You may have to make use of the gpu_memory_limit and/or lora_on_cpu config alternatives to prevent jogging away from memory. If you still run away from CUDA memory, you'll be able to attempt to merge in method RAM https://bookmark-template.com/story19994623/https-im-token-io-com-secrets