Skip to content

Discussion / FYI : Smooth performance on long context #21

@dan-and

Description

@dan-and

since the early builds have massive performance degradations, I have redone the test with the current code base with the CUDA implementation:

From 0 to 260000 tokens, the performance is moving down very smoothly and predictably when running:

Unsloths Qwen3.5-4B-Q4_K_M.gguf Model with -ctk turbo3 -ctv turbo3

Great Job :-)


| model          |             test |            t/s |     peak t/s |        ttfr (ms) |     est_ppt (ms) |    e2e_ttft (ms) |
|:---------------|-----------------:|---------------:|-------------:|-----------------:|-----------------:|-----------------:|
| llamacpp-model |           pp2048 | 3357.96 ± 0.00 |              |    714.35 ± 0.00 |    578.33 ± 0.00 |    714.40 ± 0.00 |
| llamacpp-model |             tg32 |   83.42 ± 0.00 | 86.26 ± 0.00 |                  |                  |                  |
| llamacpp-model |   pp2048 @ d4096 | 3426.76 ± 0.00 |              |   1801.74 ± 0.00 |   1665.71 ± 0.00 |   1801.78 ± 0.00 |
| llamacpp-model |     tg32 @ d4096 |   75.85 ± 0.00 | 78.53 ± 0.00 |                  |                  |                  |
| llamacpp-model |   pp2048 @ d8192 | 3446.56 ± 0.00 |              |   2841.33 ± 0.00 |   2705.31 ± 0.00 |   2841.37 ± 0.00 |
| llamacpp-model |     tg32 @ d8192 |   69.58 ± 0.00 | 72.12 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d12288 | 3382.78 ± 0.00 |              |   4016.56 ± 0.00 |   3880.53 ± 0.00 |   4016.61 ± 0.00 |
| llamacpp-model |    tg32 @ d12288 |   65.45 ± 0.00 | 67.88 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d16384 | 3353.15 ± 0.00 |              |   5137.89 ± 0.00 |   5001.86 ± 0.00 |   5137.93 ± 0.00 |
| llamacpp-model |    tg32 @ d16384 |   60.70 ± 0.00 | 63.03 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d20480 | 3272.50 ± 0.00 |              |   6239.61 ± 0.00 |   6103.59 ± 0.00 |   6239.66 ± 0.00 |
| llamacpp-model |    tg32 @ d20480 |   57.20 ± 0.00 | 59.40 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d24576 | 3206.25 ± 0.00 |              |   7647.60 ± 0.00 |   7511.58 ± 0.00 |   7647.65 ± 0.00 |
| llamacpp-model |    tg32 @ d24576 |   53.72 ± 0.00 | 55.82 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d28672 | 3137.46 ± 0.00 |              |   8999.56 ± 0.00 |   8863.53 ± 0.00 |   8999.60 ± 0.00 |
| llamacpp-model |    tg32 @ d28672 |   50.54 ± 0.00 | 52.52 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d32768 | 3065.26 ± 0.00 |              |  10406.61 ± 0.00 |  10270.58 ± 0.00 |  10406.65 ± 0.00 |
| llamacpp-model |    tg32 @ d32768 |   47.16 ± 0.00 | 49.03 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d36864 | 2995.70 ± 0.00 |              |  11860.16 ± 0.00 |  11724.14 ± 0.00 |  11860.20 ± 0.00 |
| llamacpp-model |    tg32 @ d36864 |   45.48 ± 0.00 | 47.32 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d40960 | 2936.61 ± 0.00 |              |  13334.58 ± 0.00 |  13198.55 ± 0.00 |  13334.61 ± 0.00 |
| llamacpp-model |    tg32 @ d40960 |   43.19 ± 0.00 | 44.95 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d45056 | 2860.09 ± 0.00 |              |  15032.78 ± 0.00 |  14896.75 ± 0.00 |  15032.81 ± 0.00 |
| llamacpp-model |    tg32 @ d45056 |   40.95 ± 0.00 | 42.60 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d49152 | 2803.67 ± 0.00 |              |  16652.93 ± 0.00 |  16516.90 ± 0.00 |  16652.97 ± 0.00 |
| llamacpp-model |    tg32 @ d49152 |   38.75 ± 0.00 | 40.33 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d53248 | 2733.08 ± 0.00 |              |  18462.59 ± 0.00 |  18326.57 ± 0.00 |  18462.63 ± 0.00 |
