feat(draft) - Add Dynamic CP Support

[humairafirdowse18]

Title

Add initial Megatron dynamic context parallel support

Description

This draft adds dense-model Dynamic CP support to NeMo RL’s Megatron backend while keeping Ray responsible for data distribution.

The Ray driver schedules packed samples across the DP×CP domain and sends workers serializable rank/CP plans. Workers resolve the corresponding MCore process groups locally. Dynamic CP outputs are collected from all CP lanes, reordered into the original sample order, and normalized using unique global token counts.

The current implementation:

• Supports active CP sizes including CP1 and runtime CP transitions.
• Handles sequence packing, padding, logprobs, loss scaling, and output ownership.
• Uses zero-masked placeholders to keep forward/backward collective ordering aligned.
• Adds a Ray tensor serializer to avoid requiring MCore in the driver.
• Supports dense models with PP1; MoE and PP>1 remain follow-up work.

Validation

• Distributed logprob, loss, gradient, and optimizer-update parity tests
• Active CP1/2/4 with static CP1/2 and TP1/2
• Qwen3-32B: 10 GRPO steps on 4 nodes × 4 GB200 GPUs
• Nsight traces from all 16 policy workers
• Finite metrics and expected importance-ratio/KL values

[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.