LLM Suite Troubleshooting

Troubleshooting

Engine

llama-server failed to start

The engine card in Settings → Engine, or an error when loading a model, says llama-server failed to start, sometimes with the explanation "A DLL that llama-server.exe needs could not be found (Windows error 0xC0000135)".

llama-server.exe links against DLLs such as the CUDA runtime and OpenSSL. They may work in your terminal but be missing from the smaller PATH an app started from Explorer gets.

RulZ already adds %CUDA_PATH%\bin\x64, %CUDA_PATH%\bin and the usual OpenSSL folders to the PATH it gives llama-server. Check that those folders exist on your machine.
Otherwise copy the missing DLLs next to llama-server.exe. For OpenSSL that is libcrypto-3-x64.dll and libssl-3-x64.dll.
Press Test in Settings → Engine to try again.

Other codes: 0xC000007B means a DLL of the wrong architecture (32-bit versus 64-bit); 0xC0000139 means a DLL of the wrong version.

The engine is "not available"

RulZ couldn't find llama-server.exe. Set the folder that contains it in Settings → Engine and press Test; the card lists every place it looked. See Building the Engine.

No GPUs are listed

The card says No GPU devices reported. Your llama.cpp build was probably made without CUDA (a CPU-only build runs, slowly). Rebuild with -DGGML_CUDA=ON or use a CUDA release, and check that llama-server.exe --list-devices lists your cards.

A model fails with "unknown model architecture"

Your llama.cpp is too old for that model. Update and rebuild llama.cpp, then press Test.

Loading models

Out of memory, or the load fails

Look at the memory estimate in the load panel and the log (the file icon on the model's chip in the top bar).
Lower the context length; the KV cache grows with it.
Set the K cache and V cache to Q8_0.
Turn off the vision projector if you don't need images.
Use Only GPU n if you want the model to stay on one card, or let it split.
Check nvidia-smi for other programs using VRAM. Another LLM app that is still running, a game or a browser can hold gigabytes.
Try a smaller quantization of the model.

It loads but generates slowly

The estimate probably said Partly on CPU. Layers that don't fit in VRAM run on the CPU, which is far slower. Reduce the context or cache size, or pick a smaller model or quantization. See also speculative decoding for models that have a draft head.

My model isn't in the list

Check the folder in Settings → Models, and press Rescan on My Models.
Hidden folders (names starting with a dot) and anything deeper than about eight folders are skipped.
Only the first shard (…-00001-of-0000N.gguf) of a split model is listed.
A file that can't be read shows as Unreadable with the reason in its details.

A projector or draft head isn't picked up

Sidecar files must be in the same folder as the model, and their names must follow the usual pattern (mmproj…, mtp-…). See Model Library.

Chat

Images are ignored

A message says the model has no vision projector loaded. Use a model with a projector (mmproj) file in its folder and keep Vision projector switched on in the load panel, then reload the model.

The model forgot the start of a long chat

The context window filled up and Truncate middle dropped older messages. Load the model with a larger context, use a handoff, or switch the behaviour to Stop and show an error in the chat settings.

"The enabled MCP tools take about … tokens"

The tool descriptions alone use most of the model's context. On the MCP Tools page set Tool loading to On demand (or Auto), disable tools you don't need, or load the model with a larger context. See MCP Tools.

MCP

A server shows an error

Open the server card's Log. Common causes:

npx or uvx isn't found: install Node.js or uv.
A wrong path or missing environment variable in mcp.json; check args and env.
The server needs a login or key it doesn't have.
For remote servers, check the URL and the headers.

Fix it and press the reconnect button. A syntax error in mcp.json is reported when you press Save & reconnect.

The model keeps calling tools

RulZ stops after 12 rounds of tool calls in one reply. Tighten your request, disable tools the model doesn't need, or deny the calls.

Local API server

Port 1234 is already in use

Another program, often LM Studio's server, is using it. Stop that program or change the Port on the Developer page and apply it.

"System message must be at the beginning"

The model's chat template only allows a system message at the start, and the program sent one later. Claude Code does this. Update to 0.9.1 or later, which moves those messages into the system prompt for Anthropic-style (/v1/messages) requests.

"Model … was not found"

The program asked for a model by a name that isn't in your library. Turn on Answer any model name under Model loading on the Developer page, and either load exactly one model or pick a Default model. See A model you don't have.

A model I loaded disappeared

With Settings → Models → Unload the previous model when loading another on (the default), a request that makes the API server load a different model unloads the others first. Turn the setting off to keep several models loaded. Models the server loaded itself are also unloaded after the idle time; see Local API Server.

401 Unauthorized

An API key is set. Send it as Authorization: Bearer <key> or x-api-key: <key>.

Installing

Windows SmartScreen blocks the installer

The installer isn't code-signed. Choose More info → Run anyway.

`npm run build:win` fails with "Cannot create symbolic link"

electron-builder unpacks signing tools that contain symlinks, which Windows only allows with Developer Mode on or an elevated terminal. Enable Developer Mode (Settings → System → For developers) and run the build again.

Still stuck?

Press F12 to open the developer tools and look at the Console, check the model's log from the top bar, and open an issue with the error text. Remove any API keys and tokens first.