fit-params: accept --spec-type, -md, -ngld arguments

[claudiubelu]

Overview

llama-server accepts the --spec-type, -md, -ngld parameters and fits the draft / MTP context with the target, but llama-fit-params does not. Anything that estimates a server config with llama-fit-params will get a wrong estimate.

llama-fit-params now accepts the same flags, improving the estimates.

Additional information

Fixes: #30424

Repro:

PS D:\work\llama.cpp> .\build\bin\Release\llama-fit-params.exe -m ..\models\Qwen3.5-9B\Qwen3.5-9B-IQ4_XS-mtp-latin-outq4k.gguf --fit off --fit-print on
0.01.957.464 I llama_fit_params: printing estimated memory in MiB to stdout (device, model, context, compute) ...
CUDA0 3895 8242 344
Host 318 0 272
PS D:\work\llama.cpp> .\build\bin\Release\llama-fit-params.exe -m ..\models\Qwen3.5-9B\Qwen3.5-9B-IQ4_XS-mtp-latin-outq4k.gguf --fit off --fit-print on --spec-type draft-mtp
error: invalid argument: --spec-type

# With this fix:
PS D:\work\llama.cpp> .\build-dev\bin\Release\llama-fit-params.exe -m ..\models\Qwen3.5-9B\Qwen3.5-9B-IQ4_XS-mtp-latin-outq4k.gguf --fit off --fit-print on
0.01.941.872 I llama_fit_params: printing estimated memory in MiB to stdout (device, model, context, compute) ...
CUDA0 3895 8242 344
Host 318 0 272
PS D:\work\llama.cpp> .\build-dev\bin\Release\llama-fit-params.exe -m ..\models\Qwen3.5-9B\Qwen3.5-9B-IQ4_XS-mtp-latin-outq4k.gguf --fit off --fit-print on --spec-type draft-mtp
0.00.745.028 I llama_fit_params: printing estimated memory in MiB to stdout (device, model, context, compute) ...
CUDA0 4039 9417 644
Host 318 0 544

PS D:\work\llama.cpp> .\build-dev\bin\Release\llama-fit-params.exe -m ..\models\Qwen3.5-9B\Qwen3.5-9B-Q4_K_M.gguf --fit off --fit-print on --spec-type draft-mtp
0.01.925.228 I llama_fit_params: printing estimated memory in MiB to stdout (device, model, context, compute) ...
0.03.728.047 W common_fit_print: failed to measure the memory of the extra model, printing the main model alone: the model has no MTP layers (needed by --spec-type draft-mtp)
CUDA0 4861 8393 501
Host 545 0 272

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, it discovered mismatch between memory usage estimates and actual usage, discovered that it was due to MTP not being accounted for, and it proposed a solution. I have reviewed its changed, tested the binaries, ran the CI (https://pastebin.com/zxaUZDvD).