| llamacpp-model |    tg32 @ d53248 |   37.29 ± 0.00 | 38.79 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d57344 | 2668.39 ± 0.00 |              |  20194.57 ± 0.00 |  20058.55 ± 0.00 |  20194.61 ± 0.00 |
| llamacpp-model |    tg32 @ d57344 |   35.67 ± 0.00 | 37.12 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d61440 | 2604.57 ± 0.00 |              |  22247.17 ± 0.00 |  22111.14 ± 0.00 |  22247.21 ± 0.00 |
| llamacpp-model |    tg32 @ d61440 |   33.79 ± 0.00 | 35.21 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d65536 | 2532.17 ± 0.00 |              |  24346.91 ± 0.00 |  24210.89 ± 0.00 |  24346.96 ± 0.00 |
| llamacpp-model |    tg32 @ d65536 |   32.59 ± 0.00 | 33.92 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d69632 | 2457.23 ± 0.00 |              |  26618.68 ± 0.00 |  26482.65 ± 0.00 |  26618.72 ± 0.00 |
| llamacpp-model |    tg32 @ d69632 |   31.49 ± 0.00 | 32.78 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d73728 | 2399.10 ± 0.00 |              |  28717.61 ± 0.00 |  28581.58 ± 0.00 |  28717.65 ± 0.00 |
| llamacpp-model |    tg32 @ d73728 |   30.22 ± 0.00 | 31.00 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d77824 | 2350.01 ± 0.00 |              |  30949.93 ± 0.00 |  30813.90 ± 0.00 |  30949.96 ± 0.00 |
| llamacpp-model |    tg32 @ d77824 |   28.95 ± 0.00 | 30.00 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d81920 | 2294.49 ± 0.00 |              |  33227.09 ± 0.00 |  33091.07 ± 0.00 |  33227.13 ± 0.00 |
| llamacpp-model |    tg32 @ d81920 |   28.00 ± 0.00 | 29.00 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d86016 | 2242.76 ± 0.00 |              |  35746.63 ± 0.00 |  35610.61 ± 0.00 |  35746.69 ± 0.00 |
| llamacpp-model |    tg32 @ d86016 |   27.03 ± 0.00 | 28.00 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d90112 | 2204.01 ± 0.00 |              |  38000.70 ± 0.00 |  37864.68 ± 0.00 |  38000.76 ± 0.00 |
| llamacpp-model |    tg32 @ d90112 |   26.17 ± 0.00 | 27.00 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d94208 | 2158.79 ± 0.00 |              |  40471.18 ± 0.00 |  40335.16 ± 0.00 |  40471.23 ± 0.00 |
| llamacpp-model |    tg32 @ d94208 |   25.55 ± 0.00 | 27.00 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d98304 | 2121.21 ± 0.00 |              |  42979.55 ± 0.00 |  42843.53 ± 0.00 |  42979.59 ± 0.00 |
| llamacpp-model |    tg32 @ d98304 |   24.80 ± 0.00 | 26.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d102400 | 2080.53 ± 0.00 |              |  45673.90 ± 0.00 |  45537.88 ± 0.00 |  45673.94 ± 0.00 |
| llamacpp-model |   tg32 @ d102400 |   24.11 ± 0.00 | 25.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d106496 | 2047.42 ± 0.00 |              |  48177.01 ± 0.00 |  48040.98 ± 0.00 |  48177.05 ± 0.00 |
| llamacpp-model |   tg32 @ d106496 |   23.35 ± 0.00 | 24.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d110592 | 2019.84 ± 0.00 |              |  50714.38 ± 0.00 |  50578.36 ± 0.00 |  50714.42 ± 0.00 |
| llamacpp-model |   tg32 @ d110592 |   22.81 ± 0.00 | 24.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d114688 | 1985.74 ± 0.00 |              |  53406.45 ± 0.00 |  53270.42 ± 0.00 |  53406.53 ± 0.00 |
| llamacpp-model |   tg32 @ d114688 |   22.12 ± 0.00 | 23.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d118784 | 1947.66 ± 0.00 |              |  56508.32 ± 0.00 |  56372.30 ± 0.00 |  56508.38 ± 0.00 |
