A housing-inspection counter places a refined classifier into the live Triton repository, but the lobby kiosk still sends the old tensor names. How do they roll forward without breaking that kiosk?
Select an answer to reveal the explanation.
Short Explanation
The live kiosk still sends the old tensor names. Add version 2 beside version 1 and point only new clients at 2. Deleting version 1 the same afternoon, merging with NCCL, or turning the classifier into an image generator breaks the lobby.
Full Explanation
Triton versions let a refined model sit beside the old one, so pinned clients keep calling version 1 while new clients move to version 2. Instantly replacing or deleting version 1 breaks the still-running kiosk. NCCL tensor parallelism does not merge serving versions, and turning the classifier into an image generator abandons the inspection-label job. Versioning is the associate roll-forward, not a training rewrite.