Skip to content

fix: Normalize ASR MRoPE metadata for runtime - #146

Open
suharvest wants to merge 1 commit into
NVIDIA:mainfrom
suharvest:pr/fix-asr-mrope-normalization
Open

fix: Normalize ASR MRoPE metadata for runtime#146
suharvest wants to merge 1 commit into
NVIDIA:mainfrom
suharvest:pr/fix-asr-mrope-normalization

Conversation

@suharvest

Copy link
Copy Markdown

What does this PR do?

Extends the existing runtime RoPE normalization to recognize checkpoint
metadata that combines rope_type: linear with mrope_section.

That combination describes an MRoPE layout used by Qwen3-ASR checkpoints, but
the runtime accepts the existing normalized default + mrope_section form.
Without this normalization, export preserves linear and the runtime does not
select its MRoPE path.

The input dictionary is copied rather than mutated. Ordinary linear RoPE
metadata without mrope_section remains unchanged.

Scope

  • One production-line change in the existing normalization helper.
  • Two focused regression tests.
  • No model-name, local path, runtime, engine, or product-specific branches.

Validation

  • Focused regression tests: 2 passed.
  • Positive case verifies linear + mrope_section becomes
    default + mrope_section without mutating the source.
  • Negative case verifies ordinary linear RoPE keeps linear semantics.
  • Python syntax, focused formatting checks, git diff --check, and merge-tree
    validation pass.

This is a configuration-normalization regression test. A complete
export → build → inference validation requires the model and GPU environment
and is still requested from CI/device testing.

Closes #142.

Signed-off-by: suharvest <suharvest@gmail.com>
@suharvest
suharvest requested a review from a team July 25, 2026 12:00
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

[Bug][Export]: Qwen3-ASR export can preserve rope_type=linear and disable MRoPE at runtime

1 participant