Feature Request: Router mode - Do not reload model files if they are already loaded.

[mcr-ksh]

Prerequisites

• I am running the latest code. Mention the version if possible as well.
• I carefully followed the README.md.
• I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
• I reviewed the Discussions, and have a new and useful enhancement to share.

Feature Description

Dont (re-)load the model weights if they are already loaded.

[Qwen3.5-9b]
m = o:\models\llama.cpp\Qwen3.5-9B-SOMPOA-heresy-MTP-Q6_K.gguf
mmproj = o:\models\llama.cpp\mmproj-Qwen3.5-9b-SOMPOA-heresy-MTP-BF16.gguf
chat-template-file = o:\models\llama.cpp\qwen_chat_template.jinja

[Qwen3.5-9b-noreason]
m = o:\models\llama.cpp\Qwen3.5-9B-SOMPOA-heresy-MTP-Q6_K.gguf
mmproj = o:\models\llama.cpp\mmproj-Qwen3.5-9b-SOMPOA-heresy-MTP-BF16.gguf
chat-template-file = o:\models\llama.cpp\qwen_chat_template.jinja
chat-template-kwargs = {"enable_thinking":false}
reasoning = 0
reasoning-budget = 0
;ctx-size = 111072
temp = 0.7
top-p = 0.95
top-k = 20

Lets say I want to have the same model with different parameters to it but a full reload is actually not required and would terminate the other models requests even though they are going to the same model.

Motivation

efficiency. running one model on a small GPU with different parameters.

Possible Implementation

store the hash or filepath of the currently loaded model and compare if a reload is required or not.

[Kononnable]

It is not as simple as storing hash/path of a model. Different configuration options may require different layouts in memory - i.e. some layers may be moved between backends to fit in memory, some config changes would require much more (e.g. changing KV data types).

If what you care is just the parameters around reasoning and sampling (based on example) you should change them on per-request basis. In your client app you should be able to define their values and pass them in request to LLM.
Example: https://huggingface.co/unsloth/Qwen3.5-9B-MTP-GGUF#instruct-or-non-thinking-mode

[mcr-ksh]

thats what i ended up doing as a workaround. However, not all apps allow you to specify extra parameters for the llm, thats why such a "trick" would be nice to have the engine set the parameters considering that some parameters like kv cache require a new memory layout, some dont.