* add 128k yarn context for Qwen * added property for model tensors * removing useless line
* imatrix-combine-only idea * ensured that behavior consistent with log