Skip to content

Instantly share code, notes, and snippets.

@mostlygeek
Last active August 17, 2026 17:54
Show Gist options
  • Select an option

  • Save mostlygeek/7c2e5a8f0f5e840fb24bd33d6c2d5d37 to your computer and use it in GitHub Desktop.

Select an option

Save mostlygeek/7c2e5a8f0f5e840fb24bd33d6c2d5d37 to your computer and use it in GitHub Desktop.
Qwen 3.8 - Dual 3090, llama-server, Q8
Summary
Range 2026-08-15 12:38:30 → 2026-08-15 13:44:00
Duration 1526.64s
Cached 11,089,449
Prompt 56,807 (290.43 t/s)
Generated 51,260 (40.48 t/s)
Drafted 91.4% 33461/36617

Perf (top 10 fastest decode)

Duration Status Cached Prompt Prefill Generated Decode Drafted
10.13s 200 56,000 219 259.30 t/s 711 79.11 t/s 99.8% (531/532)
28.44s 200 56,719 119 147.38 t/s 2,079 76.15 t/s 99.3% (1549/1560)
16.61s 200 49,763 1,054 597.89 t/s 1,090 74.84 t/s 97.8% (806/824)
6.55s 200 52,087 19 29.06 t/s 411 73.45 t/s 97.1% (303/312)
4.62s 200 60,854 24 20.01 t/s 223 72.68 t/s 98.8% (164/166)
13.01s 200 58,639 506 412.90 t/s 832 72.63 t/s 99.2% (614/619)
46.10s 200 110,463 57 56.83 t/s 3,190 71.70 t/s 98.9% (2366/2393)
7.24s 200 46,217 992 625.66 t/s 379 70.25 t/s 99.3% (277/279)
9.47s 200 60,893 508 390.55 t/s 550 70.22 t/s 97.8% (398/407)
7.41s 200 59,978 506 410.70 t/s 410 70.19 t/s 99.0% (299/302)

Conf:

models:
  Q3.8-27B-MTP:
    sendLoadingState: false
    env:
      - "CUDA_VISIBLE_DEVICES=GPU-6f0,GPU-f10"
    filters:
      stripParams: "temperature, top_k, top_p, repeat_penalty, min_p, presence_penalty"
      setParamsByID:
        "${MODEL_ID}:thinking-coding":
          temperature: 0.6
          presence_penalty: 0.0

        # test if preserve_thinking improves performance
        "${MODEL_ID}:preserve-thinking-coding":
          temperature: 0.6
          presence_penalty: 0.0
          chat_template_kwargs:
            preserve_thinking: true
            enable_thinking: true
            reasoning_effort: "xhigh"

        "${MODEL_ID}:instruct":
          chat_template_kwargs:
            enable_thinking: false
          temperature: 0.7
          top_p: 0.8
          presence_penalty: 1.0

    cmd: |
      ${qwen-27B-server}
      --model /path/to/models/Qwen3.8-27B-UD-Q8_K_XL.gguf
      --ctx-size 218000
      --kv-unified
      --spec-type draft-mtp
      --spec-draft-n-max 3
      --spec-draft-p-min 0.75
      --tensor-split 50,50
      -sm tensor
      -np 2
      --batch-size 4096
      --ubatch-size 1024

