokens": 43, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3967.1898219999093, "method": "dspark_v9", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 23, "repeat": 0, "selection_seconds": 0.006057920001239836, "selector_pair_scores": 576, "target_forward_passes": 18, "target_seconds": 3.939147271001275, "tokens_per_second": 6.049622296092024, "tokens_per_target_pass": 1.3333333333333333, "total_guidance_scores": 576, "wall_seconds": 3.9671898219999093 }, { "acceptance_rate": 0.30434782608695654, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.003407439000056911, "draft_forward_passes": 18, "draft_seconds": 0.01797252300002583, "dspark_confidence_evaluations": 66, "dspark_markov_candidate_scores": 576, "dspark_mean_confidence": 0.41720192822994606, "dspark_survival_floor": 0.35, "dspark_trimmed_tokens": 43, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4128.249485000197, "method": "dspark_v9", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 23, "repeat": 1, "selection_seconds": 0.006218838001132099, "selector_pair_scores": 576, "target_forward_passes": 18, "target_seconds": 4.099301538999498, "tokens_per_second": 5.813602130201406, "tokens_per_target_pass": 1.3333333333333333, "total_guidance_scores": 576, "wall_seconds": 4.128249485000197 }, { "acceptance_rate": 0.30434782608695654, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.0032436420001431543, "draft_forward_passes": 18, "draft_seconds": 0.017410426999958872, "dspark_confidence_evaluations": 66, "dspark_markov_candidate_scores": 576, "dspark_mean_confidence": 0.41720192822994606, "dspark_survival_floor": 0.35, "dspark_trimmed_tokens": 43, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4055.6359709999015, "method": "dspark_v9", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 23, "repeat": 2, "selection_seconds": 0.006191841999680037, "selector_pair_scores": 576, "target_forward_passes": 18, "target_seconds": 4.027452451000499, "tokens_per_second": 5.917690880447264, "tokens_per_target_pass": 1.3333333333333333, "total_guidance_scores": 576, "wall_seconds": 4.0556359709999015 }, { "acceptance_rate": 0.21739130434782608, "accepted_draft_tokens": 5, "boltzmann_candidate_scores": 0, "context_seconds": 0.003636558000380319, "draft_forward_passes": 20, "draft_seconds": 0.019757422000111546, "dspark_confidence_evaluations": 74, "dspark_markov_candidate_scores": 640, "dspark_mean_confidence": 0.36538131484711495, "dspark_survival_floor": 0.35, "dspark_trimmed_tokens": 51, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3999.5622540000113, "method": "dspark_v9", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 23, "repeat": 0, "selection_seconds": 0.0069999719999032095, "selector_pair_scores": 640, "target_forward_passes": 20, "target_seconds": 3.96765444499988, "tokens_per_second": 6.000656690865933, "tokens_per_target_pass": 1.2, "total_guidance_scores": 640, "wall_seconds": 3.9995622540000113 }, { "acceptance_rate": 0.21739130434782608, "accepted_draft_tokens": 5, "boltzmann_candidate_scores": 0, "context_seconds": 0.0035709030016732868, "draft_forward_passes": 20, "draft_seconds": 0.019654777001051116, "dspark_confidence_evaluations": 74, "dspark_markov_candidate_scores": 640, "dspark_mean_confidence": 0.36538131484711495, "dspark_survival_floor": 0.35, "dspark_trimmed_tokens": 51, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4089.333798000098, "method": "dspark_v9", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 23, "repeat": 1, "selection_seconds": 0.00679037400004745, "selector_pair_scores": 640, "target_forward_passes": 20, "target_seconds": 4.057742958999825, "tokens_per_second": 5.86892662363177, "tokens_per_target_pass": 1.2, "total_guidance_scores": 640, "wall_seconds": 4.089333798000098 }, { "acceptance_rate": 0.21739130434782608, "accepted_draft_tokens": 5, "boltzmann_candidate_scores": 0, "context_seconds": 0.0036236900004951167, "draft_forward_passes": 20, "draft_seconds": 0.019984681001460558, "dspark_confidence_evaluations": 74, "dspark_markov_candidate_scores": 640, "dspark_mean_confidence": 0.36538131484711495, "dspark_survival_floor": 0.35, "dspark_trimmed_tokens": 51, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4048.0708590002905, "method": "dspark_v9", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 23, "repeat": 2, "selection_seconds": 0.006978468999477627, "selector_pair_scores": 640, "target_forward_passes": 20, "target_seconds": 4.0160016189997805, "tokens_per_second": 5.92874997398811, "tokens_per_target_pass": 1.2, "total_guidance_scores": 640, "wall_seconds": 4.0480708590002905 }, { "acceptance_rate": 0.07692307692307693, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.0037079259991514846, "draft_forward_passes": 21, "draft_seconds": 0.020225930999913544, "dspark_confidence_evaluations": 78, "dspark_markov_candidate_scores": 672, "dspark_mean_confidence": 0.5499186745056739, "dspark_survival_floor": 0.35, "dspark_trimmed_tokens": 39, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4429.449243999898, "method": "dspark_v9", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 39, "repeat": 0, "selection_seconds": 0.007108283999059495, "selector_pair_scores": 672, "target_forward_passes": 21, "target_seconds": 4.396794478998345, "tokens_per_second": 5.418280846656103, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 672, "wall_seconds": 4.429449243999898 }, { "acceptance_rate": 0.07692307692307693, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.0035797939995063643, "draft_forward_passes": 21, "draft_seconds": 0.01975681599924428, "dspark_confidence_evaluations": 78, "dspark_markov_candidate_scores": 672, "dspark_mean_confidence": 0.5499186745056739, "dspark_survival_floor": 0.35, "dspark_trimmed_tokens": 39, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4432.130612000037, "method": "dspark_v9", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 39, "repeat": 1, "selection_seconds": 0.006954094999400695, "selector_pair_scores": 672, "target_forward_passes": 21, "target_seconds": 4.400334347000808, "tokens_per_second": 5.415002873565992, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 672, "wall_seconds": 4.432130612000037 }, { "acceptance_rate": 0.07692307692307693, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.0036795269993490365, "draft_forward_passes": 21, "draft_seconds": 0.019764788999964367, "dspark_confidence_evaluations": 78, "dspark_markov_candidate_scores": 672, "dspark_mean_confidence": 0.5499186745056739, "dspark_survival_floor": 0.35, "dspark_trimmed_tokens": 39, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4420.863072999964, "method": "dspark_v9", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 39, "repeat": 2, "selection_seconds": 0.006984557998748642, "selector_pair_scores": 672, "target_forward_passes": 21, "target_seconds": 4.388900081999054, "tokens_per_second": 5.428804195854404, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 672, "wall_seconds": 4.420863072999964 }, { "acceptance_rate": 0.35, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.0029342360007831303, "draft_forward_passes": 17, "draft_seconds": 0.016492530000050465, "dspark_confidence_evaluations": 65, "dspark_markov_candidate_scores": 544, "dspark_mean_confidence": 0.4400164400155728, "dspark_survival_floor": 0.35, "dspark_trimmed_tokens": 45, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3350.7558279998193, "method": "dspark_v9", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 20, "repeat": 0, "selection_seconds": 0.0057300720004604955, "selector_pair_scores": 544, "target_forward_passes": 17, "target_seconds": 3.3243507419992966, "tokens_per_second": 7.162563084856714, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 544, "wall_seconds": 3.3507558279998193 }, { "acceptance_rate": 0.35, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.0029321180004444614, "draft_forward_passes": 17, "draft_seconds": 0.016026861002046644, "dspark_confidence_evaluations": 65, "dspark_markov_candidate_scores": 544, "dspark_mean_confidence": 0.4400164400155728, "dspark_survival_floor": 0.35, "dspark_trimmed_tokens": 45, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3287.1020470001895, "method": "dspark_v9", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 20, "repeat": 1, "selection_seconds": 0.00552593300017179, "selector_pair_scores": 544, "target_forward_passes": 17, "target_seconds": 3.2613505550007176, "tokens_per_second": 7.301264048648082, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 544, "wall_seconds": 3.2871020470001895 }, { "acceptance_rate": 0.35, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.0029702219999308, "draft_forward_passes": 17, "draft_seconds": 0.016192623000279127, "dspark_confidence_evaluations": 65, "dspark_markov_candidate_scores": 544, "dspark_mean_confidence": 0.4400164400155728, "dspark_survival_floor": 0.35, "dspark_trimmed_tokens": 45, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3294.750819000001, "method": "dspark_v9", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 20, "repeat": 2, "selection_seconds": 0.0056556430008640746, "selector_pair_scores": 544, "target_forward_passes": 17, "target_seconds": 