That has already improved a lot. Initially they were generating new samples w/ transformers, and were talking in github issues about using vLLM to batch generat

That has already improved a lot. Initially they were generating new samples w/ transformers, and were talking in github issues about using vLLM to batch generate samples. Lower in the blog post it seems they already did that.