| """Conditional cost sensitivities, never a live speed forecast. |
| |
| Known costs are solo SM86 proxies. Unknown K2/eager/peer/GB10/proxy work is |
| explicitly null; every displayed rate names an assumed extra cost. Fixed5 |
| replay populations and historical adaptive windows stay separate. |
| """ |
| from __future__ import annotations |
|
|
| import argparse |
| import json |
| import math |
| from pathlib import Path |
| from typing import Any |
|
|
|
|
| def wire_bytes(batch: int, depth: int) -> int: |
| if batch not in range(1,5) or depth not in (4,5): |
| raise ValueError('B1..B4 and fixed depth4/5 required') |
| return 32 + 32 + 8*batch + 8*batch*depth + 16 + 4*batch + 8*batch*depth*32 + 4 |
|
|
|
|
| def rate(retained: float, base_ms: float, delta: float, known_ms: float, |
| unknown_ms: float) -> float: |
| values = (retained, base_ms, delta, known_ms, unknown_ms) |
| if not all(math.isfinite(x) for x in values) or retained <= 0 or base_ms <= 0 or min(known_ms,unknown_ms) < 0: |
| raise ValueError('invalid cost sensitivity inputs') |
| return (retained+delta)*1000/(base_ms+known_ms+unknown_ms) |
|
|
|
|
| def estimate(producer: dict[str, Any], gpu: dict[str, Any], pp2: dict[str, Any]) -> dict[str, Any]: |
| if not producer['complete'] or not gpu['complete']: |
| raise ValueError('incomplete GPU receipts cannot price a candidate') |
| row = next(r for r in producer['rows'] if r['batch']==1 and r['depth']==5) |
| parts: dict[str, float] = { |
| 'selector_addition_solo_synthetic_p95_ms': row['sampled']['p95_ms'], |
| 'sparse1280_D2H_solo_p95_ms': gpu['sparse1280_d2h']['p95_ms'], |
| 'sparse1280_H2D_solo_p95_ms': gpu['sparse1280_h2d']['p95_ms'], |
| 'full_dense_p_standard_verifier_solo_p95_ms': gpu['dense_p_verifier']['p95_ms'], |
| 'dense_q_expand_solo_p95_ms': gpu['dense_q_expansion']['p95_ms'], |
| 'two_serialized_framed_replies_40gbps_ms': 2*wire_bytes(1,5)*8/(pp2['wire_gbps']*1e6), |
| } |
| known = sum(parts.values()) |
| fixed: dict[str, Any] = {} |
| for key, view in pp2['views'].items(): |
| retained = 1+view['greedy_EA'] |
| delta = view['delta_EA'] |
| interval = view['CI95_delta'] |
| base = pp2['fixed5_archived_price_ms'] |
| fixed[key] = { |
| 'retained_greedy': retained, 'retained_sampled': retained+delta, |
| 'archived_price_ms_not_live_mean': base, |
| 'break_even_total_added_ms': base*delta/retained, |
| 'break_even_total_added_ms_CI_lower': None if interval is None else base*interval[0]/retained, |
| 'remaining_unmeasured_budget_ms_CI_lower': None if interval is None else base*interval[0]/retained-known, |
| 'true_net_rate_change': None, |
| 'sensitivity': [{'assumed_unmeasured_ms': u, |
| 'conditional_tps': rate(retained,base,delta,known,u), |
| 'conditional_pct_change': (rate(retained,base,delta,known,u)/(retained*1000/base)-1)*100} |
| for u in (0.,.1,.5,1.,2.,3.)], |
| } |
| historical = [] |
| for cell in pp2['w8_historical_adaptive']: |
| name = cell['cell'] |
| |
| |
| view_key = 'context:<8K' if name.startswith('short') else 'context:16-32K' if name.startswith('32k') else 'all' |
| view = pp2['views'][view_key] |
| relative_gain = view['delta_EA']/(1+view['greedy_EA']) |
| retained = 1+cell['historical_accepted_per_round'] |
| base = cell['historical_step_p50_ms'] |
| historical.append({'cell':name, 'historical_adaptive_baseline_tps':retained*1000/base, |
| 'acceptance_proxy':view_key, 'measured_acceptance_at150k':None if name.startswith('150k') else 'different fixed-depth replay population', |
| 'extra_assumption':'transplant relative fixed5 token-count gain to adaptive window; 150K extrapolates pooled <=47K anchors', |
| 'expected_live_tps':None, |
| 'sensitivity':[{'assumed_unmeasured_ms':u, |
| 'conditional_tps':rate(retained,base,retained*relative_gain,known,u), |
| 'conditional_pct_change':(rate(retained,base,retained*relative_gain,known,u)/(retained*1000/base)-1)*100} |
| for u in (0.,.5,1.,2.,3.)]}) |
| return {'schema':'pp3.cost-sensitivity.v1','classification':'ESTIMATED conditional sensitivities from MEASURED proxies', |
| 'decision':'STOP: absolute selector target missed; full K2/pair/GB10 costs and long-context acceptance unmeasured', |
| 'single_launch_goal_met':False, 'selector_absolute_005ms_goal_met':row['sampled']['p95_ms']<=.05, |
| 'parts':parts,'known_proxy_sum_ms':known,'additional_unmeasured_ms':None, |
| 'unknown_components':['extra Markov materialization and duplicate greedy selector work', |
| 'eager graph/CPU launch gaps in sampled full K2 path','full-vocabulary worker peer copies and completion waits', |
| 'CPU CRC/q validation/proxy queue and registered pool overhead','full dense-q slot index-copy/TP synchronization', |
| 'GB10 H2D/expansion/rejection differences','policy-induced future occupancy and150K acceptance'], |
| 'fixed5_replay_sensitivities':fixed,'historical_adaptive_sensitivities':historical, |
| 'interpretation':'No unknown component is assigned zero in the decision. u=0 rows are illustrative only; no speedup is claimed.'} |
|
|
|
|
| def main() -> None: |
| p = argparse.ArgumentParser(description=__doc__) |
| p.add_argument('--directory',type=Path,default=Path(__file__).parent) |
| p.add_argument('--out',type=Path,required=True) |
| a=p.parse_args() |
| files=('producer-final.json','gpu-qualification.json','pp2-frozen.json') |
| result=estimate(*(json.loads((a.directory/name).read_text()) for name in files)) |
| a.out.write_text(json.dumps(result,indent=2)+'\n') |
|
|
|
|
| if __name__ == '__main__': |
| main() |
|
|