3.268702413999108, "tokens_per_second": 7.28431414648963, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 544, "wall_seconds": 3.2947508190000008 }, { "acceptance_rate": 0.15384615384615385, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 0, "context_seconds": 0.00363178099905781, "draft_forward_passes": 20, "draft_seconds": 0.019514867999532726, "dspark_confidence_evaluations": 75, "dspark_markov_candidate_scores": 640, "dspark_mean_confidence": 0.39661210517088574, "dspark_survival_floor": 0.35, "dspark_trimmed_tokens": 49, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4280.0865600002, "method": "dspark_v9", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 26, "repeat": 0, "selection_seconds": 0.006815378000737837, "selector_pair_scores": 640, "target_forward_passes": 20, "target_seconds": 4.24861435800085, "tokens_per_second": 5.6073632305228145, "tokens_per_target_pass": 1.2, "total_guidance_scores": 640, "wall_seconds": 4.2800865600002 }, { "acceptance_rate": 0.15384615384615385, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 0, "context_seconds": 0.003538685000876285, "draft_forward_passes": 20, "draft_seconds": 0.01890168399859249, "dspark_confidence_evaluations": 75, "dspark_markov_candidate_scores": 640, "dspark_mean_confidence": 0.39661210517088574, "dspark_survival_floor": 0.35, "dspark_trimmed_tokens": 49, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4237.5819499998215, "method": "dspark_v9", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 26, "repeat": 1, "selection_seconds": 0.006671196998922824, "selector_pair_scores": 640, "target_forward_passes": 20, "target_seconds": 4.206962030999421, "tokens_per_second": 5.663607284338421, "tokens_per_target_pass": 1.2, "total_guidance_scores": 640, "wall_seconds": 4.2375819499998215 }, { "acceptance_rate": 0.15384615384615385, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 0, "context_seconds": 0.0036137769989181834, "draft_forward_passes": 20, "draft_seconds": 0.019297581999126123, "dspark_confidence_evaluations": 75, "dspark_markov_candidate_scores": 640, "dspark_mean_confidence": 0.39661210517088574, "dspark_survival_floor": 0.35, "dspark_trimmed_tokens": 49, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4286.423983999612, "method": "dspark_v9", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 26, "repeat": 2, "selection_seconds": 0.006605012999443716, "selector_pair_scores": 640, "target_forward_passes": 20, "target_seconds": 4.255241210000804, "tokens_per_second": 5.59907281444564, "tokens_per_target_pass": 1.2, "total_guidance_scores": 640, "wall_seconds": 4.286423983999612 } ], "minop_v13": [ { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 22, "context_seconds": 0.003606081999350863, "draft_forward_passes": 21, "draft_seconds": 0.0203638939983648, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4772.017110999968, "method": "minop_v13", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 0, "selection_seconds": 0.0022226920009416062, "selector_pair_scores": 0, "target_forward_passes": 21, "target_seconds": 4.744149666999419, "tokens_per_second": 5.029319770182224, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 22, "v13_candidate_scores": 22, "v13_fast_argmax_positions": 73, "v13_policy_source": "selected V10 temperature/cutoff/slope; hard routing budget=1", "v13_routed_positions": 11, "v13_top2_values_transferred": 168, "wall_seconds": 4.772017110999968 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 22, "context_seconds": 0.003514879998874676, "draft_forward_passes": 21, "draft_seconds": 0.01940387699960411, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4575.909009000043, "method": "minop_v13", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 1, "selection_seconds": 0.0021077780006635294, "selector_pair_scores": 0, "target_forward_passes": 21, "target_seconds": 4.549357138000687, "tokens_per_second": 5.244859535623641, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 22, "v13_candidate_scores": 22, "v13_fast_argmax_positions": 73, "v13_policy_source": "selected V10 temperature/cutoff/slope; hard routing budget=1", "v13_routed_positions": 11, "v13_top2_values_transferred": 168, "wall_seconds": 4.575909009000043 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 22, "context_seconds": 0.003639066998857743, "draft_forward_passes": 21, "draft_seconds": 0.01989739300097426, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4663.756385999932, "method": "minop_v13", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 2, "selection_seconds": 0.002146769999853859, "selector_pair_scores": 0, "target_forward_passes": 21, "target_seconds": 4.636475068999971, "tokens_per_second": 5.146066392328142, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 22, "v13_candidate_scores": 22, "v13_fast_argmax_positions": 73, "v13_policy_source": "selected V10 temperature/cutoff/slope; hard routing budget=1", "v13_routed_positions": 11, "v13_top2_values_transferred": 168, "wall_seconds": 4.663756385999932 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 26, "context_seconds": 0.0033087489987337904, "draft_forward_passes": 19, "draft_seconds": 0.018045733000690234, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4580.499734000114, "method": "minop_v13", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 70, "repeat": 0, "selection_seconds": 0.002162509999834583, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.555539938000038, "tokens_per_second": 5.239602967740158, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 26, "v13_candidate_scores": 26, "v13_fast_argmax_positions": 63, "v13_policy_source": "selected V10 temperature/cutoff/slope; hard routing budget=1", "v13_routed_positions": 13, "v13_top2_values_transferred": 152, "wall_seconds": 4.580499734000114 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 26, "context_seconds": 0.003334902000005968, "draft_forward_passes": 19, "draft_seconds": 0.017947023999568046, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4696.199011000317, "method": "minop_v13", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 70, "repeat": 1, "selection_seconds": 0.002147006999621226, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.671309129001202, "tokens_per_second": 5.110515960627457, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 26, "v13_candidate_scores": 26, "v13_fast_argmax_positions": 63, "v13_policy_source": "selected V10 temperature/cutoff/slope; hard routing budget=1", "v13_routed_positions": 13, "v13_top2_values_transferred": 152, "wall_seconds": 4.696199011000317 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 26, "context_seconds": 0.00336136299847567, "draft_forward_passes": 19, "draft_seconds": 0.017879337000067608, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4666.03900400014, "method": "minop_v13", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 70, "repeat": 2, "selection_seconds": 0.002184780998959468, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.641160324998509, "tokens_per_second": 5.14354894578144, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 26, "v13_candidate_scores": 26, "v13_fast_argmax_positions": 63, "v13_policy_source": "selected V10 temperature/cutoff/slope; hard routing budget=1", "v13_routed_positions": 13, "v13_top2_values_transferred": 152, "wall_seconds": 4.66603900400014 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 28, "context_seconds": 0.003244535999783693, "draft_forward_passes": 19, "draft_seconds": 0.018253123999784293, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4205.775476000326, "method": "minop_v13", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 0, "selection_seconds": 0.0021486460009327857, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.180720538000514, "tokens_per_second": 5.7064387143233555, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 28, "v13_candidate_scores": 28, "v13_fast_argmax_positions": 62, "v13_policy_source": "selected V10 temperature/cutoff/slope; hard routing budget=1", "v13_routed_positions": 14, "v13_top2_values_transferred": 152, "wall_seconds": 4.205775476000326 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 28, "context_seconds": 0.0033715740005391126, "draft_forward_passes": 19, "draft_seconds": 0.018411359000765515, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4279.328707999866, "method": "minop_v13", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 1, "selection_seconds": 0.00215914000000339, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.253914580000583, "tokens_per_second": 5.608356272126023, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 28, "v13_candidate_scores": 28, "v13_fast_argmax_positions": 62, "v13_policy_source": "selected V10 temperature/cutoff/slope; hard routing budget=1", "v13_routed_positions": 14, "v13_top2_values_transferred": 152, "wall_seconds": 4.279328707999866 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 28, "context_seconds": 0.003249199999117991, "draft_forward_passes": 19, "draft_seconds": 0.018124733000149718, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4242.457967000064, "method": "minop_v13", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 2, "selection_seconds": 0.0021414380003079714, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.217500477000613, "tokens_per_second": 5.6570978868108694, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 28, "v13_candidate_scores": 28, "v13_fast_argmax_positions": 62, "v13_policy_source": "selected V10 temperature/cutoff/slope; hard routing budget=1", "v13_routed_positions": 14, "v13_top2_values_transferred": 152, "wall_seconds": 4.242457967000064 }, { "acceptance_rate": 0.05405405405405406, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 32, "context_seconds": 0.003374722998614743, "draft_forward_passes": 20, "draft_seconds": 0.018709213000875025, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4496.5865879999, "method": "minop_v13", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 74, "repeat": 0, "selection_seconds": 0.0022697810009049135, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.470639302000109, "tokens_per_second": 5.337381929672858, "tokens_per_target_pass": 1.2, "total_guidance_scores": 32, "v13_candidate_scores": 32, "v13_fast_argmax_positions": 64, "v13_policy_source": "selected V10 temperature/cutoff/slope; hard routing budget=1", "v13_routed_positions": 16, "v13_top2_values_transferred": 160, "wall_seconds": 4.4965865879999 }, { "acceptance_rate": 0.05405405405405406, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 32, "context_seconds": 0.003371568998773, "draft_forward_passes": 20, "draft_seconds": 0.018597540000428125, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4532.575020999957, "method": "minop_v13", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 74, "repeat": 1, "selection_seconds": 0.002252628000405821, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.506809695999436, "tokens_per_second": 5.29500336757917, "tokens_per_target_pass": 1.2, "total_guidance_scores": 32, "v13_candidate_scores": 32, "v13_fast_argmax_positions": 64, "v13_policy_source": "selected V10 temperature/cutoff/slope; hard routing budget=1", "v13_routed_positions": 16, "v13_top2_values_transferred": 160, "wall_seconds": 4.532575020999957 }, { "acceptance_rate": 0.05405405405405406, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 32, "context_seconds": 0.003359692000231007, "draft_forward_passes": 20, "draft_seconds": 0.018546208999850933, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4487.3252920001505, "method": "minop_v13", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 74, "repeat": 2, "selection_seconds": 0.0023719989999335667, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.461570152000604, "tokens_per_second": 5.348397639633208, "tokens_per_target_pass": 1.2, "total_guidance_scores": 32, "v13_candidate_scores": 32, "v13_fast_argmax_positions": 64, "v13_policy_source": "selected V10 temperature/cutoff/slope; hard routing budget=1", "v13_routed_positions": 16, "v13_top2_values_transferred": 160, "wall_seconds": 4.4873252920001505 }, { "acceptance_rate": 0.15789473684210525, "accepted_draft_tokens": 9, "boltzmann_candidate_scores": 22, "context_seconds": 0.002583904000402981, "draft_forward_passes": 15, "draft_seconds": 0.013981220999994548, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3314.1236269998444, "method": "minop_v13", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 57, "repeat": 0, "selection_seconds": 0.0017712539988679055, "selector_pair_scores": 0, "target_forward_passes": 15, "target_seconds": 3.294608380000227, "tokens_per_second": 7.241733472002771, "tokens_per_target_pass": 1.6, "total_guidance_scores": 22, "v13_candidate_scores": 22, "v13_fast_argmax_positions": 49, "v13_policy_source": "selected V10 temperature/cutoff/slope; hard routing budget=1", "v13_routed_positions": 11, "v13_top2_values_transferred": 120, "wall_seconds": 3.3141236269998444 }, { "acceptance_rate": 0.15789473684210525, "accepted_draft_tokens": 9, "boltzmann_candidate_scores": 22, "context_seconds": 0.00246367499948974, "draft_forward_passes": 15, "draft_seconds": 0.014073860999815224, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3354.183937000016, "method": "minop_v13", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 57, "repeat": 1, "selection_seconds": 0.0016846699995767267, "selector_pair_scores": 0, "target_forward_passes": 15, "target_seconds": 3.33482393300028, "tokens_per_second": 7.1552426613388445, "tokens_per_target_pass": 1.6, "total_guidance_scores": 22, "v13_candidate_scores": 22, "v13_fast_argmax_positions": 49, "v13_policy_source": "selected V10 temperature/cutoff/slope; hard routing budget=1", "v13_routed_positions": 11, "v13_top2_values_transferred": 120, "wall_seconds": 3.354183937000016 }, { "acceptance_rate": 0.15789473684210525, "accepted_draft_tokens": 9, "boltzmann_candidate_scores": 22, "context_seconds": 0.002431261999390699, "draft_forward_passes": 15, "draft_seconds": 0.014045688998976402, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3324.466798999765, "method": "minop_v13", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 57, "repeat": 2, "selection_seconds": 0.0016896629999791912, "selector_pair_scores": 0, "target_forward_passes": 15, "target_seconds": 3.3051592860006167, "tokens_per_second": 7.219202792827078, "tokens_per_target_pass": 1.6, "total_guidance_scores": 22, "v13_candidate_scores": 22, "v13_fast_argmax_positions": 49, "v13_policy_source": "selected V10 temperature/cutoff/slope; hard routing budget=1", "v13_routed_positions": 11, "v13_top2_values_transferred": 120, "wall_seconds": 3.324466798999765 }, { "acceptance_rate": 0.05333333333333334, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 30, "context_seconds": 0.0033952739995584125, "draft_forward_passes": 20, "draft_seconds": 0.01854917899981956, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4687.1426060001795, "method": "minop_v13", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 75, "repeat": 0, "selection_seconds": 0.002227828998456971, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.661462063000727, "tokens_per_second": 5.120390399318497, "tokens_per_target_pass": 1.2, "total_guidance_scores": 30, "v13_candidate_scores": 30, "v13_fast_argmax_positions": 65, "v13_policy_source": "selected V10 temperature/cutoff/slope; hard routing budget=1", "v13_routed_positions": 15, "v13_top2_values_transferred": 160, "wall_seconds": 4.6871426060001795 }, { "acceptance_rate": 0.05333333333333334, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 30, "context_seconds": 0.0034833219997381093, "draft_forward_passes": 20, "draft_seconds": 0.01944112200089876, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4836.959102000037, "method": "minop_v13", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 75, "repeat": 1, "selection_seconds": 0.002294504000019515, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.810224958999697, "tokens_per_second": 4.961795105953289, "tokens_per_target_pass": 1.2, "total_guidance_scores": 30, "v13_candidate_scores": 30, "v13_fast_argmax_positions": 65, "v13_policy_source": "selected V10 temperature/cutoff/slope; hard routing budget=1", "v13_routed_positions": 15, "v13_top2_values_transferred": 160, "wall_seconds": 4.836959102000037 }, { "acceptance_rate": 0.05333333333333334, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 30, "context_seconds": 0.0035222340002292185, "draft_forward_passes": 20, "draft_seconds": 0.01908783499902711, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4665.733471999829, "method": "minop_v13", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 75, "repeat": 2, "selection_seconds": 0.0023633070004507317, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.639265684000748, "tokens_per_second": 5.143885767163873, "tokens_per_target_pass": 1.2, "total_guidance_scores": 30, "v13_candidate_scores": 30, "v13_fast_argmax_positions": 65, "v13_policy_source": "selected V10 temperature/cutoff/slope; hard routing budget=1", "v13_routed_positions": 15, "v13_top2_values_transferred": 160, "wall_seconds": 4.665733471999829 } ], "normal": [ { "acceptance_rate": 0.0, "accepted_draft_tokens": 0, "boltzmann_candidate_scores": 0, "context_seconds": 0.0, "draft_forward_passes": 0, "draft_seconds": 0.0, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4383.117909000248, "method": "normal", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 0, "repeat": 0, "selection_seconds": 0.0, "selector_pair_scores": 0, "target_forward_passes": 24, "target_seconds": 4.382232491001105, "tokens_per_second": 5.475554273983517, "tokens_per_target_pass": 1.0, "total_guidance_scores": 0, "wall_seconds": 4.383117909000248 }, { "acceptance_rate": 0.0, "accepted_draft_tokens": 0, "boltzmann_candidate_scores": 0, "context_seconds": 0.0, "draft_forward_passes": 0, "draft_seconds": 0.0, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4351.91076800038, "method": "normal", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 0, "repeat": 1, "selection_seconds": 0.0, "selector_pair_scores": 0, "target_forward_passes": 24, "target_seconds": 4.351019577999978, "tokens_per_second": 5.514818956416136, "tokens_per_target_pass": 1.0, "total_guidance_scores": 0, "wall_seconds": 4.35191076800038 }, { "acceptance_rate": 0.0, "accepted_draft_tokens": 0, "boltzmann_candidate_scores": 