| llamacpp-model |   tg32 @ d118784 |   21.45 ± 0.00 | 22.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d122880 | 1921.86 ± 0.00 |              |  58975.36 ± 0.00 |  58839.33 ± 0.00 |  58975.41 ± 0.00 |
| llamacpp-model |   tg32 @ d122880 |   21.15 ± 0.00 | 22.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d126976 | 1885.92 ± 0.00 |              |  62158.18 ± 0.00 |  62022.16 ± 0.00 |  62158.22 ± 0.00 |
| llamacpp-model |   tg32 @ d126976 |   20.60 ± 0.00 | 21.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d131072 | 1861.36 ± 0.00 |              |  64886.48 ± 0.00 |  64750.46 ± 0.00 |  64886.57 ± 0.00 |
| llamacpp-model |   tg32 @ d131072 |   20.08 ± 0.00 | 21.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d135168 | 1831.51 ± 0.00 |              |  67914.49 ± 0.00 |  67778.47 ± 0.00 |  67914.55 ± 0.00 |
| llamacpp-model |   tg32 @ d135168 |   19.65 ± 0.00 | 21.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d139264 | 1799.84 ± 0.00 |              |  71136.37 ± 0.00 |  71000.35 ± 0.00 |  71136.43 ± 0.00 |
| llamacpp-model |   tg32 @ d139264 |   19.22 ± 0.00 | 20.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d143360 | 1771.14 ± 0.00 |              |  74602.37 ± 0.00 |  74466.35 ± 0.00 |  74602.42 ± 0.00 |
| llamacpp-model |   tg32 @ d143360 |   18.77 ± 0.00 | 20.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d147456 | 1744.12 ± 0.00 |              |  77818.14 ± 0.00 |  77682.12 ± 0.00 |  77818.19 ± 0.00 |
| llamacpp-model |   tg32 @ d147456 |   18.19 ± 0.00 | 19.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d151552 | 1723.22 ± 0.00 |              |  80809.77 ± 0.00 |  80673.74 ± 0.00 |  80809.80 ± 0.00 |
| llamacpp-model |   tg32 @ d151552 |   17.99 ± 0.00 | 19.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d155648 | 1696.27 ± 0.00 |              |  84377.23 ± 0.00 |  84241.21 ± 0.00 |  84377.28 ± 0.00 |
| llamacpp-model |   tg32 @ d155648 |   17.63 ± 0.00 | 19.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d159744 | 1673.22 ± 0.00 |              |  87788.61 ± 0.00 |  87652.58 ± 0.00 |  87788.65 ± 0.00 |
| llamacpp-model |   tg32 @ d159744 |   17.25 ± 0.00 | 18.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d163840 | 1646.63 ± 0.00 |              |  91424.09 ± 0.00 |  91288.06 ± 0.00 |  91424.14 ± 0.00 |
| llamacpp-model |   tg32 @ d163840 |   17.07 ± 0.00 | 18.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d167936 | 1622.49 ± 0.00 |              |  95189.04 ± 0.00 |  95053.02 ± 0.00 |  95189.08 ± 0.00 |
| llamacpp-model |   tg32 @ d167936 |   16.61 ± 0.00 | 18.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d172032 | 1593.15 ± 0.00 |              |  99303.49 ± 0.00 |  99167.46 ± 0.00 |  99303.53 ± 0.00 |
| llamacpp-model |   tg32 @ d172032 |   16.32 ± 0.00 | 17.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d176128 | 1575.59 ± 0.00 |              | 102785.84 ± 0.00 | 102649.81 ± 0.00 | 102785.88 ± 0.00 |
| llamacpp-model |   tg32 @ d176128 |   16.05 ± 0.00 | 17.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d180224 | 1552.26 ± 0.00 |              | 106711.51 ± 0.00 | 106575.49 ± 0.00 | 106711.56 ± 0.00 |
| llamacpp-model |   tg32 @ d180224 |   15.68 ± 0.00 | 17.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d184320 | 1534.39 ± 0.00 |              | 110310.06 ± 0.00 | 110174.04 ± 0.00 | 110310.10 ± 0.00 |
| llamacpp-model |   tg32 @ d184320 |   15.48 ± 0.00 | 16.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d188416 | 1516.98 ± 0.00 |              | 113721.50 ± 0.00 | 113585.47 ± 0.00 | 113721.54 ± 0.00 |
