Huawei Computing said its AsyncFlow asynchronous streaming RL decouples rollout
(inference) and trainer via a TransferQueue data hub and CheckpointEngine weight
engine, addressing long-tail rollout stalls that leave hardware idle and raise
costs. In tests on Ascend Atlas 900 A3 SuperPoD liquid-cooled and Atlas 800 A3
A3 air-cooled supernodes, training throughput on long-sequence workloads (prompt
2k → response 16k) rose from 59.3 to 226.8 — a 3.81x increase — while
aligned maintaining convergence accuracy. AsyncFlow currently supports the verl
framework.