0, "context_seconds": 0.0, "draft_forward_passes": 0, "draft_seconds": 0.0, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4453.03997200017, "method": "normal", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 0, "repeat": 2, "selection_seconds": 0.0, "selector_pair_scores": 0, "target_forward_passes": 24, "target_seconds": 4.452125929998601, "tokens_per_second": 5.389576592823606, "tokens_per_target_pass": 1.0, "total_guidance_scores": 0, "wall_seconds": 4.45303997200017 }, { "acceptance_rate": 0.0, "accepted_draft_tokens": 0, "boltzmann_candidate_scores": 0, "context_seconds": 0.0, "draft_forward_passes": 0, "draft_seconds": 0.0, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4826.232691000314, "method": "normal", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 0, "repeat": 0, "selection_seconds": 0.0, "selector_pair_scores": 0, "target_forward_passes": 24, "target_seconds": 4.825363332000961, "tokens_per_second": 4.9728228074775265, "tokens_per_target_pass": 1.0, "total_guidance_scores": 0, "wall_seconds": 4.826232691000314 }, { "acceptance_rate": 0.0, "accepted_draft_tokens": 0, "boltzmann_candidate_scores": 0, "context_seconds": 0.0, "draft_forward_passes": 0, "draft_seconds": 0.0, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 5019.533017999947, "method": "normal", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 0, "repeat": 1, "selection_seconds": 0.0, "selector_pair_scores": 0, "target_forward_passes": 24, "target_seconds": 5.0186697469994215, "tokens_per_second": 4.781321273101795, "tokens_per_target_pass": 1.0, "total_guidance_scores": 0, "wall_seconds": 5.019533017999947 }, { "acceptance_rate": 0.0, "accepted_draft_tokens": 0, "boltzmann_candidate_scores": 0, "context_seconds": 0.0, "draft_forward_passes": 0, "draft_seconds": 0.0, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 5034.519902999818, "method": "normal", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 0, "repeat": 2, "selection_seconds": 0.0, "selector_pair_scores": 0, "target_forward_passes": 24, "target_seconds": 5.033655763999832, "tokens_per_second": 4.767088116127935, "tokens_per_target_pass": 1.0, "total_guidance_scores": 0, "wall_seconds": 5.034519902999818 }, { "acceptance_rate": 0.0, "accepted_draft_tokens": 0, "boltzmann_candidate_scores": 0, "context_seconds": 0.0, "draft_forward_passes": 0, "draft_seconds": 0.0, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4324.049445000128, "method": "normal", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 0, "repeat": 0, "selection_seconds": 0.0, "selector_pair_scores": 0, "target_forward_passes": 24, "target_seconds": 4.323183959000744, "tokens_per_second": 5.550352812859495, "tokens_per_target_pass": 1.0, "total_guidance_scores": 0, "wall_seconds": 4.324049445000128 }, { "acceptance_rate": 0.0, "accepted_draft_tokens": 0, "boltzmann_candidate_scores": 0, "context_seconds": 0.0, "draft_forward_passes": 0, "draft_seconds": 0.0, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4447.412288000123, "method": "normal", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 0, "repeat": 1, "selection_seconds": 0.0, "selector_pair_scores": 0, "target_forward_passes": 24, "target_seconds": 4.446546141001818, "tokens_per_second": 5.396396476386076, "tokens_per_target_pass": 1.0, "total_guidance_scores": 0, "wall_seconds": 4.447412288000123 }, { "acceptance_rate": 0.0, "accepted_draft_tokens": 0, "boltzmann_candidate_scores": 0, "context_seconds": 0.0, "draft_forward_passes": 0, "draft_seconds": 0.0, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4322.368250000181, "method": "normal", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 0, "repeat": 2, "selection_seconds": 0.0, "selector_pair_scores": 0, "target_forward_passes": 24, "target_seconds": 4.32149138800014, "tokens_per_second": 5.55251163525898, "tokens_per_target_pass": 1.0, "total_guidance_scores": 0, "wall_seconds": 4.322368250000181 }, { "acceptance_rate": 0.0, "accepted_draft_tokens": 0, "boltzmann_candidate_scores": 0, "context_seconds": 0.0, "draft_forward_passes": 0, "draft_seconds": 0.0, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4368.374973999835, "method": "normal", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 0, "repeat": 0, "selection_seconds": 0.0, "selector_pair_scores": 0, "target_forward_passes": 24, "target_seconds": 4.367510915999901, "tokens_per_second": 5.4940338553457035, "tokens_per_target_pass": 1.0, "total_guidance_scores": 0, "wall_seconds": 4.368374973999835 }, { "acceptance_rate": 0.0, "accepted_draft_tokens": 0, "boltzmann_candidate_scores": 0, "context_seconds": 0.0, "draft_forward_passes": 0, "draft_seconds": 0.0, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4367.182100999798, "method": "normal", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 0, "repeat": 1, "selection_seconds": 0.0, "selector_pair_scores": 0, "target_forward_passes": 24, "target_seconds": 4.366295213000285, "tokens_per_second": 5.49553452202178, "tokens_per_target_pass": 1.0, "total_guidance_scores": 0, "wall_seconds": 4.367182100999798 }, { "acceptance_rate": 0.0, "accepted_draft_tokens": 0, "boltzmann_candidate_scores": 0, "context_seconds": 0.0, "draft_forward_passes": 0, "draft_seconds": 0.0, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4368.850736999775, "method": "normal", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 0, "repeat": 2, "selection_seconds": 0.0, "selector_pair_scores": 0, "target_forward_passes": 24, "target_seconds": 4.368018621000829, "tokens_per_second": 5.493435561152071, "tokens_per_target_pass": 1.0, "total_guidance_scores": 0, "wall_seconds": 4.368850736999775 }, { "acceptance_rate": 0.0, "accepted_draft_tokens": 0, "boltzmann_candidate_scores": 0, "context_seconds": 0.0, "draft_forward_passes": 0, "draft_seconds": 0.0, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4304.382584999985, "method": "normal", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 0, "repeat": 0, "selection_seconds": 0.0, "selector_pair_scores": 0, "target_forward_passes": 24, "target_seconds": 4.303497779000281, "tokens_per_second": 5.5757125501891425, "tokens_per_target_pass": 1.0, "total_guidance_scores": 0, "wall_seconds": 4.304382584999985 }, { "acceptance_rate": 0.0, "accepted_draft_tokens": 0, "boltzmann_candidate_scores": 0, "context_seconds": 0.0, "draft_forward_passes": 0, "draft_seconds": 0.0, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4303.385703999993, "method": "normal", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 0, "repeat": 1, "selection_seconds": 0.0, "selector_pair_scores": 0, "target_forward_passes": 24, "target_seconds": 4.302473852999356, "tokens_per_second": 5.577004166206162, "tokens_per_target_pass": 1.0, "total_guidance_scores": 0, "wall_seconds": 4.303385703999993 }, { "acceptance_rate": 0.0, "accepted_draft_tokens": 0, "boltzmann_candidate_scores": 0, "context_seconds": 0.0, "draft_forward_passes": 0, "draft_seconds": 0.0, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4754.860301000008, "method": "normal", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 0, "repeat": 2, "selection_seconds": 0.0, "selector_pair_scores": 0, "target_forward_passes": 24, "target_seconds": 4.753882917001192, "tokens_per_second": 5.047466903486627, "tokens_per_target_pass": 1.0, "total_guidance_scores": 0, "wall_seconds": 4.754860301000008 }, { "acceptance_rate": 0.0, "accepted_draft_tokens": 0, "boltzmann_candidate_scores": 0, "context_seconds": 0.0, "draft_forward_passes": 0, "draft_seconds": 0.0, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4760.44611000043, "method": "normal", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 0, "repeat": 0, "selection_seconds": 0.0, "selector_pair_scores": 0, "target_forward_passes": 24, "target_seconds": 4.759603336999589, "tokens_per_second": 5.041544310223865, "tokens_per_target_pass": 1.0, "total_guidance_scores": 0, "wall_seconds": 4.76044611000043 }, { "acceptance_rate": 0.0, "accepted_draft_tokens": 0, "boltzmann_candidate_scores": 0, "context_seconds": 0.0, "draft_forward_passes": 0, "draft_seconds": 0.0, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4676.918714999829, "method": "normal", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 0, "repeat": 1, "selection_seconds": 0.0, "selector_pair_scores": 0, "target_forward_passes": 24, "target_seconds": 4.67608191500085, "tokens_per_second": 5.13158373333005, "tokens_per_target_pass": 1.0, "total_guidance_scores": 0, "wall_seconds": 4.676918714999829 }, { "acceptance_rate": 0.0, "accepted_draft_tokens": 0, "boltzmann_candidate_scores": 0, "context_seconds": 0.0, "draft_forward_passes": 0, "draft_seconds": 0.0, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4667.659938999805, "method": "normal", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 0, "repeat": 2, "selection_seconds": 0.0, "selector_pair_scores": 0, "target_forward_passes": 24, "target_seconds": 4.666760914999941, "tokens_per_second": 5.141762749139511, "tokens_per_target_pass": 1.0, "total_guidance_scores": 0, "wall_seconds": 4.667659938999805 } ], "parareal_linear_v12": [ { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.003616013000737439, "draft_forward_passes": 21, "draft_seconds": 