Comment by joaquincabezas

are there also experiments around image embedding generation to use in combination with the LLM?

maybe for this use-case is it better to execute the vision tower on a GPU and leave the LPU for the language part?