Hi team, I have a question regarding TimeGPT in AW...
# squadsm
Mariana Menchero
03/07/2025, 5:57 PMHi team, I have a question regarding TimeGPT in AWS.
h
Han Wang
03/07/2025, 6:38 PMfor example, we can ask for the typical timeseries length and parameters they use
Han Wang
03/07/2025, 6:39 PMand we create 200-300 synthetic ts to test on a 8gb machine
Han Wang
03/07/2025, 6:39 PMto see if we can reproduce the issue
m
Mariana Menchero
03/07/2025, 6:40 PMWe actually have some of their data, so we can use that for testing.
h
Han Wang
03/07/2025, 6:54 PMok, so how do we reproduce their issue? How to run timegpt using their spec?
m
Mariana Menchero
03/07/2025, 7:17 PMWe have almost everything (data and code), so if it is ok with you, we can work on that next Monday.
h
Han Wang
03/07/2025, 7:34 PMOh I am totally ok
Han Wang
03/07/2025, 7:35 PMI am just curious how you are going to run timegpt using a 8gb machine that has a gpu?
m
Mariana Menchero
03/07/2025, 8:01 PMyes, I was planning to do two things:
1. Ask the client for details on their AWS configuration.
2. Try to replicate their issue using the instances Nixtla has available to see if we hit the same memory limitations.
Mariana Menchero
03/07/2025, 8:03 PMand as I said before, I think this is a difficult case to debug, but they’re one of our new enterprise clients, so we’re trying to help them implement TimeGPT in production asap
h
Han Wang
03/07/2025, 8:05 PMThe reason I am asking is I think we may need a simple way to create similar environment in order to debug issues
Han Wang
03/07/2025, 8:06 PMOne thing I have tried is to use databricks notebook, you can select whatever aws instance with gpu
Han Wang
03/07/2025, 8:07 PMbut how to pull a certain version of timegpt is still a question
Han Wang
03/07/2025, 8:08 PMalternatively, I was thinking that given a timegpt docker, can we quickly deploy to aws ecs and test on it interactively?
m
Mariana Menchero
03/07/2025, 8:15 PMyes, I think we can do that. Plus, we already have some of the client's data and even their code, as they sent it to us for review.