0.01978740699996706, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4917.799918000128, "method": "parareal_linear_v12", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 0, "selection_seconds": 0.017630034999911004, "selector_pair_scores": 0, "target_forward_passes": 21, "target_seconds": 4.874972664000779, "tokens_per_second": 4.880231078973997, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 0, "v12_correction_rounds": 2, "v12_linear_candidate_scores": 1344, "v12_mean_update_rms": 0.4823379576554428, "wall_seconds": 4.917799918000128 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.0035397699989516695, "draft_forward_passes": 21, "draft_seconds": 0.01901735299907159, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4662.097800999618, "method": "parareal_linear_v12", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 1, "selection_seconds": 0.017027156000040122, "selector_pair_scores": 0, "target_forward_passes": 21, "target_seconds": 4.620798637000007, "tokens_per_second": 5.147897153692072, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 0, "v12_correction_rounds": 2, "v12_linear_candidate_scores": 1344, "v12_mean_update_rms": 0.4823379576554428, "wall_seconds": 4.662097800999618 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.0036216130010870984, "draft_forward_passes": 21, "draft_seconds": 0.019148064000546583, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4683.264980999866, "method": "parareal_linear_v12", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 2, "selection_seconds": 0.016966807000244444, "selector_pair_scores": 0, "target_forward_passes": 21, "target_seconds": 4.641782322999916, "tokens_per_second": 5.12462995311362, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 0, "v12_correction_rounds": 2, "v12_linear_candidate_scores": 1344, "v12_mean_update_rms": 0.4823379576554428, "wall_seconds": 4.683264980999866 }, { "acceptance_rate": 0.10606060606060606, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.0031114670000533806, "draft_forward_passes": 18, "draft_seconds": 0.016075639999144187, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4392.68641200033, "method": "parareal_linear_v12", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 66, "repeat": 0, "selection_seconds": 0.014519049000682571, "selector_pair_scores": 0, "target_forward_passes": 18, "target_seconds": 4.357466494999699, "tokens_per_second": 5.463626981073512, "tokens_per_target_pass": 1.3333333333333333, "total_guidance_scores": 0, "v12_correction_rounds": 2, "v12_linear_candidate_scores": 1152, "v12_mean_update_rms": 0.41664244799683936, "wall_seconds": 4.39268641200033 }, { "acceptance_rate": 0.10606060606060606, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.0031349020000561723, "draft_forward_passes": 18, "draft_seconds": 0.016403013999934046, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4496.079192999787, "method": "parareal_linear_v12", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 66, "repeat": 1, "selection_seconds": 0.01474569399988468, "selector_pair_scores": 0, "target_forward_passes": 18, "target_seconds": 4.460319606000212, "tokens_per_second": 5.337984268018906, "tokens_per_target_pass": 1.3333333333333333, "total_guidance_scores": 0, "v12_correction_rounds": 2, "v12_linear_candidate_scores": 1152, "v12_mean_update_rms": 0.41664244799683936, "wall_seconds": 4.496079192999787 }, { "acceptance_rate": 0.10606060606060606, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.0030788859994572704, "draft_forward_passes": 18, "draft_seconds": 0.016196672001569823, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4357.730534999973, "method": "parareal_linear_v12", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 66, "repeat": 2, "selection_seconds": 0.01459099099974992, "selector_pair_scores": 0, "target_forward_passes": 18, "target_seconds": 4.322442628998942, "tokens_per_second": 5.507453893084774, "tokens_per_target_pass": 1.3333333333333333, "total_guidance_scores": 0, "v12_correction_rounds": 2, "v12_linear_candidate_scores": 1152, "v12_mean_update_rms": 0.41664244799683936, "wall_seconds": 4.357730534999973 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.0032307620008396043, "draft_forward_passes": 19, "draft_seconds": 0.01744606200054477, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4243.486347000271, "method": "parareal_linear_v12", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 0, "selection_seconds": 0.015529243000401038, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.205720649000341, "tokens_per_second": 5.655726927686629, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 0, "v12_correction_rounds": 2, "v12_linear_candidate_scores": 1216, "v12_mean_update_rms": 0.5209882357437378, "wall_seconds": 4.243486347000271 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.0033645649991740356, "draft_forward_passes": 19, "draft_seconds": 0.017858846999388334, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4417.608871999619, "method": "parareal_linear_v12", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 1, "selection_seconds": 0.01566032400023687, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.379199747000712, "tokens_per_second": 5.432803286891368, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 0, "v12_correction_rounds": 2, "v12_linear_candidate_scores": 1216, "v12_mean_update_rms": 0.5209882357437378, "wall_seconds": 4.417608871999619 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.0033436200014875794, "draft_forward_passes": 19, "draft_seconds": 0.017315874999439984, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4284.818375000214, "method": "parareal_linear_v12", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 2, "selection_seconds": 0.015452231999461219, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.247145923998232, "tokens_per_second": 5.601170901438454, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 0, "v12_correction_rounds": 2, "v12_linear_candidate_scores": 1216, "v12_mean_update_rms": 0.5209882357437378, "wall_seconds": 4.284818375000214 }, { "acceptance_rate": 0.05405405405405406, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 0, "context_seconds": 0.0033417450003980775, "draft_forward_passes": 20, "draft_seconds": 0.017627325999001187, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4468.9576440000565, "method": "parareal_linear_v12", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 74, "repeat": 0, "selection_seconds": 0.015898804000698874, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.430470566000622, "tokens_per_second": 5.370379831686697, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "v12_correction_rounds": 2, "v12_linear_candidate_scores": 1280, "v12_mean_update_rms": 0.42320805429889097, "wall_seconds": 4.4689576440000565 }, { "acceptance_rate": 0.05405405405405406, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 0, "context_seconds": 0.0033448580011281592, "draft_forward_passes": 20, "draft_seconds": 0.017929200998423767, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4560.789374999786, "method": "parareal_linear_v12", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 74, "repeat": 1, "selection_seconds": 0.01627755900062766, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.521643652000421, "tokens_per_second": 5.262246954783157, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "v12_correction_rounds": 2, "v12_linear_candidate_scores": 1280, "v12_mean_update_rms": 0.42320805429889097, "wall_seconds": 4.560789374999786 }, { "acceptance_rate": 0.05405405405405406, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 0, "context_seconds": 0.0033728589996826486, "draft_forward_passes": 20, "draft_seconds": 0.017762027000117087, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4475.151363999885, "method": "parareal_linear_v12", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 74, "repeat": 2, "selection_seconds": 0.01592777100040621, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.436361807000594, "tokens_per_second": 5.362947093380285, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "v12_correction_rounds": 2, "v12_linear_candidate_scores": 1280, "v12_mean_update_rms": 0.42320805429889097, "wall_seconds": 4.475151363999885 }, { "acceptance_rate": 0.1076923076923077, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.0028302519995122566, "draft_forward_passes": 17, "draft_seconds": 0.01497811499848467, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3709.176538000065, "method": "parareal_linear_v12", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 65, "repeat": 0, "selection_seconds": 0.013534207999782666, "selector_pair_scores": 0, "target_forward_passes": 17, "target_seconds": 3.676428121999834, "tokens_per_second": 6.470438857283524, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 0, "v12_correction_rounds": 2, "v12_linear_candidate_scores": 1088, "v12_mean_update_rms": 0.40768313370915243, "wall_seconds": 3.709176538000065 }, { "acceptance_rate": 0.1076923076923077, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.002922952000517398, "draft_forward_passes": 17, "draft_seconds": 