Guest User

command-a

a guest
May 3rd, 2025
79
0
Never
Not a member of Pastebin yet? Sign Up, it unlocks many cool features!
text 3.40 KB | None | 0 0
  1. main: n_kv_max = 32768, n_batch = 2048, n_ubatch = 512, flash_attn = 1, n_gpu_layers = 999, n_threads = 28, n_threads_batch = 28
  2.  
  3. | PP | TG | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s |
  4. |-------|--------|--------|----------|----------|----------|----------|
  5. | 512 | 128 | 0 | 1.267 | 404.22 | 10.195 | 12.56 |
  6. | 512 | 128 | 512 | 1.279 | 400.38 | 10.349 | 12.37 |
  7. | 512 | 128 | 1024 | 1.303 | 392.84 | 10.506 | 12.18 |
  8. | 512 | 128 | 1536 | 1.326 | 386.02 | 10.684 | 11.98 |
  9. | 512 | 128 | 2048 | 1.351 | 379.08 | 10.850 | 11.80 |
  10. | 512 | 128 | 2560 | 1.376 | 372.18 | 11.043 | 11.59 |
  11. | 512 | 128 | 3072 | 1.400 | 365.71 | 11.250 | 11.38 |
  12. | 512 | 128 | 3584 | 1.426 | 359.08 | 11.470 | 11.16 |
  13. | 512 | 128 | 4096 | 1.446 | 354.00 | 11.666 | 10.97 |
  14. | 512 | 128 | 4608 | 1.469 | 348.57 | 11.852 | 10.80 |
  15. | 512 | 128 | 5120 | 1.494 | 342.79 | 12.044 | 10.63 |
  16. | 512 | 128 | 5632 | 1.514 | 338.26 | 12.230 | 10.47 |
  17. | 512 | 128 | 6144 | 1.540 | 332.57 | 12.423 | 10.30 |
  18. | 512 | 128 | 6656 | 1.563 | 327.52 | 12.625 | 10.14 |
  19. | 512 | 128 | 7168 | 1.587 | 322.66 | 12.823 | 9.98 |
  20. | 512 | 128 | 7680 | 1.608 | 318.37 | 13.018 | 9.83 |
  21. | 512 | 128 | 8192 | 1.630 | 314.06 | 13.205 | 9.69 |
  22. | 512 | 128 | 8704 | 1.655 | 309.39 | 13.398 | 9.55 |
  23. | 512 | 128 | 9216 | 1.680 | 304.83 | 13.592 | 9.42 |
  24. | 512 | 128 | 9728 | 1.701 | 300.92 | 13.780 | 9.29 |
  25. | 512 | 128 | 10240 | 1.730 | 295.90 | 13.968 | 9.16 |
  26. | 512 | 128 | 10752 | 1.756 | 291.54 | 14.152 | 9.04 |
  27. | 512 | 128 | 11264 | 1.776 | 288.30 | 14.321 | 8.94 |
  28. | 512 | 128 | 11776 | 1.803 | 284.00 | 14.498 | 8.83 |
  29. | 512 | 128 | 12288 | 1.823 | 280.89 | 14.672 | 8.72 |
  30. | 512 | 128 | 12800 | 1.849 | 276.86 | 14.850 | 8.62 |
  31.  
  32. main: n_kv_max = 32768, n_batch = 2048, n_ubatch = 1024, flash_attn = 1, n_gpu_layers = 999, n_threads = 28, n_threads_batch = 28
  33.  
  34. | PP | TG | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s |
  35. |-------|--------|--------|----------|----------|----------|----------|
  36. | 1024 | 256 | 0 | 2.592 | 395.13 | 20.309 | 12.61 |
  37. | 1024 | 256 | 1024 | 2.684 | 381.54 | 20.984 | 12.20 |
  38. | 1024 | 256 | 2048 | 2.780 | 368.29 | 21.702 | 11.80 |
  39. | 1024 | 256 | 3072 | 2.883 | 355.21 | 22.480 | 11.39 |
  40. | 1024 | 256 | 4096 | 2.970 | 344.75 | 23.300 | 10.99 |
  41.  
  42. with mmap
  43.  
  44. main: n_kv_max = 32768, n_batch = 2048, n_ubatch = 512, flash_attn = 1, n_gpu_layers = 999, n_threads = 28, n_threads_batch = 28
  45.  
  46. | PP | TG | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s |
  47. |-------|--------|--------|----------|----------|----------|----------|
  48. | 512 | 128 | 0 | 1.279 | 400.44 | 10.186 | 12.57 |
  49. | 512 | 128 | 512 | 1.288 | 397.41 | 10.342 | 12.38 |
  50. | 512 | 128 | 1024 | 1.308 | 391.42 | 10.500 | 12.19 |
  51. | 512 | 128 | 1536 | 1.334 | 383.80 | 10.680 | 11.99 |
  52. | 512 | 128 | 2048 | 1.352 | 378.72 | 10.850 | 11.80 |
Advertisement
Add Comment
Please, Sign In to add comment