Before llama.cpp, there was whisper.cpp, a port of OpenAI's whisper and the initial foundation for ggml. Also before llama were gpt-neo and gpt-j (IIRC, was the first widely deployed model to use RoPE, which llama also implemented). llama also might have borrowed from nvidia's megatron, my memory is vague there. Thanks to chatgpt 3.5, llama's release was well placed to channel rising enthusiasm about LLMs into catapulting the local scene to the next level (before this it was relegated to just researchers, ML engineers and hardcore AI enthusiasts).

Which is to say, there was a local LLM ecosystem before llama-1 and for practical uses of the time, Google's FlanT5-11B was better. Llama-1 was also initially released only to a selected few researchers, until it was leaked on 4chan, after which LeCun was key in championing a move to proper opensource. Without that leak and LeCun's subsequent championing of opensource LLMs, I'm not sure how the open LLM scene would have gone and what Meta's role would have been.