This repository was archived by the owner on Oct 5, 2026. It is now read-only.
Repository navigation
This repository was archived by the owner on Oct 5, 2026. It is now read-only.
Better suppor for interactive apps in Multi-Node #1228
Description
Activity
or also safe single node execution that allows one rank to trigger compute in others?
Legion already supports this, but there will be limits to its scalability depending on task granularity. You will not be able to program 1000s (and probably not even 100s) of nodes like this.
Opened nv-legate/legate#1008 for one possible solution to this problem. As was mentioned already, providing a single-controller execution mode would be another (orthogonal) way to solve the same problem, but that's a longer-term item.
- added a commit that references this issue
on Dec 17, 2025
When wrapping cuPyNumeric (and legate code in general) in interactive Apps (streamlit) or APIs (FastAPI) cuPyNumeric code will get executed in different order, when clicking in the app or making a request.
This works fine in Single-Node but for Multi-Node the execution of the code will diverge and creates issues.
We should support this scenario. with either a client-server architecture or also safe single node execution that allows one rank to trigger compute in others?