Skip to content

[gemma4]: track 26B-A4B shared MoE L1-L2 integration #6573

Open

Description

@richardoo-707

Problem

Phase-two task 2C3 requires Gemma 4 26B-A4B L1-L2 validation using the shared MoE implementation delivered by #6477. The current Gemma policy rejects MoE models, and the numerical model registry only includes dense E2B/E4B.

Scope and acceptance

  • Preserve the native Gemma router, GELU, dense/routed branches, parameter identities and HF checkpoint keys while reusing shared dispatch and grouped GEMM.
  • Explicitly keep EP=1 inside DP process groups; validate single/two-GPU training, lazy initialization and non-reentrant checkpointing.
  • Register a structural tiny model with an independent unmodified HF reference, token-weighted DP, exact resume/HF readback and a standard L2 report.
  • Match HF autocast and expert-contribution rounding. A declared HF blank self-control validates the reference recipe; independent checks must continue to reject a faulty candidate.
  • Retain dense E2B/E4B and checkpoint regressions.

This task covers structural tiny text-model L1-L2, not full-size 26B training, convergence, performance, multimodal behavior or EP/TP/PP. Manual same-SHA validation is complete; PR CI and merge are tracked separately in the linked implementation PR.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions