search

skypilot-org / skypilot-catalog

20 stars 30 forks source link

[Runpod] Use consistent naming for A100s #64

Open romilbhardwaj opened 7 months ago

romilbhardwaj commented 7 months ago

On runpod, A100-80GB is listed separately asA100-80GB-SXM. Since SkyPilot doesn't currently distinguish between PCIE and SXM4 variants, we should probably change A100-80GB-SXM to A100-80GB in our catalog.

image image
Michaelvll commented 7 months ago

On runpod, A100-80GB is listed separately asA100-80GB-SXM. Since SkyPilot doesn't currently distinguish between PCIE and SXM4 variants, we should probably change A100-80GB-SXM to A100-80GB in our catalog.

image image

On RunPod, there are both PCIE and SXM version on the same region. Changing to A100-80GB will make it fail to choose those SXM instance types unless a user specify the instance type directly. Does that sound good to you?

romilbhardwaj commented 7 months ago

I see... I was confused because A100-80GB is SXM4 on gcp, but looks like on runpod we will provision PCIE unless user explicitly asks for A100-80GB-SXM.

For consistency, should we make A100-80GB on runpod correspond to the SXM4 instance, and list A100-80GB-PCIE separately?