llm-fine-tuning - LLM Fine-Tuning Infrastructure
Set up infrastructure for fine-tuning LLMs using QLoRA, LoRA, Hugging Face TRL, Axolotl, DeepSpeed, or FSDP.
Tags
Updated: 2026-09-21Capabilities
What this skill does
- Fine-tune LLMs using QLoRA
- Configure Axolotl for fine-tuning
- Launch distributed DeepSpeed ZeRO-3 training
- Execute DPO preference alignment training
- Merge LoRA adapters into model
- Deploy Kubernetes LLM training jobs
Inputs
- NVIDIA GPU with 24GB+ VRAM
- Training dataset
- Base LLM model weights
- Hugging Face access token
- Axolotl configuration YAML file
- DeepSpeed JSON configuration file
Outputs
- Fine-tuned model checkpoints
- Merged model and tokenizer files
- Training log metrics
- Kubernetes training job state
Requirements
- NVIDIA GPU with 24GB+ VRAM
- CUDA 12.1+ and working nvidia-smi
- Python 3.10+ with pip
- 500GB+ disk storage
- Privileged system access where noted