[Roadmap][VLM] Support VLM Multi-Turn

TODOs for VLM Multi-Turn Support

Generalize multi-turn support for both FSDP and Megatron.
All implementation should remain under examples/.

Single-turn reference: https://github.com/THUDM/slime/pull/501

Dataset

• Add support for OpenCUA: https://opencua.xlang.ai/

Rollout (--custom-generate-function-path)

Create a custom rollout function to support multi-turn interactions.

Multi-Turn Config

• Add configs such as max_turns.
• Add early-stop logic (e.g., stop on max tokens).

Loss Mask

(ref: https://github.com/THUDM/slime/blob/v0.2.0/slime/ray/rollout.py#L236).

Potential behavior:

• Prepare tokens and loss_mask inside the custom rollout.
• Initialize with empty tokens and empty loss_mask.
• For each input turn:
  • Append input tokens to tokens.
  • Append 0s of equal length to loss_mask.
• For each generated output:
  • Append output tokens to tokens.
  • Append 1s of equal length to loss_mask.

Start in examples/ and review the single-turn reference in PR 501, then compare the rollout behavior at slime/ray/rollout.py#L236. Check the OpenCUA dataset requirements and the custom rollout entry point. Done means multi-turn support works for both FSDP and Megatron with turn limits, early stopping, and input/output loss masks.