[aoshen02]

Summary

• register the vendored Qwen3-ASR Auto classes only when Transformers does not already provide the qwen3_asr model type
• preserve Transformers' native Qwen3-ASR registrations on newer releases
• add unit coverage for both the native-present and native-absent paths

Motivation

megatron.bridge imports all model registrations eagerly. Transformers 5.13.1 now ships a native qwen3_asr config, so the unconditional vendored registration raises:

ValueError: 'qwen3_asr' is already used by a Transformers config, pick another name.

That prevents every Bridge consumer from importing, even when Qwen3-ASR is not selected. This is visible with the vLLM 0.25.1 base image, which ships Transformers 5.13.1.

Using exist_ok=True would replace the native global registration with Bridge's vendored schema. This change instead leaves the native registry untouched and retains the existing vendored fallback for Transformers releases without native Qwen3-ASR support.

This follows the import-safety goal in #4314: an unused model integration should not make import megatron.bridge fail.

Validation

• ruff format --check and ruff check on the changed files
• focused registration unit tests: 2 passed
• Transformers 5.13.1 probe: import megatron.bridge succeeds and the native Qwen3-ASR config remains registered
• Transformers 5.8.1 probe: import megatron.bridge succeeds and the vendored Qwen3-ASR config is registered

[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.

[ISEEKYAN]

/ok to test 6019e59