daekeun-ml/genai-ko-LLM

Covers QLoRA parameter-efficient fine-tuning using SageMaker training instances and multiple model serving approaches including DeepSpeed, TGI (Text Generation Inference), and NVIDIA FasterTransformer for distributed inference. Integrates with HuggingFace Hub and SageMaker's DJL container, supporting various Korean models (KULLM-Polyglot, KoAlpaca) with notebooks for local debugging before production deployment. Includes RAG implementation examples alongside inference optimization techniques.

No commits in the last 6 months.