Self-Hosting GPU-Accelerated LLM (Mistral 7B) on Kubernetes (EKS)
1–4 of 4 posts
Re: Self-Hosting GPU-Accelerated LLM (Mistral 7B) on Kubernetes (EKS)
#2Guide based on my experience running Mistral 7B LLM on EKS. Warning, some opinionated tech choices: AWS, EKS, Karpenter, NVIDIA GPUs, Hugging Face.
If you try this, be sure not to forget the GPU nodes sitting idle!
Re: Self-Hosting GPU-Accelerated LLM (Mistral 7B) on Kubernetes (EKS)
#3Guide based on my experience running Mistral 7B LLM on EKS. Warning, some opinionated tech choices: AWS, EKS, Karpenter, NVIDIA GPUs, Hugging Face. If you try this, be sure not to forget the GPU nodes sitting idle!
Hey thanks for this article. It covered a lot of things I had to figure out myself, at great pain and over the course of many weeks, to get a similar-ish workload running on EKS. It also has some ideas that I hadn’t really thought about, especially around using PVs to amortize the cost of model fetching across pod lifecycles. I like that an awful lot.
Re: Self-Hosting GPU-Accelerated LLM (Mistral 7B) on Kubernetes (EKS)
#4Guide based on my experience running Mistral 7B LLM on EKS. Warning, some opinionated tech choices: AWS, EKS, Karpenter, NVIDIA GPUs, Hugging Face. If you try this, be sure not to forget the GPU nodes sitting idle!
Hey thanks for this article. It covered a lot of things I had to figure out myself, at great pain and over the course of many weeks, to get a similar-ish workload running on EKS. It also has some ideas that I hadn’t really thought about, especially around using PVs to amortize the cost of model fetching across pod lifecycles. I like that an awful lot.
Thanks for the feedback. Glad you got something out of it.
> covered a lot of things I had to figure out myself, at great pain
My starting point for this was from Hugging Face docs, which don't really offer much for how to deploy to a k8s environment. Even the fact that you need GPUs for the model I was trying to run was not immediately apparent to me from the Mistral 7B HF docs (I'm sure this can vary a lot for different models).
> PVs to amortize the cost of model fetching across pod lifecycles
I'd love to pull more on that thread and figure out how to build a production quality inference service.