Models path change behaviour

[prawel]

When I change models path in settings, it overwrites any models.ini file that is there, without any confirmation. This is wrong.

Also, app doesn't scan subdirectories for GGUFs.

Ideally, it would be great that when pointed to directory that contains models.ini, in llama-server --models-preset syntax, menubar should display that models, with possibility to load/unload them, just like llama-server webui does, along with loaded/unloaded status.

It should also follow settings from models.ini file, e.g.:

version = 1

[*]
fit = on ; Automatic memory fitting
ctk = q8_0 ; KV cache key quantization
ctv = q8_0 ; KV cache value quantization
fa = on ; Enable flash attention
mlock = on ; Lock model in RAM
kvu = on ; Unified KV cache buffer
fit = on
ctx-size = 0
jinja = on

[GPT-OSS-20B]
m = /Users/pk/test/gguf/ggml-org/gpt-oss-20b-GGUF/gpt-oss-20b-mxfp4.gguf
grammar-file = /Users/pk/test/gpt-oss-grammar.gbnf
temp = 1.0
top-p = 1.0
top-k = 0
chat-template-kwargs = {"reasoning_effort":"low"}

[erusev]

The current models.ini is written by the app as an internal file — it wasn't designed to be user-edited, which is why it gets overwritten. We're rethinking that, since a text-based config that users can edit directly (for per-model settings like ctx-size, temp, etc.) is something we want to support.

On subdirectory scanning: as of 0.27.0, the app uses the standard HF cache layout (~/.cache/huggingface/hub/), which includes scanning subdirectories. We're also adding support for "sideloaded" models — any GGUF in the HF cache will show up in the app, not just models from our catalog. That should land in an upcoming release.