What does this PR do ?

This PR introduces a "Router Replay" feature for Mixture-of-Experts (MoE) layers. This functionality provides a deterministic routing mechanism, which is essential for debugging, controlled experimentation, and reproducing model behavior.

Inspired by recent approaches in stabilizing MoE models Router Replay(R2) and Rollout Router Replay(R3), RouterReplay implementation allows developers to easily save and set the router's replay information, providing precise control over the expert selection process to mitigate routing inconsistencie

Implementation Details:

1. Configuration Flag:

• A new boolean flag, enable_routing_replay , has been added to TransformerConfig . This allows users to enable or disable the feature globally.

2. RouterReplay Class:

• A new class, RouterReplay , is introduced in moe_utils.py to manage the state and data for the replay functionality.
• It operates in three modes, defined by the RoutingMode enum:
  • None : The default state, where standard routing occurs.
  • RECORD : The router records the expert indices ( topk_ids ) for each token.
  • REPLAY : The router bypasses the standard Top-K logic and instead uses the previously recorded expert indices to route tokens.

3. Integration with TopKRouter :

• In router.py , the TopKRouter now initializes a RouterReplay instance if config.enable_routing_replay is True .
• This instance is then passed to the topk_routing_with_score_function during the routing process.

4. Core Logic in moe_utils.py :

• The topk_routing_with_score_function has been updated to handle the router_replay object.
• When the mode is RECORD , it captures the topk_ids after they are computed.
• When the mode is REPLAY , it retrieves the stored topk_ids from the RouterReplay object and uses them to construct the routing gates and probabilities, effectively bypassing the dynamic routing calculation.

⚠️ For major changes (either in lines of code or in its impact), please make sure to first share discuss a design-doc with the team.

Contribution process

flowchart LR
    A[Pre-checks] --> B[PR Tests]
    subgraph Code Review/Approval
        C1[Expert Review] --> C2[Final Review]
    end
    B --> C1
    C2 --> D[Merge]

Pre-checks

• I want this PR in a versioned release and have added the appropriate Milestone (e.g., Core 0.8)
• I have added relevant unit tests
• I have added relevant functional tests
• I have added proper typing to my code Typing guidelines
• I have added relevant documentation
• I have run the autoformatter.sh on my PR

Code review

The following process is enforced via the CODEOWNERS file for changes into megatron/core. For changes outside of megatron/core, it is up to the PR author whether or not to tag the Final Reviewer team.

Merging your PR

Any member of core-adlr and core-nemo will be able to merge your PR.