multiturn_eval

[albertimff]

Summary

• Purpose: reuse the training AgentLoop rollout for single-run multi-turn eval without PPO, decoupling eval from training.
• How to run: vLLM async GSM8K example script with key Hydra knobs, outputs, and checkpoint handling.
• Checkpoints: supports loading FSDP/FSDP2 training checkpoints via built-in DeviceMesh/process-group compatibility patches.
• Validated scenarios: GSM8K, Geo3K, and multimodal; async + vLLM with TP1/TP2 all pass on FSDP checkpoints.
• Extensibility: add custom metrics via aggregate_summary / collect_sample_records or AgentLoop agent_metrics.
• Note: sglang TP compatibility is still missing (tp=1 fails, multi-TP untested); to be fixed later.
• No source code changes—documentation-only addition.

Reviewer Notes

• Please skim the new doc for correctness of run instructions and limitations.