Remove unused RolloutServer.all_engines for encapsulations

[fzyzcjy]

No description provided.

[gemini-code-assist]

Code Review

This pull request removes the all_engines property from the RolloutServer class. Feedback indicates that this removal is premature, as the property is necessary for maintaining consistency across all ranks in multi-node configurations and provides better encapsulation for external callers compared to accessing internal server groups directly.

I am having trouble creating individual review comments. Click here to see my feedback.

miles/ray/rollout/rollout_server.py (161-163)

Removing all_engines from RolloutServer appears to be premature and potentially counter-productive to the goal of encapsulation for the following reasons:

1. Correctness in Multi-node: Currently, methods like offload, onload, and check_weights in ServerGroup (and the weight update logic in RolloutManager) rely on the engines property, which only includes node-0 actors. In multi-node Tensor Parallel (TP) configurations where nodes_per_engine > 1, this results in operations only being applied to a subset of actors, leaving others (on different nodes) in an inconsistent state. all_engines is the correct collection to use for these operations to ensure all ranks are reached.
2. Encapsulation: In RolloutManager._try_ci_fault_injection, the code currently reaches into self._server.server_groups[0].all_engines. If RolloutServer.all_engines were available, it would provide a cleaner, higher-level interface that hides the internal group structure. Removing it forces external callers to depend on the internal server_groups implementation details.

I recommend retaining this property and using it to address the existing multi-node inconsistencies in the rollout orchestration logic.