This allows for setting the --no-context-shift value in llama-imatrix which is required for models like DeepSeek