| llamacpp-model |   tg32 @ d188416 |   15.24 ± 0.00 | 16.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d192512 | 1491.30 ± 0.00 |              | 118509.15 ± 0.00 | 118373.13 ± 0.00 | 118509.19 ± 0.00 |
| llamacpp-model |   tg32 @ d192512 |   14.95 ± 0.00 | 16.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d196608 | 1471.26 ± 0.00 |              | 122805.39 ± 0.00 | 122669.37 ± 0.00 | 122805.43 ± 0.00 |
| llamacpp-model |   tg32 @ d196608 |   14.80 ± 0.00 | 16.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d200704 | 1458.37 ± 0.00 |              | 125987.13 ± 0.00 | 125851.11 ± 0.00 | 125987.18 ± 0.00 |
| llamacpp-model |   tg32 @ d200704 |   14.41 ± 0.00 | 15.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d204800 | 1437.43 ± 0.00 |              | 130774.22 ± 0.00 | 130638.19 ± 0.00 | 130774.27 ± 0.00 |
| llamacpp-model |   tg32 @ d204800 |   14.18 ± 0.00 | 15.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d208896 | 1416.41 ± 0.00 |              | 135347.23 ± 0.00 | 135211.20 ± 0.00 | 135347.27 ± 0.00 |
| llamacpp-model |   tg32 @ d208896 |   14.02 ± 0.00 | 15.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d212992 | 1400.21 ± 0.00 |              | 139079.05 ± 0.00 | 138943.02 ± 0.00 | 139079.09 ± 0.00 |
| llamacpp-model |   tg32 @ d212992 |   13.68 ± 0.00 | 14.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d217088 | 1382.22 ± 0.00 |              | 143739.22 ± 0.00 | 143603.19 ± 0.00 | 143739.26 ± 0.00 |
| llamacpp-model |   tg32 @ d217088 |   13.61 ± 0.00 | 15.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d221184 | 1360.31 ± 0.00 |              | 149030.98 ± 0.00 | 148894.96 ± 0.00 | 149031.03 ± 0.00 |
| llamacpp-model |   tg32 @ d221184 |   13.30 ± 0.00 | 14.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d225280 | 1344.43 ± 0.00 |              | 153261.20 ± 0.00 | 153125.17 ± 0.00 | 153261.24 ± 0.00 |
| llamacpp-model |   tg32 @ d225280 |   13.12 ± 0.00 | 14.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d229376 | 1325.67 ± 0.00 |              | 158416.95 ± 0.00 | 158280.93 ± 0.00 | 158417.01 ± 0.00 |
| llamacpp-model |   tg32 @ d229376 |   12.84 ± 0.00 | 14.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d233472 | 1313.42 ± 0.00 |              | 162637.25 ± 0.00 | 162501.22 ± 0.00 | 162637.31 ± 0.00 |
| llamacpp-model |   tg32 @ d233472 |   12.78 ± 0.00 | 14.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d237568 | 1297.20 ± 0.00 |              | 167530.82 ± 0.00 | 167394.79 ± 0.00 | 167530.85 ± 0.00 |
| llamacpp-model |   tg32 @ d237568 |   12.35 ± 0.00 | 13.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d241664 | 1281.79 ± 0.00 |              | 172606.08 ± 0.00 | 172470.05 ± 0.00 | 172606.13 ± 0.00 |
| llamacpp-model |   tg32 @ d241664 |   12.40 ± 0.00 | 13.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d245760 | 1266.15 ± 0.00 |              | 177761.94 ± 0.00 | 177625.92 ± 0.00 | 177761.98 ± 0.00 |
| llamacpp-model |   tg32 @ d245760 |   12.25 ± 0.00 | 13.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d249856 | 1252.94 ± 0.00 |              | 182221.82 ± 0.00 | 182085.79 ± 0.00 | 182221.88 ± 0.00 |
| llamacpp-model |   tg32 @ d249856 |   12.07 ± 0.00 | 13.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d253952 | 1239.44 ± 0.00 |              | 187327.68 ± 0.00 | 187191.66 ± 0.00 | 187327.72 ± 0.00 |
| llamacpp-model |   tg32 @ d253952 |   11.93 ± 0.00 | 13.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d258048 | 1227.05 ± 0.00 |              | 192256.83 ± 0.00 | 192120.81 ± 0.00 | 192256.88 ± 0.00 |
| llamacpp-model |   tg32 @ d258048 |   11.78 ± 0.00 | 13.00 ± 0.00 |                  |                  |                  |

