When it doesn't work
The failures people actually hit, with fixes ordered cheapest-first in quality terms rather than alphabetically.
- ComfyUI runs out of memory
- A diffusion pipeline is several models at once, and the text encoder is usually the one to move.
- Ollama is slower than expected
- Usually the model didn't fully fit, and Ollama ran it anyway across GPU and system RAM.
- CUDA out of memory
- The model, its KV cache and the runtime's working buffers don't fit in VRAM together.
- It loaded, but generation is painfully slow
- Part of the model is in system RAM. Spilling costs 5–20×, not 10%.
- The GGUF file won't load
- Your runtime is older than the model's architecture, or the file is a partial download.