Deterministic true on-policy GRPO wit the inference-optimized megatron generation (EP8/TP1)

[utkarsh530]

Generation bitwise-identical to training/scoring (gen_kl == 0.0 x18 full-lr, prod bed, 1.02x throughput tax; cert job 2333154). Requires the companion Megatron-LM branch utkarsh530/Megatron-LM:det-inference-ep8tp1-certified. Tested model: Qwen30B-A3B

NeMo-RL side:

• megatron_generation: fail fast when colocated generation would silently ignore model-side mcore_generation_config keys
• grpo/utils: non-colocated (dedicated-node) megatron generation
• setup: batch-invariant activation hook + det attention num_splits pin
• train: fused logprob scoring path (float -> log_softmax -> gather) + per-EP-rank fp64 combine mirror env gates
• megatron_worker: det wiring + refit param-gather sync barrier
• router_replay: megatron-backend support (feature OFF in the recipe)
• recipe YAML (certified config) + gen_kl_harness determinism gate + docs

What does this PR do ?

Add a one line overview of what this PR aims to accomplish.

Issues

List issues that this PR closes (syntax):

Usage

• You can potentially add a usage example below

# Add a code snippet demonstrating how to use this

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you run the unit tests and functional tests locally? Visit our Testing Guide for how to run tests
• Did you add or update any necessary documentation? Visit our Document Development Guide for how to write, build and test the docs.

Additional Information

• ...
  This branch is in active development and will be pushed for further updates. I need to clip out some of the redundant paths and will try to get the relevant paths in the Megatron-LM we're merging to make it completely usable.

[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.