llama-benchy (0.3.2.dev1+g17b42667a)
date: 2026-03-29 09:31:13 | latency mode: generation

Also turbo2 works fine:

| model          |             test |            t/s |     peak t/s |        ttfr (ms) |     est_ppt (ms) |    e2e_ttft (ms) |
|:---------------|-----------------:|---------------:|-------------:|-----------------:|-----------------:|-----------------:|
| llamacpp-model |           pp2048 | 3363.06 ± 0.00 |              |    700.56 ± 0.00 |    564.37 ± 0.00 |    700.61 ± 0.00 |
| llamacpp-model |             tg32 |   85.18 ± 0.00 | 88.04 ± 0.00 |                  |                  |                  |
| llamacpp-model |   pp2048 @ d4096 | 3437.04 ± 0.00 |              |   1719.24 ± 0.00 |   1583.05 ± 0.00 |   1719.28 ± 0.00 |
| llamacpp-model |     tg32 @ d4096 |   80.56 ± 0.00 | 83.39 ± 0.00 |                  |                  |                  |
| llamacpp-model |   pp2048 @ d8192 | 3451.68 ± 0.00 |              |   2871.38 ± 0.00 |   2735.19 ± 0.00 |   2871.42 ± 0.00 |
| llamacpp-model |     tg32 @ d8192 |   76.03 ± 0.00 | 78.80 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d12288 | 3388.71 ± 0.00 |              |   3982.50 ± 0.00 |   3846.30 ± 0.00 |   3982.55 ± 0.00 |
| llamacpp-model |    tg32 @ d12288 |   71.38 ± 0.00 | 74.06 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d16384 | 3361.10 ± 0.00 |              |   5150.32 ± 0.00 |   5014.13 ± 0.00 |   5150.36 ± 0.00 |
| llamacpp-model |    tg32 @ d16384 |   68.72 ± 0.00 | 71.37 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d20480 | 3285.20 ± 0.00 |              |   6390.61 ± 0.00 |   6254.42 ± 0.00 |   6390.65 ± 0.00 |
| llamacpp-model |    tg32 @ d20480 |   65.54 ± 0.00 | 68.14 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d24576 | 3223.35 ± 0.00 |              |   7587.13 ± 0.00 |   7450.94 ± 0.00 |   7587.17 ± 0.00 |
| llamacpp-model |    tg32 @ d24576 |   63.20 ± 0.00 | 65.72 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d28672 | 3144.74 ± 0.00 |              |   9011.01 ± 0.00 |   8874.81 ± 0.00 |   9011.05 ± 0.00 |
| llamacpp-model |    tg32 @ d28672 |   60.18 ± 0.00 | 62.61 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d32768 | 3075.11 ± 0.00 |              |  10489.95 ± 0.00 |  10353.76 ± 0.00 |  10489.99 ± 0.00 |
| llamacpp-model |    tg32 @ d32768 |   57.62 ± 0.00 | 59.98 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d36864 | 3012.57 ± 0.00 |              |  11863.07 ± 0.00 |  11726.88 ± 0.00 |  11863.12 ± 0.00 |
| llamacpp-model |    tg32 @ d36864 |   55.69 ± 0.00 | 58.04 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d40960 | 2952.43 ± 0.00 |              |  13386.63 ± 0.00 |  13250.43 ± 0.00 |  13386.66 ± 0.00 |
| llamacpp-model |    tg32 @ d40960 |   53.80 ± 0.00 | 56.05 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d45056 | 2878.70 ± 0.00 |              |  15032.48 ± 0.00 |  14896.29 ± 0.00 |  15032.52 ± 0.00 |
| llamacpp-model |    tg32 @ d45056 |   51.63 ± 0.00 | 53.84 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d49152 | 2807.42 ± 0.00 |              |  16613.96 ± 0.00 |  16477.76 ± 0.00 |  16614.00 ± 0.00 |
| llamacpp-model |    tg32 @ d49152 |   49.93 ± 0.00 | 52.11 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d53248 | 2742.86 ± 0.00 |              |  18394.85 ± 0.00 |  18258.65 ± 0.00 |  18394.90 ± 0.00 |
| llamacpp-model |    tg32 @ d53248 |   47.99 ± 0.00 | 50.08 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d57344 | 2667.60 ± 0.00 |              |  20288.00 ± 0.00 |  20151.81 ± 0.00 |  20288.05 ± 0.00 |
