Misc. bug: --mmproj doesn't respect the -mg GPU specified

[samteezy]

Name and Version

build: d1e2adb (6382)

Operating systems

Linux

Which llama.cpp modules do you know to be affected?

llama-server

Command line

/root/llama-builds/llama.cpp/bin/llama-server
    --metrics
    --port ${PORT}
    --api-key needakey
    -sm none 
    -mg 1
    -t 10
    --swa-full
    --model /mnt/models/unsloth/gemma-3-4b-it-qat-Q4_0.gguf
    --mmproj /mnt/models/unsloth/gemma-3-4b-it-qat-mmproj-F16.gguf
    --jinja
    -ngl 99
    --temp 1.0
    --top_k 64
    --top_p 0.95
    --min_p 0.0
    --ctx-size 32000

Problem description & steps to reproduce

When I have -sm none -mg 1 specified to load a model that also comes with a separate projector file, the file loads on what would be the "0" GPU, even if there's plenty of VRAM to spare on 1.

I'd expect this to respect the -mg 1 setting and have both files load on 1.

First Bad Commit

No response

Relevant log output

[samteezy]

Actually, I'm realizing this may be a duplicate of #15061.

However, I think this also affects -nkvo. I'm attaching a screenshot because printing out amdgpu_top --smi output is giving me formatting hell here... but you can see that the GTT is associated with card 0 while card 1 is actually running the model. Nothing is running on card 0.

This is running Qwen3-4B with a fairly large prompt, and I can see the KV cache memory usage grow on card 0 with -nkvo set. If there's a better way of proving/demonstrating this, let me know.

Is this an actual issue? I don't know if it presents a problem for me at the moment, but it might be introducing unnecessary overhead?

(And yes, I need to figure out a better cooling solution for my V620.)

[hjc4869]

There's an envvar to specify the mmproj device: #14236

The help strings could be improved so that people find this envvar, or adding --mmproj-device might be better as stated in the original PR.