feat(ppo): add DeepSeek R1 7B single-controller recipe

[snowmanwwg]

Summary

• add an 8-node x 4-GPU SingleController PPO recipe for DeepSeek-R1-Distill-Qwen-7B on DAPOMath
• allocate four nodes to non-colocated vLLM generation and four nodes to colocated policy/value training
• use 2,048 prompts per RL step with policy and critic global batch sizes of 2,048, satisfying the current SingleController one-optimizer-step-per-RL-step constraint
• configure an 18,432-token context, sequence packing, and no system prompt

Testing

• git diff --check
• YAML syntax parsed successfully
• GPU training not run as part of this PR preparation

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[snowmanwwg]

/ok to test 54d8d7f

[snowmanwwg]

/ok to test 003db41

[snowmanwwg]

/ok to test e0a4514