Fix Ollama CUDA Out of Memory in 5 Minutes

CUDA out of memory is usually not a single problem. It is a budget mismatch between model size, context window, and runtime overhead.

Fast fix order

1. Lower quantization
2. Reduce context size
3. Reduce GPU layers
4. Retry with smaller output length

Why this works

Each step reduces memory pressure from a different axis. Most users only change one variable and stop too early.

Prevent repeated OOM

• Keep a per-model context cap
• Save known-good launch commands
• Use a fit calculator before pulling new large models

The fastest stable workflow is: estimate -> verify -> lock known-safe parameters.