llama-bench : accept ranges for integer parameters

[slaren]

Accept ranges with the syntax start-end[+step] for all integer parameters.

Example:

$ llama-bench -n 0 -p 1-5,10-100+10,256
| model                          |       size |     params | backend    | ngl |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | CUDA       |  99 |             pp1 |        171.12 ± 5.78 |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | CUDA       |  99 |             pp2 |        229.50 ± 6.88 |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | CUDA       |  99 |             pp3 |       343.37 ± 12.77 |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | CUDA       |  99 |             pp4 |        433.23 ± 9.29 |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | CUDA       |  99 |             pp5 |       494.69 ± 12.29 |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | CUDA       |  99 |            pp10 |        795.65 ± 5.96 |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | CUDA       |  99 |            pp20 |      1489.93 ± 19.02 |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | CUDA       |  99 |            pp30 |      2103.04 ± 61.36 |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | CUDA       |  99 |            pp40 |      2589.35 ± 37.22 |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | CUDA       |  99 |            pp50 |      2904.41 ± 48.25 |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | CUDA       |  99 |            pp60 |      3393.16 ± 19.55 |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | CUDA       |  99 |            pp70 |      3612.60 ± 35.55 |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | CUDA       |  99 |            pp80 |      4003.98 ± 16.02 |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | CUDA       |  99 |            pp90 |      4101.49 ± 18.51 |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | CUDA       |  99 |           pp100 |      4288.92 ± 19.39 |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | CUDA       |  99 |           pp256 |      5728.99 ± 13.87 |

[slaren]

The only real change here is replacing string_split<int>(argv[i], split_delim) with parse_int_range(argv[i]), the rest of the diff is due to the change in indentation.

[JohannesGaessler]

You should specify that the ranges are inclusive.

[slaren]

I changed it to first-last+step. Is that clear enough?

[JohannesGaessler]

I guess that also works but I think it would make sense to also rename the variables in the code to match the help.

[JohannesGaessler]

Would it be possible to support not just linear step sizes but exponential ones as well? So something like -ub 16-128*2 would translate to -ub 16,32,64,128.