mirror of
https://github.com/ollama/ollama.git
synced 2025-07-29 00:33:10 +02:00
Most tensor backends try to optimize performance by using a lower precision for matmuls. However, some operations (such as kq) on some models are sensitive to this and require full precision.