LLM Finetune with PEFT

[avishniakov]

This PR brings in the pipeline to run Mistral model fine-tuning using PEFT library on Viggio dataset.

Key highlights:

• Project has one pipeline to run the full cycle of LLM fine-tuning:
  • Data extraction and preparation
  • Finetuning base model on new data
  • Evaluation of the performances for base and fine-tuned model
  • Promotion of the model to the target environment, if it outperforms the base and currently promoted one
• Model can be trained on local(remote) orchestrator in full
• Model can be trained on step operators, where only GPU intense steps will be pushed to step operator (we used Vertex AI step operator)
• Here we also move around the previous LitGPT fine-tune project to llm-litgpt-finetuning and this new project becomes llm-lora-finetuning

P.S. Diff is a bit off due to project movement. LitGPT one was lift and shift - no changes in it.

[dagshub[bot]]

Join the discussion on DagsHub!

[strickvl]

@coderabbitai review

[avishniakov]

@htahir1 are we ok to merge this? It will replace the current LORA example with PEFT one and LitGPT will be moved to a new directory. Any blockers?

[schustmi]

Just one more thing that came to my mind: We should probably also rename the llm-lora-finetuning template right?