Originally published at: Llama 3.2 Full-Stack Optimizations Unlock High Performance on NVIDIA GPUs | NVIDIA Technical Blog
Meta recently released its Llama 3.2 series of vision language models (VLMs), which come in 11B parameter and 90B parameter variants. These models are multimodal, supporting both text and image inputs. In addition, Meta has launched text-only small language model (SLM) variants of Llama 3.2 with 1B and 3B parameters. NVIDIA has optimized the Llama…
1 Like
Great article. Learned that the vision language model’s encoder and decoder can be optimized separately. For example, use FP8 post-training on the decoder model. Can disaggregating the encoder and decoder further boost the performance for VLMs? Huge potential!