| llamacpp-model |    tg32 @ d57344 |   45.92 ± 0.00 | 47.93 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d61440 | 2603.33 ± 0.00 |              |  22152.22 ± 0.00 |  22016.03 ± 0.00 |  22152.26 ± 0.00 |
| llamacpp-model |    tg32 @ d61440 |   44.33 ± 0.00 | 46.27 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d65536 | 2538.95 ± 0.00 |              |  24205.15 ± 0.00 |  24068.96 ± 0.00 |  24205.19 ± 0.00 |
| llamacpp-model |    tg32 @ d65536 |   43.12 ± 0.00 | 45.02 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d69632 | 2465.64 ± 0.00 |              |  26411.76 ± 0.00 |  26275.57 ± 0.00 |  26411.80 ± 0.00 |
| llamacpp-model |    tg32 @ d69632 |   41.87 ± 0.00 | 43.71 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d73728 | 2390.93 ± 0.00 |              |  28804.07 ± 0.00 |  28667.88 ± 0.00 |  28804.11 ± 0.00 |
| llamacpp-model |    tg32 @ d73728 |   40.38 ± 0.00 | 42.22 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d77824 | 2343.44 ± 0.00 |              |  30971.25 ± 0.00 |  30835.06 ± 0.00 |  30971.31 ± 0.00 |
| llamacpp-model |    tg32 @ d77824 |   39.29 ± 0.00 | 41.04 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d81920 | 2304.18 ± 0.00 |              |  33270.39 ± 0.00 |  33134.20 ± 0.00 |  33270.44 ± 0.00 |
| llamacpp-model |    tg32 @ d81920 |   38.07 ± 0.00 | 39.77 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d86016 | 2257.04 ± 0.00 |              |  35594.20 ± 0.00 |  35458.01 ± 0.00 |  35594.24 ± 0.00 |
| llamacpp-model |    tg32 @ d86016 |   37.10 ± 0.00 | 38.75 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d90112 | 2219.84 ± 0.00 |              |  37789.44 ± 0.00 |  37653.25 ± 0.00 |  37789.48 ± 0.00 |
| llamacpp-model |    tg32 @ d90112 |   36.15 ± 0.00 | 37.79 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d94208 | 2179.44 ± 0.00 |              |  40145.48 ± 0.00 |  40009.29 ± 0.00 |  40145.59 ± 0.00 |
| llamacpp-model |    tg32 @ d94208 |   35.18 ± 0.00 | 36.81 ± 0.00 |                  |                  |                  |
| llamacpp-model |  pp2048 @ d98304 | 2142.30 ± 0.00 |              |  42581.17 ± 0.00 |  42444.97 ± 0.00 |  42581.21 ± 0.00 |
| llamacpp-model |    tg32 @ d98304 |   34.37 ± 0.00 | 35.93 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d102400 | 2104.26 ± 0.00 |              |  45206.73 ± 0.00 |  45070.54 ± 0.00 |  45206.78 ± 0.00 |
| llamacpp-model |   tg32 @ d102400 |   33.29 ± 0.00 | 34.81 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d106496 | 2069.37 ± 0.00 |              |  47754.02 ± 0.00 |  47617.83 ± 0.00 |  47754.07 ± 0.00 |
| llamacpp-model |   tg32 @ d106496 |   32.35 ± 0.00 | 33.84 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d110592 | 2041.54 ± 0.00 |              |  50176.25 ± 0.00 |  50040.05 ± 0.00 |  50176.28 ± 0.00 |
| llamacpp-model |   tg32 @ d110592 |   31.99 ± 0.00 | 33.47 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d114688 | 2005.05 ± 0.00 |              |  52968.85 ± 0.00 |  52832.66 ± 0.00 |  52968.90 ± 0.00 |
| llamacpp-model |   tg32 @ d114688 |   31.17 ± 0.00 | 32.60 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d118784 | 1969.64 ± 0.00 |              |  55882.88 ± 0.00 |  55746.68 ± 0.00 |  55882.93 ± 0.00 |
| llamacpp-model |   tg32 @ d118784 |   30.25 ± 0.00 | 31.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d122880 | 1941.63 ± 0.00 |              |  58532.56 ± 0.00 |  58396.36 ± 0.00 |  58532.59 ± 0.00 |
| llamacpp-model |   tg32 @ d122880 |   29.72 ± 0.00 | 31.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d126976 | 1910.64 ± 0.00 |              |  61508.30 ± 0.00 |  61372.11 ± 0.00 |  61508.34 ± 0.00 |
