Surface oMLX configuration settings in setup flow

[eeshansrivastava89]

Context

oMLX exposes per-model configuration via its app UI and admin API:

Sampling: max tokens, temperature, top-p, top-k, min-p, repetition penalty, presence penalty

Advanced: thinking budget, force sampling

Experimental: turbo quant, kv cache, spec pre-fill, dflash, native MTP, VLM MTP

Currently, offgrid-ai's oMLX setup flow (configureManagedProfile in profile-setup.mjs) only picks a model and saves. The full glass-box setup flow (explanation cards, prompts, defaults) is only implemented for llama.cpp.

What to do

• Use the oMLX admin API to configure per-model settings during setup, same glass-box pattern as llama.cpp
• Surface sampling parameters with explanation cards and sensible defaults
• Surface advanced/experimental settings with appropriate warnings
• Investigate which settings are per-model vs. server-level via the admin API
• Maintain parity as oMLX evolves (check omlx --version and API changes)

Design considerations

• oMLX is a managed server — unlike llama.cpp where we build the server command, oMLX has its own settings layer
• Need to determine which settings can be configured via the admin API (/admin/api/models/<id>/settings) vs. only through the app UI
• Some experimental settings may not have API endpoints — handle gracefully
• Keep the glass-box principle: every setting gets an explanation, not just a prompt

References

• MTP enable/disable already implemented via admin API: ensureOmlxMtpSetting in process.mjs
• llama.cpp glass-box flow: configureLocalProfile in profile-setup.mjs (17 prompts with explanation cards)
• oMLX admin API: PUT /admin/api/models/<id>/settings