[Bug] Applying profile with presence_penalty to VLM-MTP-enabled model returns HTTP 500

[jakeuj]

Summary

Applying a sampling profile that sets a non-zero presence_penalty to a model whose base settings have vlm_mtp_enabled: true returns an unhandled HTTP 500 from the admin API.

The incompatibility itself appears intentional because the VLM MTP decode path cannot apply the per-request logits processor. The problem is that the profile UI/API allows the conflicting effective configuration to reach ModelSettings construction and exposes the validation ValueError as an internal server error.

Environment

• oMLX: 0.6.1
• Installation: macOS App
• macOS: 26.6.1
• Architecture: Apple Silicon (arm64)
• Model: Qwen3.8-27B-8bit

Reproduction

1. Enable vlm_mtp_enabled in the model's base settings.
2. Create or select a profile with presence_penalty: 1.5.
3. Apply the profile from the admin UI.

Relevant effective settings:

{
  "base_model_settings": {
    "vlm_mtp_enabled": true
  },
  "profile_settings": {
    "presence_penalty": 1.5,
    "repetition_penalty": 1.0
  }
}

Actual behavior

The request fails:

POST /admin/api/models/Qwen3.8-27B-8bit/profiles/qwen3-r-general/apply -> 500

ValueError: vlm_mtp_enabled cannot be combined with presence_penalty; these
settings require per-request logits processors, which the vlm_mtp decode path
does not apply

The failure occurs while constructing the merged ModelSettings, before the new settings are assigned or persisted, so the profile is not partially activated.

Expected behavior

One of the following would make the failure actionable:

1. Prevent the incompatible combination in the UI and explain the tradeoff.
2. Return HTTP 400/422 with the validation message instead of HTTP 500.
3. Resolve the merge consistently with the existing conflict-resolution paths, for example by disabling the speed-only vlm_mtp_enabled setting when an output-shaping penalty is applied.

The report is about profile validation and API/UI error handling, not about removing the VLM MTP restriction.

Source observation

The current implementation has resolve_vlm_mtp_conflicts() for some saved-settings and runtime-profile merge paths. ModelSettingsManager.apply_profile() appears to pass its merged dictionary directly to ModelSettings.from_dict(), allowing this expected validation error to propagate through the admin route as HTTP 500.

[jakeuj]

Moved to issue #2801 for canonical bug tracking: #2801. Closing this Discussion to avoid splitting updates between two threads.