[aniskumar-nv]

What does this PR do?

Type of change: Bug fix.

Streaming EAGLE3/DFlash training on meta-models/Muse-Glimmer-30B fails because its text stack is nested as text_config with model_type: muse_glimmer_text, which _FINAL_NORM_TYPE_BY_MODEL_TYPE does not list, so FakeBaseModel cannot reconstruct the base logits from the streamed pre-norm hidden state.

This registers muse_glimmer_text (and muse_glimmer, for a text-only checkpoint) as plain rmsnorm. The model is a Gemma-2 derivative whose per-layer sandwich norms carry the (1 + weight) offset, but its final norm does not: on the released checkpoint, normed * weight reproduces HF hidden_states[-1] at cos 0.999999 (max-abs 0.21, bf16 noise), while the (1 + weight) form gives cos 0.9207 / max-abs 62.7. Same pattern as the Gemma 4 entries.

Usage

No new flags. Streaming training on this target works once the entry exists.

Testing

pytest tests/unit/torch/speculative/plugins/test_modeling_final_norm.py -q

• 21 passed (CPU, PyTorch 2.13.0, transformers 5.15.1); the parametrized test_select_final_norm_type gains both keys.
• Norm form measured on meta-models/Muse-Glimmer-30B (bf16, B300): forward hook on the final MuseGlimmerRMSNorm input, both candidate forms compared with the module output and with hidden_states[-1].

Before your PR is "Ready for review"

• Is this change backward compatible?: ✅
• If you copied code from any other sources or added a new PIP dependency, did you follow guidance in CONTRIBUTING.md: N/A
• Did you write any new necessary tests?: ✅
• Did you update Changelog?: ✅
• Did you get Claude approval on this PR?: ❌ (not run yet)

Additional Information

Draft until a Muse Glimmer EAGLE3/DFlash training smoke with this entry has run end to end.

🤖 Generated with Claude Code

[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.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}