1{2 "actions": [3 "A1",4 "A2|budget_tokens=512",5 "A3|budget_tokens=512|blind",6 "A3|budget_tokens=512|aware",7 "A4|budget_tokens=512|blind",8 "A5|budget_tokens=512|blind",9 "A6|budget_tokens=512|max_fetch_bytes=2000|max_lookups=6|blind",10 "A7|codec=zstd",11 "A8|budget_tokens=512|blind",12 "B0",13 "B1",14 "B2|bits=8",15 "B3|keep_frac=0.25",16 "B5|keep_frac=0.25|aware",17 "B9|bits=8|keep_frac=0.25|policy=attention|aware"18 ],19 "cost_col": "cost_amortized_ms",20 "tau": 0.9,21 "table": {22 "exec|short|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|aware": {23 "chosen": "A3|budget_tokens=512|aware",24 "best_static": "A3|budget_tokens=512|aware",25 "candidates": [26 {27 "action": "A3|budget_tokens=512|aware",28 "q_mean": 0.5,29 "q_lo": 0.0,30 "cost": 1329.8871670267545,31 "n": 232 },33 {34 "action": "B5|keep_frac=0.25|aware",35 "q_mean": 0.0,36 "q_lo": 0.0,37 "cost": 1476.8745410256088,38 "n": 239 },40 {41 "action": "B9|bits=8|keep_frac=0.25|policy=attention|aware",42 "q_mean": 0.0,43 "q_lo": 0.0,44 "cost": 1518.9385620178655,45 "n": 246 }47 ]48 },49 "exec|short|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|blind": {50 "chosen": "A3|budget_tokens=512|blind",51 "best_static": "A1",52 "candidates": [53 {54 "action": "A1",55 "q_mean": 1.0,56 "q_lo": 1.0,57 "cost": 1631.058249855414,58 "n": 259 },60 {61 "action": "A2|budget_tokens=512",62 "q_mean": 1.0,63 "q_lo": 1.0,64 "cost": 1417.3036039574072,65 "n": 266 },67 {68 "action": "A3|budget_tokens=512|blind",69 "q_mean": 1.0,70 "q_lo": 1.0,71 "cost": 1332.8896454768255,72 "n": 273 },74 {75 "action": "A4|budget_tokens=512|blind",76 "q_mean": 0.5,77 "q_lo": 0.0,78 "cost": 8919.545584125444,79 "n": 280 },81 {82 "action": "A5|budget_tokens=512|blind",83 "q_mean": 0.0,84 "q_lo": 0.0,85 "cost": 8856.286415655632,86 "n": 287 },88 {89 "action": "A6|budget_tokens=512|max_fetch_bytes=2000|max_lookups=6|blind",90 "q_mean": 0.0,91 "q_lo": 0.0,92 "cost": 1261.0194170265459,93 "n": 294 },95 {96 "action": "A7|codec=zstd",97 "q_mean": 1.0,98 "q_lo": 1.0,99 "cost": 1632.7006250503473,100 "n": 2101 },102 {103 "action": "A8|budget_tokens=512|blind",104 "q_mean": 0.0,105 "q_lo": 0.0,106 "cost": 2142.073229479138,107 "n": 2108 },109 {110 "action": "B0",111 "q_mean": 1.0,112 "q_lo": 1.0,113 "cost": 1853.2067709602416,114 "n": 2115 },116 {117 "action": "B1",118 "q_mean": 1.0,119 "q_lo": 1.0,120 "cost": 1809.8062080680393,121 "n": 2122 },123 {124 "action": "B2|bits=8",125 "q_mean": 1.0,126 "q_lo": 1.0,127 "cost": 1990.0899369386025,128 "n": 2129 },130 {131 "action": "B3|keep_frac=0.25",132 "q_mean": 0.0,133 "q_lo": 0.0,134 "cost": 1653.049167478457,135 "n": 2136 }137 ]138 },139 "exec|short|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|aware": {140 "chosen": "A3|budget_tokens=512|aware",141 "best_static": "A3|budget_tokens=512|aware",142 "candidates": [143 {144 "action": "A3|budget_tokens=512|aware",145 "q_mean": 1.0,146 "q_lo": 1.0,147 "cost": 1885.1518545416184,148 "n": 2149 },150 {151 "action": "B5|keep_frac=0.25|aware",152 "q_mean": 0.5,153 "q_lo": 0.0,154 "cost": 1798.074854479637,155 "n": 2156 },157 {158 "action": "B9|bits=8|keep_frac=0.25|policy=attention|aware",159 "q_mean": 0.0,160 "q_lo": 0.0,161 "cost": 1998.245208291337,162 "n": 2163 }164 ]165 },166 "exec|short|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|blind": {167 "chosen": "A3|budget_tokens=512|blind",168 "best_static": "A1",169 "candidates": [170 {171 "action": "A1",172 "q_mean": 1.0,173 "q_lo": 1.0,174 "cost": 2106.2395630870014,175 "n": 2176 },177 {178 "action": "A2|budget_tokens=512",179 "q_mean": 1.0,180 "q_lo": 1.0,181 "cost": 1777.6976449531503,182 "n": 2183 },184 {185 "action": "A3|budget_tokens=512|blind",186 "q_mean": 1.0,187 "q_lo": 1.0,188 "cost": 1548.3912916388363,189 "n": 2190 },191 {192 "action": "A4|budget_tokens=512|blind",193 "q_mean": 0.5,194 "q_lo": 0.0,195 "cost": 7460.223854461219,196 "n": 2197 },198 {199 "action": "A5|budget_tokens=512|blind",200 "q_mean": 0.0,201 "q_lo": 0.0,202 "cost": 6980.376125022303,203 "n": 2204 },205 {206 "action": "A6|budget_tokens=512|max_fetch_bytes=2000|max_lookups=6|blind",207 "q_mean": 0.0,208 "q_lo": 0.0,209 "cost": 1045.0055415276438,210 "n": 2211 },212 {213 "action": "A7|codec=zstd",214 "q_mean": 1.0,215 "q_lo": 1.0,216 "cost": 1939.9121680180542,217 "n": 2218 },219 {220 "action": "A8|budget_tokens=512|blind",221 "q_mean": 0.0,222 "q_lo": 0.0,223 "cost": 6249.4903958868235,224 "n": 2225 },226 {227 "action": "B0",228 "q_mean": 1.0,229 "q_lo": 1.0,230 "cost": 2214.8592923767865,231 "n": 2232 },233 {234 "action": "B1",235 "q_mean": 1.0,236 "q_lo": 1.0,237 "cost": 2213.8642499921843,238 "n": 2239 },240 {241 "action": "B2|bits=8",242 "q_mean": 1.0,243 "q_lo": 1.0,244 "cost": 2394.7104800026864,245 "n": 2246 },247 {248 "action": "B3|keep_frac=0.25",249 "q_mean": 0.0,250 "q_lo": 0.0,251 "cost": 1192.3931669443846,252 "n": 2253 }254 ]255 },256 "public_2wikimqa|medium|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|aware": {257 "chosen": "B5|keep_frac=0.25|aware",258 "best_static": "B5|keep_frac=0.25|aware",259 "candidates": [260 {261 "action": "B5|keep_frac=0.25|aware",262 "q_mean": 0.0,263 "q_lo": 0.0,264 "cost": 1401.3548743678257,265 "n": 2266 },267 {268 "action": "B9|bits=8|keep_frac=0.25|policy=attention|aware",269 "q_mean": 0.0,270 "q_lo": 0.0,271 "cost": 1821.9773949240334,272 "n": 2273 }274 ]275 },276 "public_2wikimqa|medium|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|blind": {277 "chosen": "A1",278 "best_static": "A1",279 "candidates": [280 {281 "action": "A1",282 "q_mean": 0.0,283 "q_lo": 0.0,284 "cost": 958.6374779464677,285 "n": 2286 },287 {288 "action": "A7|codec=zstd",289 "q_mean": 0.0,290 "q_lo": 0.0,291 "cost": 1236.2463335739449,292 "n": 2293 },294 {295 "action": "B0",296 "q_mean": 0.0,297 "q_lo": 0.0,298 "cost": 1379.270581412129,299 "n": 2300 },301 {302 "action": "B1",303 "q_mean": 0.0,304 "q_lo": 0.0,305 "cost": 1497.370915429201,306 "n": 2307 },308 {309 "action": "B2|bits=8",310 "q_mean": 0.0,311 "q_lo": 0.0,312 "cost": 1899.9604169512168,313 "n": 2314 },315 {316 "action": "B3|keep_frac=0.25",317 "q_mean": 0.5,318 "q_lo": 0.0,319 "cost": 1568.692895467393,320 "n": 2321 }322 ]323 },324 "public_2wikimqa|medium|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|aware": {325 "chosen": "B5|keep_frac=0.25|aware",326 "best_static": "B5|keep_frac=0.25|aware",327 "candidates": [328 {329 "action": "B5|keep_frac=0.25|aware",330 "q_mean": 0.0,331 "q_lo": 0.0,332 "cost": 1259.7939987899736,333 "n": 2334 },335 {336 "action": "B9|bits=8|keep_frac=0.25|policy=attention|aware",337 "q_mean": 0.0,338 "q_lo": 0.0,339 "cost": 1319.705936824903,340 "n": 2341 }342 ]343 },344 "public_2wikimqa|medium|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|blind": {345 "chosen": "B3|keep_frac=0.25",346 "best_static": "A1",347 "candidates": [348 {349 "action": "A1",350 "q_mean": 0.0,351 "q_lo": 0.0,352 "cost": 1315.37816679338,353 "n": 2354 },355 {356 "action": "A7|codec=zstd",357 "q_mean": 0.0,358 "q_lo": 0.0,359 "cost": 1215.5741045135073,360 "n": 2361 },362 {363 "action": "B0",364 "q_mean": 0.0,365 "q_lo": 0.0,366 "cost": 1290.58462503599,367 "n": 2368 },369 {370 "action": "B1",371 "q_mean": 0.0,372 "q_lo": 0.0,373 "cost": 1291.3224364165217,374 "n": 2375 },376 {377 "action": "B2|bits=8",378 "q_mean": 0.0,379 "q_lo": 0.0,380 "cost": 1230.071353900712,381 "n": 2382 },383 {384 "action": "B3|keep_frac=0.25",385 "q_mean": 0.0,386 "q_lo": 0.0,387 "cost": 1153.9108945289627,388 "n": 2389 }390 ]391 },392 "public_2wikimqa|short|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|aware": {393 "chosen": "A3|budget_tokens=512|aware",394 "best_static": "A3|budget_tokens=512|aware",395 "candidates": [396 {397 "action": "A3|budget_tokens=512|aware",398 "q_mean": 0.0,399 "q_lo": 0.0,400 "cost": 140.73666697368026,401 "n": 2402 }403 ]404 },405 "public_2wikimqa|short|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|blind": {406 "chosen": "A6|budget_tokens=512|max_fetch_bytes=2000|max_lookups=6|blind",407 "best_static": "A2|budget_tokens=512",408 "candidates": [409 {410 "action": "A2|budget_tokens=512",411 "q_mean": 0.0,412 "q_lo": 0.0,413 "cost": 136.72804203815758,414 "n": 2415 },416 {417 "action": "A3|budget_tokens=512|blind",418 "q_mean": 0.0,419 "q_lo": 0.0,420 "cost": 112.30885348049924,421 "n": 2422 },423 {424 "action": "A4|budget_tokens=512|blind",425 "q_mean": 0.0,426 "q_lo": 0.0,427 "cost": 9116.805187019054,428 "n": 2429 },430 {431 "action": "A5|budget_tokens=512|blind",432 "q_mean": 0.0,433 "q_lo": 0.0,434 "cost": 9143.03629292408,435 "n": 2436 },437 {438 "action": "A6|budget_tokens=512|max_fetch_bytes=2000|max_lookups=6|blind",439 "q_mean": 0.0,440 "q_lo": 0.0,441 "cost": 101.88310500234365,442 "n": 2443 },444 {445 "action": "A8|budget_tokens=512|blind",446 "q_mean": 0.0,447 "q_lo": 0.0,448 "cost": 728.3446455257945,449 "n": 2450 }451 ]452 },453 "public_2wikimqa|short|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|aware": {454 "chosen": "A3|budget_tokens=512|aware",455 "best_static": "A3|budget_tokens=512|aware",456 "candidates": [457 {458 "action": "A3|budget_tokens=512|aware",459 "q_mean": 0.0,460 "q_lo": 0.0,461 "cost": 178.05522889830172,462 "n": 2463 }464 ]465 },466 "public_2wikimqa|short|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|blind": {467 "chosen": "A6|budget_tokens=512|max_fetch_bytes=2000|max_lookups=6|blind",468 "best_static": "A2|budget_tokens=512",469 "candidates": [470 {471 "action": "A2|budget_tokens=512",472 "q_mean": 0.0,473 "q_lo": 0.0,474 "cost": 116.48358352249488,475 "n": 2476 },477 {478 "action": "A3|budget_tokens=512|blind",479 "q_mean": 0.5,480 "q_lo": 0.0,481 "cost": 131.05783401988447,482 "n": 2483 },484 {485 "action": "A4|budget_tokens=512|blind",486 "q_mean": 0.0,487 "q_lo": 0.0,488 "cost": 8600.43547960231,489 "n": 2490 },491 {492 "action": "A5|budget_tokens=512|blind",493 "q_mean": 0.0,494 "q_lo": 0.0,495 "cost": 8467.526854481548,496 "n": 2497 },498 {499 "action": "A6|budget_tokens=512|max_fetch_bytes=2000|max_lookups=6|blind",500 "q_mean": 0.0,501 "q_lo": 0.0,502 "cost": 112.01987491222098,503 "n": 2504 },505 {506 "action": "A8|budget_tokens=512|blind",507 "q_mean": 0.0,508 "q_lo": 0.0,509 "cost": 1496.8782920041122,510 "n": 2511 }512 ]513 },514 "public_qasper|medium|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|aware": {515 "chosen": "B5|keep_frac=0.25|aware",516 "best_static": "B5|keep_frac=0.25|aware",517 "candidates": [518 {519 "action": "B5|keep_frac=0.25|aware",520 "q_mean": 0.24999999999999997,521 "q_lo": 0.125,522 "cost": 722.1069170045666,523 "n": 2524 },525 {526 "action": "B9|bits=8|keep_frac=0.25|policy=attention|aware",527 "q_mean": 0.24999999999999997,528 "q_lo": 0.125,529 "cost": 780.5135424714535,530 "n": 2531 }532 ]533 },534 "public_qasper|medium|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|blind": {535 "chosen": "A1",536 "best_static": "A1",537 "candidates": [538 {539 "action": "A1",540 "q_mean": 0.24999999999999997,541 "q_lo": 0.125,542 "cost": 595.0978129403666,543 "n": 2544 },545 {546 "action": "A7|codec=zstd",547 "q_mean": 0.24999999999999997,548 "q_lo": 0.125,549 "cost": 604.6264575561509,550 "n": 2551 },552 {553 "action": "B0",554 "q_mean": 0.24999999999999997,555 "q_lo": 0.125,556 "cost": 1024.1511670174077,557 "n": 2558 },559 {560 "action": "B1",561 "q_mean": 0.24999999999999997,562 "q_lo": 0.125,563 "cost": 991.0490409820341,564 "n": 2565 },566 {567 "action": "B2|bits=8",568 "q_mean": 0.24999999999999997,569 "q_lo": 0.125,570 "cost": 765.4312709346414,571 "n": 2572 },573 {574 "action": "B3|keep_frac=0.25",575 "q_mean": 0.0625,576 "q_lo": 0.0,577 "cost": 600.5491039832123,578 "n": 2579 }580 ]581 },582 "public_qasper|medium|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|aware": {583 "chosen": "B9|bits=8|keep_frac=0.25|policy=attention|aware",584 "best_static": "B9|bits=8|keep_frac=0.25|policy=attention|aware",585 "candidates": [586 {587 "action": "B5|keep_frac=0.25|aware",588 "q_mean": 0.2,589 "q_lo": 0.0,590 "cost": 711.4740828401409,591 "n": 2592 },593 {594 "action": "B9|bits=8|keep_frac=0.25|policy=attention|aware",595 "q_mean": 0.2625,596 "q_lo": 0.125,597 "cost": 781.3643955742009,598 "n": 2599 }600 ]601 },602 "public_qasper|medium|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|blind": {603 "chosen": "B3|keep_frac=0.25",604 "best_static": "A1",605 "candidates": [606 {607 "action": "A1",608 "q_mean": 0.18749999999999997,609 "q_lo": 0.0,610 "cost": 722.5117080379277,611 "n": 2612 },613 {614 "action": "A7|codec=zstd",615 "q_mean": 0.18749999999999997,616 "q_lo": 0.0,617 "cost": 720.7211870700121,618 "n": 2619 },620 {621 "action": "B0",622 "q_mean": 0.18749999999999997,623 "q_lo": 0.0,624 "cost": 1075.719332555309,625 "n": 2626 },627 {628 "action": "B1",629 "q_mean": 0.18749999999999997,630 "q_lo": 0.0,631 "cost": 866.128791065421,632 "n": 2633 },634 {635 "action": "B2|bits=8",636 "q_mean": 0.0625,637 "q_lo": 0.0,638 "cost": 609.4338131370023,639 "n": 2640 },641 {642 "action": "B3|keep_frac=0.25",643 "q_mean": 0.0,644 "q_lo": 0.0,645 "cost": 524.0459991036914,646 "n": 2647 }648 ]649 },650 "public_qasper|short|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|aware": {651 "chosen": "A3|budget_tokens=512|aware",652 "best_static": "A3|budget_tokens=512|aware",653 "candidates": [654 {655 "action": "A3|budget_tokens=512|aware",656 "q_mean": 0.2932692307692307,657 "q_lo": 0.125,658 "cost": 209.8669374245219,659 "n": 2660 }661 ]662 },663 "public_qasper|short|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|blind": {664 "chosen": "A4|budget_tokens=512|blind",665 "best_static": "A4|budget_tokens=512|blind",666 "candidates": [667 {668 "action": "A2|budget_tokens=512",669 "q_mean": 0.0625,670 "q_lo": 0.0,671 "cost": 156.89179295441136,672 "n": 2673 },674 {675 "action": "A3|budget_tokens=512|blind",676 "q_mean": 0.0,677 "q_lo": 0.0,678 "cost": 136.68591663008556,679 "n": 2680 },681 {682 "action": "A4|budget_tokens=512|blind",683 "q_mean": 0.2932692307692307,684 "q_lo": 0.125,685 "cost": 8117.365646059625,686 "n": 2687 },688 {689 "action": "A5|budget_tokens=512|blind",690 "q_mean": 0.0,691 "q_lo": 0.0,692 "cost": 8089.817729953211,693 "n": 2694 },695 {696 "action": "A6|budget_tokens=512|max_fetch_bytes=2000|max_lookups=6|blind",697 "q_mean": 0.0,698 "q_lo": 0.0,699 "cost": 118.74318757327273,700 "n": 2701 },702 {703 "action": "A8|budget_tokens=512|blind",704 "q_mean": 0.0,705 "q_lo": 0.0,706 "cost": 519.0687498543411,707 "n": 2708 }709 ]710 },711 "public_qasper|short|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|aware": {712 "chosen": "A3|budget_tokens=512|aware",713 "best_static": "A3|budget_tokens=512|aware",714 "candidates": [715 {716 "action": "A3|budget_tokens=512|aware",717 "q_mean": 0.19763513513513511,718 "q_lo": 0.125,719 "cost": 298.25204209191725,720 "n": 2721 }722 ]723 },724 "public_qasper|short|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|blind": {725 "chosen": "A6|budget_tokens=512|max_fetch_bytes=2000|max_lookups=6|blind",726 "best_static": "A2|budget_tokens=512",727 "candidates": [728 {729 "action": "A2|budget_tokens=512",730 "q_mean": 0.0,731 "q_lo": 0.0,732 "cost": 163.6342714773491,733 "n": 2734 },735 {736 "action": "A3|budget_tokens=512|blind",737 "q_mean": 0.0,738 "q_lo": 0.0,739 "cost": 131.36837491765618,740 "n": 2741 },742 {743 "action": "A4|budget_tokens=512|blind",744 "q_mean": 0.0,745 "q_lo": 0.0,746 "cost": 3503.9038544055074,747 "n": 2748 },749 {750 "action": "A5|budget_tokens=512|blind",751 "q_mean": 0.0,752 "q_lo": 0.0,753 "cost": 6761.200083012227,754 "n": 2755 },756 {757 "action": "A6|budget_tokens=512|max_fetch_bytes=2000|max_lookups=6|blind",758 "q_mean": 0.0,759 "q_lo": 0.0,760 "cost": 117.29304090840742,761 "n": 2762 },763 {764 "action": "A8|budget_tokens=512|blind",765 "q_mean": 0.15789473684210525,766 "q_lo": 0.0,767 "cost": 934.2439590254799,768 "n": 2769 }770 ]771 },772 "synthetic|medium|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|aware": {773 "chosen": "B5|keep_frac=0.25|aware",774 "best_static": "B5|keep_frac=0.25|aware",775 "candidates": [776 {777 "action": "B5|keep_frac=0.25|aware",778 "q_mean": 1.0,779 "q_lo": 1.0,780 "cost": 2380.357041885145,781 "n": 5782 },783 {784 "action": "B9|bits=8|keep_frac=0.25|policy=attention|aware",785 "q_mean": 1.0,786 "q_lo": 1.0,787 "cost": 2720.084957452491,788 "n": 5789 }790 ]791 },792 "synthetic|medium|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|blind": {793 "chosen": "B2|bits=8",794 "best_static": "A1",795 "candidates": [796 {797 "action": "A1",798 "q_mean": 1.0,799 "q_lo": 1.0,800 "cost": 2029.3159250169992,801 "n": 5802 },803 {804 "action": "A7|codec=zstd",805 "q_mean": 1.0,806 "q_lo": 1.0,807 "cost": 1883.6203998420388,808 "n": 5809 },810 {811 "action": "B0",812 "q_mean": 1.0,813 "q_lo": 1.0,814 "cost": 901.3801999855787,815 "n": 5816 },817 {818 "action": "B1",819 "q_mean": 1.0,820 "q_lo": 1.0,821 "cost": 951.6737081110477,822 "n": 5823 },824 {825 "action": "B2|bits=8",826 "q_mean": 1.0,827 "q_lo": 1.0,828 "cost": 798.1130251660943,829 "n": 5830 },831 {832 "action": "B3|keep_frac=0.25",833 "q_mean": 0.2,834 "q_lo": 0.2,835 "cost": 692.6723245764151,836 "n": 5837 }838 ]839 },840 "synthetic|medium|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|aware": {841 "chosen": "B5|keep_frac=0.25|aware",842 "best_static": "B5|keep_frac=0.25|aware",843 "candidates": [844 {845 "action": "B5|keep_frac=0.25|aware",846 "q_mean": 0.8,847 "q_lo": 0.8,848 "cost": 2617.682141601108,849 "n": 5850 },851 {852 "action": "B9|bits=8|keep_frac=0.25|policy=attention|aware",853 "q_mean": 0.8,854 "q_lo": 0.8,855 "cost": 2645.6742170033976,856 "n": 5857 }858 ]859 },860 "synthetic|medium|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|blind": {861 "chosen": "B0",862 "best_static": "A1",863 "candidates": [864 {865 "action": "A1",866 "q_mean": 1.0,867 "q_lo": 1.0,868 "cost": 2015.7257752027363,869 "n": 5870 },871 {872 "action": "A7|codec=zstd",873 "q_mean": 1.0,874 "q_lo": 1.0,875 "cost": 3648.054808191955,876 "n": 5877 },878 {879 "action": "B0",880 "q_mean": 1.0,881 "q_lo": 1.0,882 "cost": 926.8670671852306,883 "n": 5884 },885 {886 "action": "B1",887 "q_mean": 1.0,888 "q_lo": 1.0,889 "cost": 964.7658923873678,890 "n": 5891 },892 {893 "action": "B2|bits=8",894 "q_mean": 1.0,895 "q_lo": 1.0,896 "cost": 955.3255414124578,897 "n": 5898 },899 {900 "action": "B3|keep_frac=0.25",901 "q_mean": 0.2,902 "q_lo": 0.2,903 "cost": 789.7272833390161,904 "n": 5905 }906 ]907 },908 "synthetic|short|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|aware": {909 "chosen": "A3|budget_tokens=512|aware",910 "best_static": "A3|budget_tokens=512|aware",911 "candidates": [912 {913 "action": "A3|budget_tokens=512|aware",914 "q_mean": 1.0,915 "q_lo": 1.0,916 "cost": 428.00330000463873,917 "n": 10918 },919 {920 "action": "B5|keep_frac=0.25|aware",921 "q_mean": 1.0,922 "q_lo": 1.0,923 "cost": 1493.95510929171,924 "n": 5925 },926 {927 "action": "B9|bits=8|keep_frac=0.25|policy=attention|aware",928 "q_mean": 1.0,929 "q_lo": 1.0,930 "cost": 1567.7744757616892,931 "n": 5932 }933 ]934 },935 "synthetic|short|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|blind": {936 "chosen": "B1",937 "best_static": "A1",938 "candidates": [939 {940 "action": "A1",941 "q_mean": 1.0,942 "q_lo": 1.0,943 "cost": 774.6145416051149,944 "n": 5945 },946 {947 "action": "A2|budget_tokens=512",948 "q_mean": 0.3,949 "q_lo": 0.2,950 "cost": 341.9635915895924,951 "n": 10952 },953 {954 "action": "A3|budget_tokens=512|blind",955 "q_mean": 0.4,956 "q_lo": 0.4,957 "cost": 363.9699418679811,958 "n": 10959 },960 {961 "action": "A4|budget_tokens=512|blind",962 "q_mean": 0.3,963 "q_lo": 0.2,964 "cost": 4784.523370966781,965 "n": 10966 },967 {968 "action": "A5|budget_tokens=512|blind",969 "q_mean": 0.0,970 "q_lo": 0.0,971 "cost": 4584.249533410184,972 "n": 10973 },974 {975 "action": "A6|budget_tokens=512|max_fetch_bytes=2000|max_lookups=6|blind",976 "q_mean": 0.3,977 "q_lo": 0.2,978 "cost": 2474.1135540185496,979 "n": 10980 },981 {982 "action": "A7|codec=zstd",983 "q_mean": 1.0,984 "q_lo": 1.0,985 "cost": 1710.8100838027894,986 "n": 5987 },988 {989 "action": "A8|budget_tokens=512|blind",990 "q_mean": 0.2,991 "q_lo": 0.0,992 "cost": 2308.3051127032377,993 "n": 10994 },995 {996 "action": "B0",997 "q_mean": 1.0,998 "q_lo": 1.0,999 "cost": 714.2056331736967,1000 "n": 