0.015632403999006783, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3878.029894000065, "method": "parareal_linear_v12", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 65, "repeat": 1, "selection_seconds": 0.013792693001050793, "selector_pair_scores": 0, "target_forward_passes": 17, "target_seconds": 3.844305724000151, "tokens_per_second": 6.188709384920383, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 0, "v12_correction_rounds": 2, "v12_linear_candidate_scores": 1088, "v12_mean_update_rms": 0.40768313370915243, "wall_seconds": 3.878029894000065 }, { "acceptance_rate": 0.1076923076923077, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.0027561400002014125, "draft_forward_passes": 17, "draft_seconds": 0.015370348999567796, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3744.839551000041, "method": "parareal_linear_v12", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 65, "repeat": 2, "selection_seconds": 0.013630427000862255, "selector_pair_scores": 0, "target_forward_passes": 17, "target_seconds": 3.71170426100025, "tokens_per_second": 6.408819302709756, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 0, "v12_correction_rounds": 2, "v12_linear_candidate_scores": 1088, "v12_mean_update_rms": 0.40768313370915243, "wall_seconds": 3.744839551000041 }, { "acceptance_rate": 0.05333333333333334, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 0, "context_seconds": 0.0032836290006343916, "draft_forward_passes": 20, "draft_seconds": 0.017956513999706658, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4674.317854000037, "method": "parareal_linear_v12", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 75, "repeat": 0, "selection_seconds": 0.016368061999401107, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.635149732001082, "tokens_per_second": 5.134439023966257, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "v12_correction_rounds": 2, "v12_linear_candidate_scores": 1280, "v12_mean_update_rms": 0.425190139064085, "wall_seconds": 4.674317854000037 }, { "acceptance_rate": 0.05333333333333334, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 0, "context_seconds": 0.0034166280006502348, "draft_forward_passes": 20, "draft_seconds": 0.017786266000257456, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4676.957423999738, "method": "parareal_linear_v12", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 75, "repeat": 1, "selection_seconds": 0.0159214899986182, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.638225522999619, "tokens_per_second": 5.131541261599764, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "v12_correction_rounds": 2, "v12_linear_candidate_scores": 1280, "v12_mean_update_rms": 0.425190139064085, "wall_seconds": 4.676957423999738 }, { "acceptance_rate": 0.05333333333333334, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 0, "context_seconds": 0.0034196579995295906, "draft_forward_passes": 20, "draft_seconds": 0.01806513600058679, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4709.783152, "method": "parareal_linear_v12", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 75, "repeat": 2, "selection_seconds": 0.016082138999081508, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.670574014000067, "tokens_per_second": 5.09577601036864, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "v12_correction_rounds": 2, "v12_linear_candidate_scores": 1280, "v12_mean_update_rms": 0.425190139064085, "wall_seconds": 4.709783152 } ], "simple_parareal_v14": [ { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.00355254299893204, "draft_forward_passes": 21, "draft_seconds": 0.019801582001491624, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4701.156849999734, "method": "simple_parareal_v14", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 0, "selection_seconds": 0.0026457540016053827, "selector_pair_scores": 0, "target_forward_passes": 21, "target_seconds": 4.673503082999105, "tokens_per_second": 5.105126411598319, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 0, "v14_damping": 1.0, "v14_estimator_ops": 504, "v14_fast_argmax_positions": 84, "v14_mean_abs_residual": 2.904665985446223, "v14_switch_threshold": 0.0, "v14_switched_positions": 0, "wall_seconds": 4.701156849999734 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.003564313999049773, "draft_forward_passes": 21, "draft_seconds": 0.01956791999782581, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4576.86222700022, "method": "simple_parareal_v14", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 1, "selection_seconds": 0.002576881000550202, "selector_pair_scores": 0, "target_forward_passes": 21, "target_seconds": 4.549522990000241, "tokens_per_second": 5.24376719456774, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 0, "v14_damping": 1.0, "v14_estimator_ops": 504, "v14_fast_argmax_positions": 84, "v14_mean_abs_residual": 2.904665985446223, "v14_switch_threshold": 0.0, "v14_switched_positions": 0, "wall_seconds": 4.57686222700022 }, { "acceptance_rate": 0.038461538461538464, "accepted_draft_tokens": 3, "boltzmann_candidate_scores": 0, "context_seconds": 0.0036445009995986766, "draft_forward_passes": 21, "draft_seconds": 0.02043707800021366, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4745.791663000091, "method": "simple_parareal_v14", "new_tokens": 24, "prompt": "The capital of Austria is", "prompt_index": 0, "proposed_draft_tokens": 78, "repeat": 2, "selection_seconds": 0.002713404000587616, "selector_pair_scores": 0, "target_forward_passes": 21, "target_seconds": 4.717369929000597, "tokens_per_second": 5.05711200664637, "tokens_per_target_pass": 1.1428571428571428, "total_guidance_scores": 0, "v14_damping": 1.0, "v14_estimator_ops": 504, "v14_fast_argmax_positions": 84, "v14_mean_abs_residual": 2.904665985446223, "v14_switch_threshold": 0.0, "v14_switched_positions": 0, "wall_seconds": 4.745791663000091 }, { "acceptance_rate": 0.10606060606060606, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.0031668910005464568, "draft_forward_passes": 18, "draft_seconds": 0.016776834999745915, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4303.10769100015, "method": "simple_parareal_v14", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 66, "repeat": 0, "selection_seconds": 0.00219785299987052, "selector_pair_scores": 0, "target_forward_passes": 18, "target_seconds": 4.279622903000472, "tokens_per_second": 5.5773644824635555, "tokens_per_target_pass": 1.3333333333333333, "total_guidance_scores": 0, "v14_damping": 1.0, "v14_estimator_ops": 432, "v14_fast_argmax_positions": 72, "v14_mean_abs_residual": 2.8808834901662763, "v14_switch_threshold": 0.0, "v14_switched_positions": 0, "wall_seconds": 4.30310769100015 }, { "acceptance_rate": 0.10606060606060606, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.0031204619995151006, "draft_forward_passes": 18, "draft_seconds": 0.01731982199999038, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4640.013465000266, "method": "simple_parareal_v14", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 66, "repeat": 1, "selection_seconds": 0.002358891999847401, "selector_pair_scores": 0, "target_forward_passes": 18, "target_seconds": 4.615791603000616, "tokens_per_second": 5.172398783114011, "tokens_per_target_pass": 1.3333333333333333, "total_guidance_scores": 0, "v14_damping": 1.0, "v14_estimator_ops": 432, "v14_fast_argmax_positions": 72, "v14_mean_abs_residual": 2.8808834901662763, "v14_switch_threshold": 0.0, "v14_switched_positions": 0, "wall_seconds": 4.640013465000266 }, { "acceptance_rate": 0.10606060606060606, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.0030398949993468705, "draft_forward_passes": 18, "draft_seconds": 0.01689837600042665, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4384.295101000134, "method": "simple_parareal_v14", "new_tokens": 24, "prompt": "Speculative decoding can improve language model inference by", "prompt_index": 1, "proposed_draft_tokens": 66, "repeat": 2, "selection_seconds": 0.002197075000822224, "selector_pair_scores": 0, "target_forward_passes": 18, "target_seconds": 4.360769094001171, "tokens_per_second": 5.474084076713992, "tokens_per_target_pass": 1.3333333333333333, "total_guidance_scores": 0, "v14_damping": 1.0, "v14_estimator_ops": 432, "v14_fast_argmax_positions": 72, "v14_mean_abs_residual": 2.8808834901662763, "v14_switch_threshold": 0.0, "v14_switched_positions": 0, "wall_seconds": 4.384295101000134 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.003292187001079583, "draft_forward_passes": 19, "draft_seconds": 0.01811086100042303, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4238.053848000163, "method": "simple_parareal_v14", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 0, "selection_seconds": 0.0023968680002326437, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.212743753999803, "tokens_per_second": 5.6629766540897135, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 0, "v14_damping": 1.0, "v14_estimator_ops": 456, "v14_fast_argmax_positions": 76, "v14_mean_abs_residual": 3.3376644179121158, "v14_switch_threshold": 0.0, "v14_switched_positions": 0, "wall_seconds": 4.238053848000163 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.0033965700004046084, "draft_forward_passes": 