-
Notifications
You must be signed in to change notification settings - Fork 8
Open
Description
Hi,
My use case is that I want to run predictions from existing, large models (like BERT). Currently on GCP I tested following products: CMLE, CloudRun, AppEngine - but all of them have limitations in terms of memory and AFAIK they are not supporting GPU.
That's why, I'm wondering if is it possible to change 'gcp-notebook-executor' for opening/closing existing VM's, instead creating every time new one, which will require time-consuming models download & installation every time.
I think it could be somehow related with #10, however my point is to avoid any downloads (even from GCS bucket), as time for prediction is priority.
PS. I'm open for suggestions, if you know what other solution could help with 'GPU' requiring predictions.
Reactions are currently unavailable
Metadata
Metadata
Assignees
Labels
No labels