[roadmap] TorchTitan Integration into Verl

[acisseJZhong]

Roadmap: Integrate TorchTitan Training Engine into Verl

We want to integrate Torchtitan as one of the training engine backends; and work seamlessly with pytorch distributed offerings.

Infrastructure Components

• Model Engine Implementation [trainer] feat: Add Torchtitan as alternative training engine #5051
  • Support text model
  • Support multimodal model
• Torchtitan sharding manager

Parallelism

• TP/SP
  • TP work with varlen attention
  • TP work with flex attention
• PP
  • relevant discussions Support pre-split microbatches in distributed pipeline parallelism pytorch/torchtitan#2370, Question about Pipeline parallelism pytorch/torchtitan#431
• EP
• CP

E2E Validation

• Integrate with Verl's SFT trainer [trainer] feat: Add Torchtitan as alternative training engine #5051
• Integrate with Verl's RL trainer