xming521/WeClone

AI Summary: The user encountered a `RuntimeError` when initiating training in the `WeClone` application. The error message indicates an unsupported operation where multiple elements of a tensor refer to a single memory location during a write, explicitly suggesting to clone the tensor before performing the operation. This typically occurs in tensor-based frameworks when an in-place modification is attempted on a tensor that shares memory with another.

AI Summary: The task involves fine-tuning a large language model (LLM) using a user's chat history to create a personalized digital avatar. The issue highlights a user's successful, albeit iterative, process of doing this on a Mac, using specific tools and techniques to overcome memory limitations and achieve a partially satisfactory result. The request implicitly asks for confirmation of the process and perhaps suggestions for improvement.

AI Summary: The user is experiencing an error, `AttributeError: Can't pickle local object 'PreTrainedModel.enable_input_require_grads.<locals>.make_inputs_require_grads'`, while running the `weclone-cli train-sft` command to train a custom digital avatar model. This error stems from the inability of the multiprocessing module to serialize a local object during the training process. The task is to debug and resolve this pickling error within the WeClone project, likely involving modifications to the training script or data handling to ensure all necessary objects are serializable.