51001 },1002 {1003 "action": "B1",1004 "q_mean": 1.0,1005 "q_lo": 1.0,1006 "cost": 599.7976336162537,1007 "n": 51008 },1009 {1010 "action": "B2|bits=8",1011 "q_mean": 1.0,1012 "q_lo": 1.0,1013 "cost": 622.288533882238,1014 "n": 51015 },1016 {1017 "action": "B3|keep_frac=0.25",1018 "q_mean": 0.2,1019 "q_lo": 0.2,1020 "cost": 638.1709832465276,1021 "n": 51022 }1023 ]1024 },1025 "synthetic|short|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|aware": {1026 "chosen": "A3|budget_tokens=512|aware",1027 "best_static": "A3|budget_tokens=512|aware",1028 "candidates": [1029 {1030 "action": "A3|budget_tokens=512|aware",1031 "q_mean": 1.0,1032 "q_lo": 1.0,1033 "cost": 492.4797534924717,1034 "n": 131035 },1036 {1037 "action": "B5|keep_frac=0.25|aware",1038 "q_mean": 0.8,1039 "q_lo": 0.8,1040 "cost": 1461.255682632327,1041 "n": 51042 },1043 {1044 "action": "B9|bits=8|keep_frac=0.25|policy=attention|aware",1045 "q_mean": 0.8,1046 "q_lo": 0.8,1047 "cost": 1528.5088498611003,1048 "n": 51049 }1050 ]1051 },1052 "synthetic|short|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|blind": {1053 "chosen": "B1",1054 "best_static": "A1",1055 "candidates": [1056 {1057 "action": "A1",1058 "q_mean": 1.0,1059 "q_lo": 1.0,1060 "cost": 834.2573178815655,1061 "n": 81062 },1063 {1064 "action": "A2|budget_tokens=512",1065 "q_mean": 0.38461538461538464,1066 "q_lo": 0.2,1067 "cost": 335.3575096788028,1068 "n": 131069 },1070 {1071 "action": "A3|budget_tokens=512|blind",1072 "q_mean": 0.46153846153846156,1073 "q_lo": 0.4,1074 "cost": 379.59068921358823,1075 "n": 131076 },1077 {1078 "action": "A4|budget_tokens=512|blind",1079 "q_mean": 0.5384615384615384,1080 "q_lo": 0.4,1081 "cost": 3687.5417852648893,1082 "n": 131083 },1084 {1085 "action": "A5|budget_tokens=512|blind",1086 "q_mean": 0.23076923076923078,1087 "q_lo": 0.2,1088 "cost": 4437.3677050999295,1089 "n": 131090 },1091 {1092 "action": "A6|budget_tokens=512|max_fetch_bytes=2000|max_lookups=6|blind",1093 "q_mean": 0.23076923076923078,1094 "q_lo": 0.2,1095 "cost": 5032.344285452452,1096 "n": 131097 },1098 {1099 "action": "A7|codec=zstd",1100 "q_mean": 1.0,1101 "q_lo": 1.0,1102 "cost": 2361.7878749937518,1103 "n": 81104 },1105 {1106 "action": "A8|budget_tokens=512|blind",1107 "q_mean": 0.2,1108 "q_lo": 0.0,1109 "cost": 29813.514716585632,1110 "n": 101111 },1112 {1113 "action": "B0",1114 "q_mean": 1.0,1115 "q_lo": 1.0,1116 "cost": 798.3306667534634,1117 "n": 51118 },1119 {1120 "action": "B1",1121 "q_mean": 1.0,1122 "q_lo": 1.0,1123 "cost": 742.7887752186507,1124 "n": 51125 },1126 {1127 "action": "B2|bits=8",1128 "q_mean": 1.0,1129 "q_lo": 1.0,1130 "cost": 804.3182000052184,1131 "n": 51132 },1133 {1134 "action": "B3|keep_frac=0.25",1135 "q_mean": 0.2,1136 "q_lo": 0.2,1137 "cost": 510.90556636918336,1138 "n": 51139 }1140 ]1141 }1142 },1143 "global_best_static": "A1",1144 "fitted_on": [1145 "E1_pilot.jsonl"1146 ],1147 "n_dev_rows": 10371148}