NVIDIA-Certified Associate: Generative AI LLMs (NCA-GENL) Practice Test

A team needs to fine-tune a 70B-parameter base model to follow a new product's instruction style. They have a single 8x H100 node available. Which fine-tuning approach is most feasible on that hardware?

• AFull fine-tuning at full precision
• BLoRA (Low-Rank Adaptation) with rank 8 to 32, applied to attention projection matrices, with the base model in 4-bit quantization (QLoRA)
• CTrain a new 70B model from scratch on the same hardware
• DDistill the 70B model into a 7B model on a single GPU

Keyboard shortcuts: Off