This is meant to be a general FAQ for the Torch Tuner CLI.
This document is a work in progress, so please be patient.
Why do I receive an exception about chat templates when merging my fine-tuned LoRA adapter?
This error usually occurs when you tuned an adapter with JSONL.
Please rerun the merge command with the --is-chat-model CLI argument set to true.
Why do I receive the following warning when tuning my model?
Setting `save_embedding_layers` to `True` as the embedding layer has been resized during finetuning.
This warning occurs when tuning your model with the --is-chat-model or --use-agent-tokens argument is set to true.
No need to worry, this is because the insertion of new tokens causes the embeddings layer size to change.
Why do I receive the error when resuming tuning of my model?
An unexpected Exception has been caught: loaded state dict contains a parameter group that doesn't match the size of optimizer's group
This error usually occurs when you resume a tuning job with different value
for the --save-embeddings argument or different target modules than that tuning job was initially started with.
How do I extend the torch-tuner CLI to support another LLM type?
If 'generic' LLM type doesn't suit your specific needs, you can follow the pattern found in the modules package to implement a custom LLM module.
- Create a python file in the modules package with the name of the LLM type
- Create the required functions, and implement the functions in the LLM Base Module
- Finally, wireup the new LLM module in the tuner_utils python file