Hey everyone! I've been tasked with looking into Tabnine Enterprise for our team. We need to keep everything on-premise for security reasons. I've seen the high-level docs, but I'm a bit lost on the actual step-by-step setup.
Could someone walk me through the main stages? Like, from getting the initial package to having it actually running and connected to our IDE? I'm especially unsure about the hardware requirements for the private model part. Do we need a beefy GPU server, or can it run on our existing Kubernetes cluster? 😅
Also, any gotchas with the licensing or the initial configuration file would be super helpful. I don't want to hit a big blocker after spending a day on installation.
Still learning
Oh, I'm looking into something similar! We're also considering the on-prem option.
The GPU question is a big one for me too. From what I gathered talking to their sales team, you *can* run the private model training on a powerful CPU setup, but it'll be extremely slow. They really push for a dedicated GPU server for that part, which got expensive fast in our quotes.
A gotcha I heard about from a colleague was the license key being tied to a specific domain or IP address list in the config. If you mess that up during initial setup, the whole thing just won't connect. Make sure you have your exact internal access URLs finalized before you start.
That GPU question is a real blocker for us too. We got a trial going without dedicated GPUs, but like user196 said, it was painfully slow for the private model training. Our team lost patience waiting for it.
Our biggest gotcha was the config file for the initial package. The documentation says to specify your IDE client URLs, but it's easy to miss the exact format. We had to reach out to support because our connection kept failing. Definitely get that list right before you start.
Did you already get the deployment package? I'm curious if your download included the pre-built container images or if you had to build them yourself. That was another unexpected step for us.
Yeah, the hardware requirements tripped us up as well. We tried using our existing K8s cluster, but the private model training just crawled. It did work, but you're talking days instead of hours. A dedicated GPU node ended up being necessary for us to make it usable.
For the stages, once you get the package, it's mostly about that initial config YAML file. Double-check the `license_key` and `client_origins` sections before you even run the first install script. Getting those wrong means starting over.
Did you get clarity on whether the training happens automatically after deployment, or is that a separate manual step you have to trigger?