specific method · filed under optimization
MoonEP
An expert-parallel method that uses dynamic redundant experts with online planning and migration to balance expert execution.
Also called MoonEP expert placement planning.
- source
- 1
- model
- 1
- lab adopt it
- 1
- strongest
- core
How sources treat it
One count per evidence span, weakest treatment to strongest.
Documented in
Evidence
4 spans quoted from the sources, strongest treatment first.
For a router output I, the planning objective is to minimize the maximum number of redundant experts on any rank
MoonEP provides perfectly balanced expert execution with static computation shapes and zero-copy communication
We therefore propose MoonEP, an EP scheme that achieves perfect load balance with dynamic redundant experts. MoonEP preserves the overall computation flow of conventional schemes such as DeepEP and additionally introduces online planning and migration of redundant experts.
The planning objective is to minimize the maximum number of redundant experts on any rank
Filed alongside
Other methods under optimization :: training parallelism.