| llamacpp-model |   tg32 @ d126976 |   29.22 ± 0.00 | 30.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d131072 | 1878.21 ± 0.00 |              |  64561.27 ± 0.00 |  64425.08 ± 0.00 |  64561.31 ± 0.00 |
| llamacpp-model |   tg32 @ d131072 |   28.46 ± 0.00 | 30.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d135168 | 1855.69 ± 0.00 |              |  67228.33 ± 0.00 |  67092.14 ± 0.00 |  67228.37 ± 0.00 |
| llamacpp-model |   tg32 @ d135168 |   28.11 ± 0.00 | 29.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d139264 | 1823.41 ± 0.00 |              |  70526.92 ± 0.00 |  70390.73 ± 0.00 |  70526.97 ± 0.00 |
| llamacpp-model |   tg32 @ d139264 |   27.65 ± 0.00 | 29.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d143360 | 1799.44 ± 0.00 |              |  73399.98 ± 0.00 |  73263.79 ± 0.00 |  73400.02 ± 0.00 |
| llamacpp-model |   tg32 @ d143360 |   26.94 ± 0.00 | 28.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d147456 | 1773.69 ± 0.00 |              |  76430.21 ± 0.00 |  76294.01 ± 0.00 |  76430.24 ± 0.00 |
| llamacpp-model |   tg32 @ d147456 |   26.64 ± 0.00 | 28.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d151552 | 1748.98 ± 0.00 |              |  79640.03 ± 0.00 |  79503.84 ± 0.00 |  79640.07 ± 0.00 |
| llamacpp-model |   tg32 @ d151552 |   26.04 ± 0.00 | 27.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d155648 | 1723.21 ± 0.00 |              |  83051.93 ± 0.00 |  82915.74 ± 0.00 |  83051.97 ± 0.00 |
| llamacpp-model |   tg32 @ d155648 |   25.44 ± 0.00 | 27.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d159744 | 1697.96 ± 0.00 |              |  86511.71 ± 0.00 |  86375.51 ± 0.00 |  86511.75 ± 0.00 |
| llamacpp-model |   tg32 @ d159744 |   24.99 ± 0.00 | 26.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d163840 | 1673.13 ± 0.00 |              |  89972.86 ± 0.00 |  89836.67 ± 0.00 |  89972.91 ± 0.00 |
| llamacpp-model |   tg32 @ d163840 |   24.62 ± 0.00 | 26.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d167936 | 1655.17 ± 0.00 |              |  93144.33 ± 0.00 |  93008.13 ± 0.00 |  93144.37 ± 0.00 |
| llamacpp-model |   tg32 @ d167936 |   24.27 ± 0.00 | 25.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d172032 | 1630.81 ± 0.00 |              |  96750.41 ± 0.00 |  96614.22 ± 0.00 |  96750.45 ± 0.00 |
| llamacpp-model |   tg32 @ d172032 |   23.87 ± 0.00 | 25.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d176128 | 1611.12 ± 0.00 |              | 100318.37 ± 0.00 | 100182.17 ± 0.00 | 100318.41 ± 0.00 |
| llamacpp-model |   tg32 @ d176128 |   23.40 ± 0.00 | 25.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d180224 | 1586.65 ± 0.00 |              | 104271.75 ± 0.00 | 104135.56 ± 0.00 | 104271.79 ± 0.00 |
| llamacpp-model |   tg32 @ d180224 |   22.77 ± 0.00 | 24.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d184320 | 1566.23 ± 0.00 |              | 107966.12 ± 0.00 | 107829.93 ± 0.00 | 107966.16 ± 0.00 |
| llamacpp-model |   tg32 @ d184320 |   22.76 ± 0.00 | 24.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d188416 | 1539.55 ± 0.00 |              | 112443.18 ± 0.00 | 112306.98 ± 0.00 | 112443.22 ± 0.00 |
| llamacpp-model |   tg32 @ d188416 |   22.09 ± 0.00 | 23.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d192512 | 1522.61 ± 0.00 |              | 116000.03 ± 0.00 | 115863.84 ± 0.00 | 116000.08 ± 0.00 |
| llamacpp-model |   tg32 @ d192512 |   21.98 ± 0.00 | 23.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d196608 | 1503.58 ± 0.00 |              | 119889.05 ± 0.00 | 119752.86 ± 0.00 | 119889.09 ± 0.00 |