19, "draft_seconds": 0.018368522999480774, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4347.38794399982, "method": "simple_parareal_v14", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 1, "selection_seconds": 0.002491295999789145, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.3216277659994375, "tokens_per_second": 5.520556322360035, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 0, "v14_damping": 1.0, "v14_estimator_ops": 456, "v14_fast_argmax_positions": 76, "v14_mean_abs_residual": 3.3376644179121158, "v14_switch_threshold": 0.0, "v14_switched_positions": 0, "wall_seconds": 4.34738794399982 }, { "acceptance_rate": 0.08571428571428572, "accepted_draft_tokens": 6, "boltzmann_candidate_scores": 0, "context_seconds": 0.003302042000086658, "draft_forward_passes": 19, "draft_seconds": 0.018156473000544793, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4262.909221999962, "method": "simple_parareal_v14", "new_tokens": 24, "prompt": "A digital twin represents", "prompt_index": 2, "proposed_draft_tokens": 70, "repeat": 2, "selection_seconds": 0.002490749000571668, "selector_pair_scores": 0, "target_forward_passes": 19, "target_seconds": 4.2374942740016195, "tokens_per_second": 5.629958028695787, "tokens_per_target_pass": 1.263157894736842, "total_guidance_scores": 0, "v14_damping": 1.0, "v14_estimator_ops": 456, "v14_fast_argmax_positions": 76, "v14_mean_abs_residual": 3.3376644179121158, "v14_switch_threshold": 0.0, "v14_switched_positions": 0, "wall_seconds": 4.262909221999962 }, { "acceptance_rate": 0.05405405405405406, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 0, "context_seconds": 0.00328836500102625, "draft_forward_passes": 20, "draft_seconds": 0.018490224999368365, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4446.584711999549, "method": "simple_parareal_v14", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 74, "repeat": 0, "selection_seconds": 0.0023682520004513208, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.420968963000178, "tokens_per_second": 5.3974008265790205, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "v14_damping": 1.0, "v14_estimator_ops": 480, "v14_fast_argmax_positions": 80, "v14_mean_abs_residual": 3.060230193579325, "v14_switch_threshold": 0.0, "v14_switched_positions": 0, "wall_seconds": 4.446584711999549 }, { "acceptance_rate": 0.05405405405405406, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 0, "context_seconds": 0.003439092000007804, "draft_forward_passes": 20, "draft_seconds": 0.018779834000724804, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4532.1038840002075, "method": "simple_parareal_v14", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 74, "repeat": 1, "selection_seconds": 0.0024738999995861377, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.505825838999499, "tokens_per_second": 5.295553812155048, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "v14_damping": 1.0, "v14_estimator_ops": 480, "v14_fast_argmax_positions": 80, "v14_mean_abs_residual": 3.060230193579325, "v14_switch_threshold": 0.0, "v14_switched_positions": 0, "wall_seconds": 4.5321038840002075 }, { "acceptance_rate": 0.05405405405405406, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 0, "context_seconds": 0.0033189570008289593, "draft_forward_passes": 20, "draft_seconds": 0.018784764000429277, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4500.85202199989, "method": "simple_parareal_v14", "new_tokens": 24, "prompt": "Reliable software systems should", "prompt_index": 3, "proposed_draft_tokens": 74, "repeat": 2, "selection_seconds": 0.002511467000203993, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.474723644000733, "tokens_per_second": 5.332323720639884, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "v14_damping": 1.0, "v14_estimator_ops": 480, "v14_fast_argmax_positions": 80, "v14_mean_abs_residual": 3.060230193579325, "v14_switch_threshold": 0.0, "v14_switched_positions": 0, "wall_seconds": 4.50085202199989 }, { "acceptance_rate": 0.1076923076923077, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.002797780000946659, "draft_forward_passes": 17, "draft_seconds": 0.015868391998992593, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3687.2341809998943, "method": "simple_parareal_v14", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 65, "repeat": 0, "selection_seconds": 0.00204594700016969, "selector_pair_scores": 0, "target_forward_passes": 17, "target_seconds": 3.6651519020006162, "tokens_per_second": 6.508943783302569, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 0, "v14_damping": 1.0, "v14_estimator_ops": 408, "v14_fast_argmax_positions": 68, "v14_mean_abs_residual": 2.5207495514920133, "v14_switch_threshold": 0.0, "v14_switched_positions": 0, "wall_seconds": 3.6872341809998943 }, { "acceptance_rate": 0.1076923076923077, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.002772139000626339, "draft_forward_passes": 17, "draft_seconds": 0.01582063400064726, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3746.550189000118, "method": "simple_parareal_v14", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 65, "repeat": 1, "selection_seconds": 0.0020738419993904245, "selector_pair_scores": 0, "target_forward_passes": 17, "target_seconds": 3.72461166300036, "tokens_per_second": 6.405893098793677, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 0, "v14_damping": 1.0, "v14_estimator_ops": 408, "v14_fast_argmax_positions": 68, "v14_mean_abs_residual": 2.5207495514920133, "v14_switch_threshold": 0.0, "v14_switched_positions": 0, "wall_seconds": 3.746550189000118 }, { "acceptance_rate": 0.1076923076923077, "accepted_draft_tokens": 7, "boltzmann_candidate_scores": 0, "context_seconds": 0.002874350002002757, "draft_forward_passes": 17, "draft_seconds": 0.016887146000044595, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 3990.959803000351, "method": "simple_parareal_v14", "new_tokens": 24, "prompt": "Water distribution networks must", "prompt_index": 4, "proposed_draft_tokens": 65, "repeat": 2, "selection_seconds": 0.0021637159998135758, "selector_pair_scores": 0, "target_forward_passes": 17, "target_seconds": 3.967669295000178, "tokens_per_second": 6.013591011855624, "tokens_per_target_pass": 1.411764705882353, "total_guidance_scores": 0, "v14_damping": 1.0, "v14_estimator_ops": 408, "v14_fast_argmax_positions": 68, "v14_mean_abs_residual": 2.5207495514920133, "v14_switch_threshold": 0.0, "v14_switched_positions": 0, "wall_seconds": 3.990959803000351 }, { "acceptance_rate": 0.05333333333333334, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 0, "context_seconds": 0.0034516589989834756, "draft_forward_passes": 20, "draft_seconds": 0.019279156000720832, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4719.947376000164, "method": "simple_parareal_v14", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 75, "repeat": 0, "selection_seconds": 0.002589095000530506, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.693131102999359, "tokens_per_second": 5.084802453949895, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "v14_damping": 1.0, "v14_estimator_ops": 480, "v14_fast_argmax_positions": 80, "v14_mean_abs_residual": 3.0506469932475158, "v14_switch_threshold": 0.0, "v14_switched_positions": 0, "wall_seconds": 4.719947376000164 }, { "acceptance_rate": 0.05333333333333334, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 0, "context_seconds": 0.003429350000260456, "draft_forward_passes": 20, "draft_seconds": 0.019043017000967666, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4828.850199000044, "method": "simple_parareal_v14", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 75, "repeat": 1, "selection_seconds": 0.0025282320002588676, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.802269233000061, "tokens_per_second": 4.970127258238391, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "v14_damping": 1.0, "v14_estimator_ops": 480, "v14_fast_argmax_positions": 80, "v14_mean_abs_residual": 3.0506469932475158, "v14_switch_threshold": 0.0, "v14_switched_positions": 0, "wall_seconds": 4.828850199000044 }, { "acceptance_rate": 0.05333333333333334, "accepted_draft_tokens": 4, "boltzmann_candidate_scores": 0, "context_seconds": 0.003400349000003189, "draft_forward_passes": 20, "draft_seconds": 0.01869258099986837, "exact_match": true, "fused_anchor_uses": 0, "jump_candidate_scores": 0, "jump_forward_passes": 0, "latency_ms": 4643.64698899999, "method": "simple_parareal_v14", "new_tokens": 24, "prompt": "Artificial intelligence can support scientific research by", "prompt_index": 5, "proposed_draft_tokens": 75, "repeat": 2, "selection_seconds": 0.0025600809999559715, "selector_pair_scores": 0, "target_forward_passes": 20, "target_seconds": 4.617437928000527, "tokens_per_second": 5.168351525611641, "tokens_per_target_pass": 1.2, "total_guidance_scores": 0, "v14_damping": 1.0, "v14_estimator_ops": 480, "v14_fast_argmax_positions": 80, "v14_mean_abs_residual": 3.0506469932475158, "v14_switch_threshold": 0.0, "v14_switched_positions": 0, "wall_seconds": 4.64364698899999 } ] }, "selected_settings": { "act_margin_threshold": 1.0, "dspark_survival_floor": 0.35, "v10": { "margin_cutoff": 0.35, "margin_slope": 