vllm.distributed.eplb ¶

Expert parallelism load balancer (EPLB).

Modules:

• async_worker –

  The async worker that transfers experts in the background.

• eplb_communicator –

  EPLB communicator implementations and factory.

• eplb_state –

  Expert parallelism load balancer (EPLB) metrics and states.

• eplb_utils –

  Utility functions for EPLB (Expert Parallel Load Balancing).

• policy –
• rebalance_execute –

  The actual execution of the rearrangement.