Data

ID Time Status Cached Prompt Generated Prefill Decode Duration
574 2026-08-15 13:44:00 200 163,438 767 492 281.10 t/s 43.47 t/s 14.96s
573 2026-08-15 13:43:44 200 162,331 295 812 97.15 t/s 33.16 t/s 28.43s
572 2026-08-15 13:43:08 200 162,009 167 154 66.00 t/s 48.00 t/s 6.68s
571 2026-08-15 13:42:51 200 161,325 428 255 184.53 t/s 33.54 t/s 10.79s
570 2026-08-15 13:42:39 200 160,661 428 235 181.17 t/s 65.73 t/s 7.45s
569 2026-08-15 13:42:31 200 159,812 428 420 208.45 t/s 56.37 t/s 10.31s
568 2026-08-15 13:42:20 200 155,066 805 3,940 229.16 t/s 32.76 t/s 125.28s
567 2026-08-15 13:40:14 200 153,129 188 1,750 80.61 t/s 24.43 t/s 75.43s
566 2026-08-15 13:38:58 200 152,229 684 217 207.40 t/s 29.39 t/s 12.15s
565 2026-08-15 13:38:45 200 152,013 92 123 31.39 t/s 37.12 t/s 7.68s
564 2026-08-15 13:38:37 200 150,955 776 282 281.91 t/s 28.53 t/s 13.40s
563 2026-08-15 13:38:23 200 149,154 541 1,261 257.00 t/s 30.94 t/s 43.67s
562 2026-08-15 13:37:39 200 149,025 19 109 15.55 t/s 47.58 t/s 4.31s
561 2026-08-15 13:37:34 200 148,441 486 99 240.61 t/s 55.84 t/s 4.59s
560 2026-08-15 13:37:29 200 148,124 135 181 95.12 t/s 38.66 t/s 6.90s
559 2026-08-15 13:37:22 200 148,016 19 90 15.84 t/s 56.16 t/s 3.58s
558 2026-08-15 13:37:18 200 147,897 21 99 16.66 t/s 52.83 t/s 3.93s
557 2026-08-15 13:37:13 200 147,747 18 131 14.71 t/s 42.06 t/s 5.12s
556 2026-08-15 13:37:08 200 146,332 357 1,058 203.14 t/s 33.98 t/s 33.67s
555 2026-08-15 13:36:33 200 146,164 15 152 12.72 t/s 45.56 t/s 5.29s
554 2026-08-15 13:36:28 200 145,972 70 121 54.67 t/s 48.66 t/s 4.53s
553 2026-08-15 13:36:23 200 145,625 239 107 161.33 t/s 52.31 t/s 4.29s
552 2026-08-15 13:36:18 200 143,474 2,036 116 395.47 t/s 41.26 t/s 8.71s
551 2026-08-15 13:36:08 200 142,621 857 2,015 325.90 t/s 31.39 t/s 67.59s
550 2026-08-15 13:35:00 200 139,905 944 1,771 361.39 t/s 34.71 t/s 54.38s
549 2026-08-15 13:34:05 200 137,894 159 1,852 56.26 t/s 33.32 t/s 59.16s
548 2026-08-15 13:31:45 200 136,799 271 824 181.90 t/s 43.49 t/s 21.18s
547 2026-08-15 13:31:23 200 135,133 990 676 382.28 t/s 47.14 t/s 17.65s
546 2026-08-15 13:31:05 200 134,853 130 151 57.31 t/s 42.85 t/s 6.51s
545 2026-08-15 13:30:53 200 134,045 643 164 318.00 t/s 46.50 t/s 6.28s
544 2026-08-15 13:30:45 200 133,782 19 245 15.62 t/s 37.63 t/s 8.44s
543 2026-08-15 13:30:36 200 133,204 311 267 168.67 t/s 45.27 t/s 8.50s
542 2026-08-15 13:30:25 200 132,918 169 117 63.00 t/s 54.99 t/s 5.53s
541 2026-08-15 13:30:08 200 132,376 428 115 225.48 t/s 64.82 t/s 4.40s
540 2026-08-15 13:30:03 200 131,791 428 156 219.41 t/s 55.34 t/s 5.51s
539 2026-08-15 13:29:57 200 130,857 442 491 176.31 t/s 36.21 t/s 16.85s
538 2026-08-15 13:29:29 200 129,884 428 544 230.53 t/s 28.34 t/s 21.80s
537 2026-08-15 13:29:07 200 128,790 440 653 230.81 t/s 58.22 t/s 13.82s
536 2026-08-15 13:28:52 200 127,816 455 519 231.38 t/s 49.36 t/s 13.20s
535 2026-08-15 13:28:38 200 126,940 435 440 218.57 t/s 54.40 t/s 10.80s
534 2026-08-15 13:28:27 200 124,939 185 1,817 122.42 t/s 37.26 t/s 50.98s
533 2026-08-15 13:27:35 200 124,196 627 117 300.14 t/s 45.82 t/s 5.34s
532 2026-08-15 13:27:30 200 120,144 1,665 2,387 458.10 t/s 42.35 t/s 60.68s
531 2026-08-15 13:26:29 200 119,293 754 98 373.64 t/s 49.90 t/s 4.64s
530 2026-08-15 13:26:23 200 119,125 20 149 16.83 t/s 44.66 t/s 5.17s
529 2026-08-15 13:26:18 200 117,002 1,656 468 472.91 t/s 56.47 t/s 12.45s
528 2026-08-15 13:26:05 200 116,856 20 125 19.43 t/s 48.34 t/s 4.25s
527 2026-08-15 13:26:00 200 116,058 271 528 192.52 t/s 68.23 t/s 9.77s
526 2026-08-15 13:25:50 200 114,377 1,367 313 437.59 t/s 52.90 t/s 9.68s
525 2026-08-15 13:25:38 200 114,260 18 99 17.37 t/s 61.19 t/s 3.27s
