Generating output from finetuned model using LLM.generate() method vs `litgpt chat` cli command

Nobody has claimed this yet.

I performed lora finetuning on Qwen2.5-3B-Instruct model on my custom dataset, and it seems that when I generate content using the `litgpt chat` cli interface the finetuned model performs flawlessly. However, when using it through the python api, following the [tutorial](https://github.com/Lightning-AI/litgpt/blob/main/tutorials/python-api.md#generatechat) the model outputs are random characters and basically rubbish. I tried using `litserve` and converting it into HuggingFace models but none worked.

Does anyone have any idea how to fix this.

P.S. Currently I am using a custom implementation of [`litgpt\chat\base.py`](https://github.com/Lightning-AI/litgpt/blob/main/litgpt/chat/base.py) to get my code working but this is pretty slow and I wish to parallelize the generation.

Thanks for any help.

Start with the Python API generate-chat tutorial and compare its generation path with litgpt/chat/base.py and the litgpt chat CLI. Reproduce the Qwen2.5-3B-Instruct LoRA case, checking why the Python API produces random characters while the CLI works. Done means the Python API generates coherent output consistent with the CLI without relying on a custom chat implementation.

Written by the indexing model from the issue text.