axolotl

You can use a YAML configuration file to drive LLM fine-tuning, which abstracts away manual tokenization, masking, and distributed training setup. This approach allows you to reliably define training workflows by specifying your model, dataset format, and tuning parameters in one place.

Yes, configuration-driven fine-tuning supports LoRA, QLoRA, FSDP, and DeepSpeed strategies for model adaptation. It enables you to launch multi-GPU distributed training workflows seamlessly by defining the desired training backend and strategy within your YAML configuration.

LLM fine-tuning supports pretraining, instruction, chat, preference, stepwise, and multimodal dataset formats. You can supply Hugging Face-style datasets directly into your configuration to handle preprocessing and masking for supervised fine-tuning or preference optimization workflows.

Yes, Axolotl supports multimodal adaptation and inference alongside standard language model workflows. You can configure multimodal dataset formats and training parameters in your YAML file to fine-tune vision-language models using the same configuration-driven lifecycle.

The best way to merge and quantize a fine-tuned LLM is through a unified training lifecycle that handles post-training export natively. After running your fine-tuning workflow, you can use the same configuration tooling to merge adapters and apply quantization before deployment.

You do not strictly need DeepSpeed or FSDP, but they are supported as PyTorch training backends for distributed GPU setups. You can select your preferred distributed training strategy within the YAML configuration to match your hardware scale and fine-tuning requirements.