1.5, "stochastic_budget": 2, "temperature": 0.16 }, "v11": { "early_position_bias": 0.15, "gate_floor": 0.16, "mobs_budget": 3, "pair_weight": 1.0, "temperature": 0.04, "top_k": 8 }, "v12": { "correction_rounds": 2, "damping": 0.75, "top_k": 8 }, "v13": { "policy_source": "V10 selected config", "routing_budget": 1, "top2_location": "Torch drafter before NumPy transfer" }, "v14": { "coefficients": { "coarse_gap_weight": 2.884912113376932, "intercept": 1.4487449748772188, "position_weight": -1.0593768039151394 }, "damping": 1.0, "max_corrections": 1, "switch_threshold": 0.0 }, "verify_widths": {} }, "study": "LFM2.5 All-14 speculative decoding study", "summary": { "boltzmann_gated_mobs_v11": { "all_exact": true, "draft_seconds_median": 0.01748556999950779, "latency_seconds_median": 4.464789268499999, "mean_acceptance_rate": 0.07421935421935422, "mean_draft_forward_passes": 19.166666666666668, "mean_guidance_work": 135.33333333333334, "mean_target_forward_passes": 19.166666666666668, "mean_tokens_per_target_pass": 1.258518846134945, "selection_seconds_median": 0.00583111299988559, "speedup_vs_normal": 0.9888687210433905, "target_seconds_median": 4.437326398998266, "tokens_per_second_median": 5.3754660168823545, "v11_fast_argmax_positions": 72.16666666666667 }, "boltzmann_v10": { "all_exact": true, "draft_seconds_median": 0.01795636049996574, "latency_seconds_median": 4.644041660500079, "mean_acceptance_rate": 0.0765966197545145, "mean_draft_forward_passes": 19.166666666666668, "mean_guidance_work": 37.666666666666664, "mean_target_forward_passes": 19.166666666666668, "mean_tokens_per_target_pass": 1.2686716791979948, "selection_seconds_median": 0.004698019500665396, "speedup_vs_normal": 0.9506963193729104, "target_seconds_median": 4.615151720999847, "tokens_per_second_median": 5.167961781389962, "v10_fast_argmax_positions": 57.833333333333336 }, "dflash": { "all_exact": true, "draft_seconds_median": 0.017471871000907413, "latency_seconds_median": 4.486416377499836, "mean_acceptance_rate": 0.07421935421935422, "mean_draft_forward_passes": 19.166666666666668, "mean_guidance_work": 0.0, "mean_target_forward_passes": 19.166666666666668, "mean_tokens_per_target_pass": 1.258518846134945, "selection_seconds_median": 0.0003873834998557868, "speedup_vs_normal": 0.9840885686395442, "target_seconds_median": 4.464233998500049, "tokens_per_second_median": 5.349481226125415 }, "dflash2": { "all_exact": true, "draft_seconds_median": 0.017428329999347625, "latency_seconds_median": 4.580972383000017, "mean_acceptance_rate": 0.07107774854253729, "mean_draft_forward_passes": 19.333333333333332, "mean_guidance_work": 3866.6666666666665, "mean_target_forward_passes": 19.333333333333332, "mean_tokens_per_target_pass": 1.2478254459678608, "selection_seconds_median": 0.008189913500473267, "speedup_vs_normal": 0.9637851538952292, "target_seconds_median": 4.550487350499452, "tokens_per_second_median": 5.239112363543155 }, "dflash3_mobs": { "all_exact": true, "draft_seconds_median": 0.017244260500092423, "latency_seconds_median": 4.533612524000091, "mean_acceptance_rate": 0.07107774854253729, "mean_draft_forward_passes": 19.333333333333332, "mean_guidance_work": 773.3333333333334, "mean_target_forward_passes": 19.333333333333332, "mean_tokens_per_target_pass": 1.2478254459678608, "selection_seconds_median": 0.007564432499975737, "speedup_vs_normal": 0.9739943089872078, "target_seconds_median": 4.504199763000315, "tokens_per_second_median": 5.2946090792245935 }, "dflash4_jump_mobs": { "all_exact": true, "draft_seconds_median": 0.01733637049983372, "latency_seconds_median": 4.505064758999879, "mean_acceptance_rate": 0.06454455570797459, "mean_draft_forward_passes": 19.666666666666668, "mean_guidance_work": 944.0, "mean_target_forward_passes": 19.666666666666668, "mean_tokens_per_target_pass": 1.2230576441102756, "selection_seconds_median": 0.01308126600019932, "speedup_vs_normal": 0.9800420653810923, "target_seconds_median": 4.470823709499655, "tokens_per_second_median": 5.3274845340568655 }, "dflash5_fused_jump_mobs": { "all_exact": true, "draft_seconds_median": 0.01735558499990475, "latency_seconds_median": 4.505628296000168, "mean_acceptance_rate": 0.07107774854253729, "mean_draft_forward_passes": 19.333333333333332, "mean_guidance_work": 928.0, "mean_target_forward_passes": 19.333333333333332, "mean_tokens_per_target_pass": 1.2478254459678608, "selection_seconds_median": 0.013150039500942512, "speedup_vs_normal": 0.9798939709362487, "target_seconds_median": 4.471018944499747, "tokens_per_second_median": 5.3266794963014945 }, "dflash6_bmobs": { "all_exact": true, "draft_seconds_median": 0.01774667949916875, "latency_seconds_median": 4.640960819499924, "mean_acceptance_rate": 0.06477358390401869, "mean_draft_forward_passes": 19.666666666666668, "mean_guidance_work": 705.3333333333334, "mean_target_forward_passes": 19.666666666666668, "mean_tokens_per_target_pass": 1.2242272347535506, "selection_seconds_median": 0.010490578000599271, "speedup_vs_normal": 0.9513778139163739, "target_seconds_median": 4.606647188999887, "tokens_per_second_median": 5.171666368946552 }, "dflash6_boltzmann": { "all_exact": true, "draft_seconds_median": 0.017645633500023905, "latency_seconds_median": 4.54780940149999, "mean_acceptance_rate": 0.07919537235326708, "mean_draft_forward_passes": 19.0, "mean_guidance_work": 608.0, "mean_target_forward_passes": 19.0, "mean_tokens_per_target_pass": 1.2781954887218046, "selection_seconds_median": 0.007728781999958301, "speedup_vs_normal": 0.9708502284700562, "target_seconds_median": 4.517905775998088, "tokens_per_second_median": 5.277517934955759 }, "dflash7_act": { "act_mean_verified_drafts_per_block": 1.3348506069094304, "act_trimmed_draft_tokens": 46.166666666666664, "all_exact": true, "draft_seconds_median": 0.019069493499046075, "latency_seconds_median": 4.069323076500041, "mean_acceptance_rate": 0.20076554671382257, "mean_draft_forward_passes": 19.333333333333332, "mean_guidance_work": 0.0, "mean_target_forward_passes": 19.333333333333332, "mean_tokens_per_target_pass": 1.2479925303454715, "selection_seconds_median": 0.002520233999575794, "speedup_vs_normal": 1.0849628874346593, "target_seconds_median": 4.043520783499616, "tokens_per_second_median": 5.897831539084201 }, "dspark_v9": { "all_exact": true, "draft_seconds_median": 0.01931573399951958, "dspark_mean_confidence": 0.4549896783455034, "latency_seconds_median": 4.182915717500009, "mean_acceptance_rate": 0.19625139353400223, "mean_draft_forward_passes": 19.5, "mean_guidance_work": 624.0, "mean_target_forward_passes": 19.5, "mean_tokens_per_target_pass": 1.238468720821662, "selection_seconds_median": 0.006802876000392644, "speedup_vs_normal": 1.0556715789160154, "target_seconds_median": 4.153131784999459, "tokens_per_second_median": 5.738604707269913 }, "minop_v13": { "all_exact": true, "draft_seconds_median": 0.018478784000308224, "latency_seconds_median": 4.554242015, "mean_acceptance_rate": 0.07919537235326708, "mean_draft_forward_passes": 19.0, "mean_guidance_work": 26.666666666666668, "mean_target_forward_passes": 19.0, "mean_tokens_per_target_pass": 1.2781954887218046, "selection_seconds_median": 0.0021608249999189866, "speedup_vs_normal": 0.9694546218252974, "target_seconds_median": 4.528083417000062, "tokens_per_second_median": 5.269931451601406, "v13_candidate_scores": 26.666666666666668, "v13_fast_argmax_positions": 62.666666666666664, "v13_routed_positions": 13.333333333333334, "v13_top2_values_transferred": 152.0 }, "normal": { "all_exact": true, "draft_seconds_median": 0.0, "latency_seconds_median": 4.415265098500186, "mean_acceptance_rate": 0.0, "mean_draft_forward_passes": 0.0, "mean_guidance_work": 0.0, "mean_target_forward_passes": 24.0, "mean_tokens_per_target_pass": 1.0, "selection_seconds_median": 0.0, "speedup_vs_normal": 1.0, "target_seconds_median": 4.414389316001461, "tokens_per_second_median": 5.435975375184796 }, "parareal_linear_v12": { "all_exact": true, "draft_seconds_median": 0.017694676499559137, "latency_seconds_median": 4.472054503999971, "mean_acceptance_rate": 0.07421935421935422, "mean_draft_forward_passes": 19.166666666666668, "mean_guidance_work": 1226.6666666666667, "mean_target_forward_passes": 19.166666666666668, "mean_tokens_per_target_pass": 1.258518846134945, "selection_seconds_median": 0.01577956400046787, "speedup_vs_normal": 0.9872494064326132, "target_seconds_median": 4.433416186500608, "tokens_per_second_median": 5.3666634625334915, "v12_mean_update_rms": 0.44600832807802476 }, "simple_parareal_v14": { "all_exact": true, "draft_seconds_median": 0.01842937399942457, "latency_seconds_median": 4.473718366999719, "mean_acceptance_rate": 0.07421935421935422, "mean_draft_forward_passes": 19.166666666666668, "mean_guidance_work": 460.0, "mean_target_forward_passes": 19.166666666666668, "mean_tokens_per_target_pass": 1.258518846134945, "selection_seconds_median": 0.002482324500078903, "speedup_vs_normal": 0.9869180603907856, "target_seconds_median": 4.447846303500455, "tokens_per_second_median": 5.364862273609452, "v14_estimator_ops": 460.0, "v14_fast_argmax_positions": 76.66666666666667, "v14_mean_abs_residual": 2.959140105307245, "v14_switched_positions": 0.0 } }, "winner": "dflash7_act" }