server : restore forwarding of base CLI args to router child instances

[jhsmith409]

Summary

In router mode (--models-preset / --models-max), the parent llama-server's own CLI flags stopped being forwarded to the child instances it spawns. This regressed between b9641 (good) and b9692 (bad). Flags that aren't derivable from the preset .ini — e.g. --parallel, --cache-type-k/v, --flash-attn, --n-gpu-layers, --threads, --numa — silently disappear from the child's argument list, so children fall back to defaults.

The most damaging case is --parallel, whose default is now -1 = auto (resolves to n_parallel = 4 on a multi-core host). A child whose ctx-size was tuned for a single slot then allocates ~4× the KV cache and OOMs.

Fixes #24762. Same underlying cause as #24735 ("build b9688 no respect context size"), which manifests via --ctx-size.

Root cause

server_models::load_models() builds each model's effective preset by cascading the [*] global preset and the per-model section, but the final pass that merged the parent's CLI args (base_preset) into every model preset was dropped in the router rework (#23976).

In b9641, after final_presets is assembled:

// server base preset from CLI args takes highest precedence
for (auto & [name, preset] : final_presets) {
    preset.merge(base_preset);
}

In b9692 this block is gone. base_preset is still constructed (load_from_args(argc, argv)) and cleaned (unset_reserved_args), but it is never merged into the per-model presets — so it has no effect on the rendered child args.

Fix

Restore the dropped merge pass right after final_presets is assembled. base_preset and common_preset::merge() (overwrite semantics → CLI takes precedence) are unchanged, so this re-establishes the pre-#23976 behavior with no other changes.

Verification

• Confirmed via the router's spawning server instance with args: log: on b9641 the child receives --parallel, --cache-type-k/v, --flash-attn, -ngl, etc.; on b9692 only preset-derived keys (--ctx-size, --model, --split-mode, --tensor-split, --main-gpu, --alias) survive, and the child logs n_parallel is set to auto, using n_parallel = 4 then OOMs the KV cache.
• The restored block is a verbatim reinstatement of code that shipped in b9641; common_preset::merge(const common_preset&) and the base_preset member are unchanged in b9692.

Note: I was unable to run a full CUDA build to runtime-test this (no CUDA toolchain on the build host — it lives only in the prebuilt server-cuda image). The change is a faithful restoration of the b9641 code path; a maintainer CI build would confirm.

[ggml-gh-bot]

Hi @jhsmith409, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• AI-generated content: This project does not accept PRs, descriptions or commit messages that are fully or predominantly AI-generated. If you have used AI to assist you in writing code, please make sure to disclose that explicitly.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.

[jhsmith409]

The PR restored the code to match the prior code in the repo. Neither I nor AI wrote it, I just put it back. Claude diagnosed it and wrote the description.

So AI didn't write the code from my perspective unless the repo was previously written by AI.

[ngxson]

check before you push - the fix is already deployed