[FEAT] - Better DMR integration

[krissetto]

What you'd like to see

Better support for DMR's feature-set, similar to the support compose provides (e.g. for runtime parameters and context length)

Why you'd like to see it

To better take advantage of local models

Workarounds?

Creating the local model runtimes with a compose file, then pointing the base url in the agentfile to the appropriate urls exposed by compose (which might even need to be proxied to localhost, not sure atm)

Additional context

I'd want to be able to:

models:
  localboi:
    provider: dmr
    model: ai/qwen3:14B-Q6_K
    ctx_length (or max_tokens): 32000
    runtime_opts: |
      <llama.cpp_args_here>
      ...
   ...
...

[vvoland]

Yes!!

[krissetto]

@vvoland @kiview have fun 😉