524 2026-08-15 13:25:34 200 110,918 945 2,396 423.12 t/s 55.06 t/s 46.37s
523 2026-08-15 13:24:47 200 110,799 20 99 19.96 t/s 51.22 t/s 3.54s
522 2026-08-15 13:24:43 200 109,540 20 1,240 21.28 t/s 40.30 t/s 32.31s
521 2026-08-15 13:24:10 200 106,300 2,269 972 518.64 t/s 59.68 t/s 21.26s
520 2026-08-15 13:23:48 200 104,333 1,840 128 517.55 t/s 46.41 t/s 6.89s
519 2026-08-15 13:23:41 200 102,502 1,706 126 521.65 t/s 44.07 t/s 6.71s
518 2026-08-15 13:23:34 200 102,334 17 152 10.04 t/s 42.40 t/s 5.86s
517 2026-08-15 13:23:25 200 102,195 19 119 19.38 t/s 44.09 t/s 4.25s
516 2026-08-15 13:23:20 200 101,942 20 233 20.36 t/s 56.71 t/s 5.67s
515 2026-08-15 13:23:14 200 101,507 18 418 19.08 t/s 67.97 t/s 7.66s
514 2026-08-15 13:23:06 200 100,841 76 589 72.65 t/s 66.14 t/s 10.52s
513 2026-08-15 13:22:55 200 99,927 271 643 215.14 t/s 67.62 t/s 11.33s
512 2026-08-15 13:22:43 200 98,318 459 1,149 297.21 t/s 47.18 t/s 26.46s
511 2026-08-15 13:22:16 200 97,700 354 264 263.20 t/s 55.15 t/s 6.68s
510 2026-08-15 13:22:09 200 97,609 45 46 44.36 t/s 57.76 t/s 2.37s
509 2026-08-15 13:22:06 200 95,817 975 817 436.62 t/s 37.82 t/s 24.39s
508 2026-08-15 13:21:41 200 93,899 1,823 96 549.04 t/s 58.19 t/s 5.52s
507 2026-08-15 13:21:35 200 89,793 3,978 128 652.25 t/s 45.16 t/s 9.47s
506 2026-08-15 13:21:25 200 86,368 605 2,819 393.74 t/s 37.37 t/s 77.50s
505 2026-08-15 13:20:08 200 85,327 946 96 494.88 t/s 58.57 t/s 4.06s
504 2026-08-15 13:20:03 200 83,608 1,619 101 598.67 t/s 54.17 t/s 5.07s
503 2026-08-15 13:19:58 200 83,036 472 101 346.79 t/s 66.87 t/s 3.37s
502 2026-08-15 13:19:54 200 81,233 1,673 129 604.94 t/s 47.81 t/s 5.95s
501 2026-08-15 13:19:48 200 81,059 68 105 79.76 t/s 48.61 t/s 3.49s
500 2026-08-15 13:19:44 200 80,893 77 89 84.90 t/s 61.04 t/s 2.84s
499 2026-08-15 13:19:40 200 80,833 17 42 20.10 t/s 46.97 t/s 2.21s
498 2026-08-15 13:19:38 200 79,105 1,578 149 608.64 t/s 43.16 t/s 6.51s
497 2026-08-15 13:19:31 200 78,076 939 91 495.78 t/s 59.71 t/s 3.88s
496 2026-08-15 13:19:27 200 76,578 1,398 99 587.64 t/s 62.46 t/s 4.42s
495 2026-08-15 13:19:22 200 75,326 1,189 64 547.57 t/s 61.56 t/s 3.66s
494 2026-08-15 13:19:18 200 75,165 39 123 18.99 t/s 48.71 t/s 5.03s
493 2026-08-15 12:42:16 200 74,418 218 530 213.83 t/s 44.45 t/s 13.33s
492 2026-08-15 12:42:03 200 73,555 406 456 326.49 t/s 40.89 t/s 12.83s
491 2026-08-15 12:41:50 200 73,060 391 104 337.29 t/s 51.78 t/s 3.60s
490 2026-08-15 12:41:46 200 72,276 643 142 247.83 t/s 48.26 t/s 5.96s
489 2026-08-15 12:41:34 200 69,297 95 2,885 109.37 t/s 35.64 t/s 82.23s
488 2026-08-15 12:40:11 200 68,782 310 204 171.22 t/s 43.95 t/s 6.85s
487 2026-08-15 12:40:01 200 68,479 168 134 84.03 t/s 59.15 t/s 4.66s
486 2026-08-15 12:39:46 200 66,602 1,580 297 598.19 t/s 41.17 t/s 10.25s
485 2026-08-15 12:39:35 200 66,280 219 102 233.98 t/s 45.25 t/s 3.57s
484 2026-08-15 12:39:31 200 66,068 18 193 24.01 t/s 67.05 t/s 4.01s
483 2026-08-15 12:39:27 200 65,748 18 301 24.04 t/s 62.81 t/s 5.90s
482 2026-08-15 12:39:20 200 65,130 422 195 361.38 t/s 58.50 t/s 4.87s
481 2026-08-15 12:39:15 200 64,334 501 294 401.88 t/s 57.95 t/s 6.68s
480 2026-08-15 12:39:08 200 63,671 501 161 406.91 t/s 65.33 t/s 4.05s
479 2026-08-15 12:39:03 200 62,609 675 386 477.05 t/s 56.46 t/s 8.61s
478 2026-08-15 12:38:54 200 61,952 501 156 402.59 t/s 44.79 t/s 5.07s
477 2026-08-15 12:38:48 200 60,893 508 550 390.55 t/s 70.22 t/s 9.47s
476 2026-08-15 12:38:38 200 59,978 506 410 410.70 t/s 70.19 t/s 7.41s
475 2026-08-15 12:38:30 200 58,639 506 832 412.90 t/s 72.63 t/s 13.01s

Exported from llama-swap at 2026-08-17 10:41:58

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment