Michael Limberger

Troubleshooting

When it breaks

Something is broken. Maybe the pull stalls, maybe generation is glacial, maybe your character started sounding like a generic help-desk bot. This part is organized by symptom so you can fix the common Magidonia setups without guessing.

Model will not download

Symptom: the pull starts, then stalls or fails. Usual causes: disk space, network, or HuggingFace congestion.

Q8_0 is about 25GB. You need that much free, plus a little buffer. On Linux check free space; on Mac use About This Mac, then Storage. Confirm you can reach HuggingFace:

ping huggingface.co

Ollama resumes interrupted downloads. Run the same pull again:

Show me

ollama pull hf.co/bartowski/TheDrummer_Magidonia-24B-v4.3-GGUF:Q8_0

HuggingFace can be slow in peak US afternoon and evening. Try off-peak, or fall back to a smaller quant:

Show me

ollama pull hf.co/bartowski/TheDrummer_Magidonia-24B-v4.3-GGUF:Q6_K

Model runs slowly

Symptom: several seconds per token. Usual causes: GPU not engaged, memory pressure, oversized context, or Rosetta on Apple Silicon.

Check Activity Monitor (Mac) or Task Manager (Windows) and confirm Ollama is using GPU memory. On Apple Silicon, Ollama should be native ARM, not Rosetta. Close browsers and other large apps if memory is full. Try Q6_K if Q8_0 is too heavy for speed, and shrink context if you do not need 16K:

PARAMETER num_ctx 8192

Confirm Ollama is actually running (menu bar on Mac, or ps aux | grep ollama on Linux).

Character breaks or goes generic

Symptom: after twenty or thirty messages the voice collapses into generic chatbot. Usual causes: vague or bloated card, missing refresh, context window too large, or a bad sample that was left in history.

Check the card: more than about 1500 tokens is often bloated, and "helpful and smart" is not a person. After twenty to thirty messages, inject a short trait reminder. Keep num_ctx in the 8K to 16K range for character work. Keep repeat_penalty at 1.0. Regenerate a bad last turn, or edit an out-of-character line so the history teaches the right tone.

Model refuses content

Symptom: "I can't do that" on creative fiction Magidonia usually handles. Usual causes: leftover negative rules in the card, UI safety layers, or a base model that is more filtered than Magidonia.

Remove "you must refuse" style lines and reframe positively (Section 08). Check OpenWebUI filters if any are enabled. For stubborn cases, try a heretic Cydonia variant, or nudge temperature slightly (for example 0.9) if refusals look like uncertainty rather than a hard filter.

Show me

ollama pull hf.co/mradermacher/Cydonia-24B-v4.3-heretic-v2-i1-GGUF:Q8_0

Repetitive responses

Symptom: the same phrase or idea loops. Usual cause: repeat penalty above 1.0, or DRY/XTC samplers fighting Magidonia.

PARAMETER repeat_penalty 1.0

Disable DRY and XTC if present. Raise temperature slightly (1.1 or 1.2) and vary your prompts. Make sure example dialogues are not all the same length and tone.

Model generates both sides

Symptom: the model answers, then invents your next line. Fix: stop sequences.

Show me

PARAMETER stop "User:"
        PARAMETER stop "\nUser:"

In OpenWebUI, add the same entries under Advanced, Stop Sequences. For named characters, also stop on their name (Mark, Nyx, and so on).

OpenWebUI does not show the model

Symptom: pull succeeded, but the model is missing from the dropdown.

Confirm Ollama is running and the model is listed:

Show me

ollama list

Confirm the API is reachable:

Show me

curl http://localhost:11434/api/tags

In OpenWebUI, Admin, Settings, API Connections, Ollama should point at http://localhost:11434. Restart OpenWebUI and hard-refresh the browser if needed.

Drift in long conversations

Symptom: fine for the first thirty messages, then forgets who they are.

Inject a character refresh a few turns back, summarize and prune, or start a new chat with a short "so far" note. Do not force every topic into one endless thread.

Output too short or too long

Symptom: one-liners when you want scenes, or walls of text when you want dialogue.

Adjust max tokens (for example 1024 shorter, 2048 default, 4096 longer). More importantly, match example reply length and first-message length to the style you want. The model copies that template.

Quick checklist

Character broken?

☐ repeat_penalty = 1.0?
        ☐ Stop sequences set ("User:", "\nUser:")?
        ☐ Context window reasonable (8K-16K)?
        ☐ Card specific (not "helpful and nice")?
        ☐ Character refresh after ~30 messages?

Model slow?

☐ GPU being used (check Activity Monitor)?
        ☐ Other apps closed?
        ☐ Ollama native version (not Rosetta on Mac)?
        ☐ Context not oversized?
        ☐ Quant matches your hardware?

OpenWebUI issues?

☐ Ollama running?
        ☐ Model in ollama list?
        ☐ OpenWebUI connected to http://localhost:11434?
        ☐ Restarted OpenWebUI after the pull?

Most of these have one-line fixes once you name the symptom. Next is a single-page quick reference you can keep open while you build.