integrations.expert_parallel.experts_fn
integrations.expert_parallel.experts_fn
The expert_parallel experts implementation for ALL_EXPERTS_FUNCTIONS.
Dispatches tokens to the ranks owning their experts (DeepEP or torch all-to-all), runs
whichever experts implementation the user configured on the received rows, and combines
the results back. Slots owned by another rank reach the local implementation as expert id
num_experts with weight 0.
Functions
| Name | Description |
|---|---|
| register_all | Register expert_parallel in ALL_EXPERTS_FUNCTIONS and whitelist it. |
| resolve_local_implementation | The experts forward registered under name (default: the configured one). |
register_all
integrations.expert_parallel.experts_fn.register_all()Register expert_parallel in ALL_EXPERTS_FUNCTIONS and whitelist it.
resolve_local_implementation
integrations.expert_parallel.experts_fn.resolve_local_implementation(name=None)The experts forward registered under name (default: the configured one).