NeMo RL has used the driver to deal with data and control. This has led to the driver being a bottleneck for data movement since all data funnels through one rank. This is no longer tenable at the scales we are now training at.
This issue tracks the separation of data plane from the driver.
NeMo RL has used the driver to deal with data and control. This has led to the driver being a bottleneck for data movement since all data funnels through one rank. This is no longer tenable at the scales we are now training at.
This issue tracks the separation of data plane from the driver.