| llamacpp-model |   tg32 @ d196608 |   21.67 ± 0.00 | 23.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d200704 | 1478.96 ± 0.00 |              | 124346.05 ± 0.00 | 124209.86 ± 0.00 | 124346.09 ± 0.00 |
| llamacpp-model |   tg32 @ d200704 |   21.31 ± 0.00 | 22.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d204800 | 1462.59 ± 0.00 |              | 128285.16 ± 0.00 | 128148.97 ± 0.00 | 128285.20 ± 0.00 |
| llamacpp-model |   tg32 @ d204800 |   21.08 ± 0.00 | 22.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d208896 | 1444.56 ± 0.00 |              | 132489.65 ± 0.00 | 132353.46 ± 0.00 | 132489.69 ± 0.00 |
| llamacpp-model |   tg32 @ d208896 |   20.70 ± 0.00 | 22.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d212992 | 1427.11 ± 0.00 |              | 136684.45 ± 0.00 | 136548.26 ± 0.00 | 136684.49 ± 0.00 |
| llamacpp-model |   tg32 @ d212992 |   20.21 ± 0.00 | 21.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d217088 | 1409.41 ± 0.00 |              | 141056.49 ± 0.00 | 140920.30 ± 0.00 | 141056.54 ± 0.00 |
| llamacpp-model |   tg32 @ d217088 |   20.06 ± 0.00 | 21.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d221184 | 1394.43 ± 0.00 |              | 145122.07 ± 0.00 | 144985.88 ± 0.00 | 145122.12 ± 0.00 |
| llamacpp-model |   tg32 @ d221184 |   19.91 ± 0.00 | 21.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d225280 | 1375.80 ± 0.00 |              | 150017.28 ± 0.00 | 149881.09 ± 0.00 | 150017.32 ± 0.00 |
| llamacpp-model |   tg32 @ d225280 |   19.43 ± 0.00 | 20.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d229376 | 1356.68 ± 0.00 |              | 154843.64 ± 0.00 | 154707.45 ± 0.00 | 154843.68 ± 0.00 |
| llamacpp-model |   tg32 @ d229376 |   19.37 ± 0.00 | 20.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d233472 | 1342.57 ± 0.00 |              | 159033.12 ± 0.00 | 158896.93 ± 0.00 | 159033.16 ± 0.00 |
| llamacpp-model |   tg32 @ d233472 |   19.13 ± 0.00 | 20.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d237568 | 1327.78 ± 0.00 |              | 163738.17 ± 0.00 | 163601.98 ± 0.00 | 163738.23 ± 0.00 |
| llamacpp-model |   tg32 @ d237568 |   18.84 ± 0.00 | 20.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d241664 | 1313.72 ± 0.00 |              | 168279.85 ± 0.00 | 168143.66 ± 0.00 | 168279.89 ± 0.00 |
| llamacpp-model |   tg32 @ d241664 |   18.76 ± 0.00 | 20.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d245760 | 1295.95 ± 0.00 |              | 173362.60 ± 0.00 | 173226.41 ± 0.00 | 173362.65 ± 0.00 |
| llamacpp-model |   tg32 @ d245760 |   18.36 ± 0.00 | 19.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d249856 | 1281.91 ± 0.00 |              | 178170.67 ± 0.00 | 178034.47 ± 0.00 | 178170.70 ± 0.00 |
| llamacpp-model |   tg32 @ d249856 |   18.10 ± 0.00 | 19.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d253952 | 1266.63 ± 0.00 |              | 183321.25 ± 0.00 | 183185.06 ± 0.00 | 183321.29 ± 0.00 |
| llamacpp-model |   tg32 @ d253952 |   17.89 ± 0.00 | 19.00 ± 0.00 |                  |                  |                  |
| llamacpp-model | pp2048 @ d258048 | 1252.60 ± 0.00 |              | 188352.62 ± 0.00 | 188216.42 ± 0.00 | 188352.67 ± 0.00 |
| llamacpp-model |   tg32 @ d258048 |   17.69 ± 0.00 | 19.00 ± 0.00 |                  |                  |                  |

llama-benchy (0.3.2.dev1+g17b42667a)
date: 2026-03-29 11:25:44 | latency mode: generation

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions