Troubleshooting

The handful of things that go wrong, and what to do about each.

"Ollama isn't running"

Remuda is a client. If it can't reach a server, it says so with a persistent banner rather than quietly failing every action — mutating actions (load, save, pull, delete) are disabled while it's up. Two ways out:

• Start Ollama, from the banner itself, spawns ollama serve for you. If Ollama isn't installed this fails with the underlying error verbatim (for example, "No such file or directory") rather than a generic message.
• Retry re-runs the health check immediately, for when Ollama was already starting up or you started it yourself in a terminal.

Remuda also health-checks on a timer, so the banner clears on its own once the server answers.

A model spills to system RAM

When a loaded model doesn't fit entirely in VRAM, the runtime chip turns amber and names the split — how much is on the GPU, how much fell back to system RAM. Expect a large drop in tokens per second; a spilled model behaves like a different, much slower model, not a slightly slower one.

Three ways out, in order of effort:

• Eject another loaded model to free VRAM for this one.
• Use a smaller quantisation — Save as… in the Modelfile editor can quantise a variant on the way in.
• Lower num_ctx, since the context window's KV cache is part of what has to fit.

Before loading, the fit predictor can flag a spill in advance, at a chosen context length, so this is visible before you wait out a load.

Context filling up

Each reply reports the context used against the model's num_ctx. As a chat gets long, that figure climbs toward the ceiling; once it's reached, the server starts dropping the oldest turns to make room, which can make a model seem to "forget" earlier context.

num_ctx is a load-time setting, not a sampling one — raising it makes Ollama reload the model with a larger KV cache, which costs more memory and can push a model that fit into a spill. Prompt-eval tokens/second, shown after every reply, is the figure that matters most when deciding whether to raise it: a slow prompt-eval rate means a large context is already expensive.