Description

This contribution provides an implementation of LLM.merge_and_unload(), which is triggered from LudwigModel.train(), based on configuration. The "merge and unload" behavior merges the fine-tuned LoRA weights into the base model so that the users can load one complete model (e.g., from HuggingFace) in a single AutoModelForCausalLM.from_pretrained() call, rather than using two calls (first loading the base model with AutoModelForCausalLM.from_pretrained() and then loading the fine-tuned weights with PeftModelForCausalLM.from_pretrained()). This capability facilitates portability of the inference function between Ludwig (LudwigModel.load() followed by model.predict() for inference) and others (e.g., HuggingFace using AutoModelForCausalLM.from_pretrained() followed by transformers.pipeline() for inference).

The configuration consists of extending the adapter section with the optional postprocessor section as follows:

adapter:
  type: lora
  postprocessor:
    merge_adapter_into_base_model: true
    progressbar: true

(If the merge_adapter_into_base_model is kept and set to false, then the progressbar directive can be omitted.)

Code Pull Requests

Please provide the following:

• a clear explanation of what your code does
• if applicable, a reference to an issue
• a reproducible test for your PR (code, config and data sample)

Documentation Pull Requests

Note that the documentation HTML files are in docs/ while the Markdown sources are in mkdocs/docs.

If you are proposing a modification to the documentation you should change only the Markdown files.

api.md is automatically generated from the docstrings in the code, so if you want to change something in that file, first modify ludwig/api.py docstring, then run mkdocs/code_docs_autogen.py, which will create mkdocs/docs/api.md .