One nice thing about Ollama vs. stock llama.cpp is Ollama supports both ggml and gguf models. If you've still got a lot of old ggml bins around you can easily create a model file and use them.
I haven't done benchmarking vs. vLLM, but it's quite fast; in my tests on an A100-80g w/ llama2 70b I was getting over 25 tok/sec which is just mind blowing.
I was even getting around 30 tok/sec on llama2 7b on an old RTX 1070, which is equally crazy.
Comments
One nice thing about Ollama vs. stock llama.cpp is Ollama supports both ggml and gguf models. If you've still got a lot of old ggml bins around you can easily create a model file and use them.
I haven't done benchmarking vs. vLLM, but it's quite fast; in my tests on an A100-80g w/ llama2 70b I was getting over 25 tok/sec which is just mind blowing. I was even getting around 30 tok/sec on llama2 7b on an old RTX 1070, which is equally crazy.