Could you detail what you mean by deploying LLMs ? Is it about integrating commercial LLMs in an enterprise context? Or running self-hosted LLM for a small company (e.g. Ollama + Ollama Web UI)? Or integrating an Agentic approach to existing software stack?
However, the issue you will quickly encounter is resources/costs. For a simple mode like llama3-7b you need at least a g5.2xlarge on AWS. If you want a 'chat gpt equivalent' model, you need something like llama3-70b or command-r-plus etc. These will require at least a g5.48xlarge that will cost you $20 an hour
An alternate approach is going hybrid: self-hosted UI which takes care of user access, shared documents (RAG), custom prompts etc but that you hook to a LLM provider where you pay per token (could be OpenAI platform or anything from Huggings).
Let me know if this helps! Also note that I'm lead dev of an open source project addressing these kind of needs: https://opengpa.org - Feel free to jump on our Discord to discuss.
Agreed, this works great for a single or small set of users (2-3), but OP is asking about self-hosting an open-source / open-weight LLM locally to serve a larger set of users (100s per post title). I suspect they may need a larger server farm with a lot of GPUs to handle the load.
However, the issue you will quickly encounter is resources/costs. For a simple mode like llama3-7b you need at least a g5.2xlarge on AWS. If you want a 'chat gpt equivalent' model, you need something like llama3-70b or command-r-plus etc. These will require at least a g5.48xlarge that will cost you $20 an hour
An alternate approach is going hybrid: self-hosted UI which takes care of user access, shared documents (RAG), custom prompts etc but that you hook to a LLM provider where you pay per token (could be OpenAI platform or anything from Huggings).
Let me know if this helps! Also note that I'm lead dev of an open source project addressing these kind of needs: https://opengpa.org - Feel free to jump on our Discord to discuss.
Do they want near-realtime responses? Will they all hit it at the same time? Can you put some workloads in an overnight batch queue?