[Fix] Qwen3-ASR config: set thinker_config before super().__init__

[AgainstEntropy]

Motivation

Launching any Qwen3-ASR model crashes during config init under huggingface_hub >= 1.12:

File "python/sglang/srt/configs/qwen3_asr.py", line 151, in __init__
    super().__init__(**kwargs)
File "huggingface_hub/dataclasses.py", line 276, in init_with_validate
    cls.validate(self)
File "transformers/configuration_utils.py", line 446, in validate_token_ids
    text_config = self.get_text_config(decoder=True)
File "python/sglang/srt/configs/qwen3_asr.py", line 164, in get_text_config
    return self.thinker_config.text_config
AttributeError: 'Qwen3ASRConfig' object has no attribute 'thinker_config'

PretrainedConfig.__init__ is wrapped by huggingface_hub.dataclasses.init_with_validate, which runs validate_token_ids → get_text_config(decoder=True) during base init. Our get_text_config reads self.thinker_config.text_config, but self.thinker_config is only set after super().__init__ returns. Result: every Qwen3-ASR launch hard-fails before model load.

This was introduced in #22073 (Qwen3-ASR support) and started crashing once huggingface_hub 1.12+ began running validation inside the base __init__. The current sglang pyproject.toml already requires huggingface_hub >= 1.0, so users hit this on any reasonably recent install.

Registered unittest for this model is covered here: #22731

Modifications

python/sglang/srt/configs/qwen3_asr.py — set self.thinker_config before calling super().__init__(**kwargs), so the base validator finds the attribute it needs. One-line move; no behavior change for callers that already construct successfully on older huggingface_hub.

Accuracy Tests

Speed Tests and Profiling

Test plan

Verified locally on H200 with transformers 5.5.4 + huggingface_hub 1.12.2:

• Without this fix: python -m sglang.launch_server --model-path Qwen/Qwen3-ASR-0.6B and the existing test/manual/models/test_qwen3_asr.py suite both crash at server startup with the AttributeError above (every test ERRORs in setUpClass).
• With this fix: server launches cleanly; all 19 tests in test/manual/models/test_qwen3_asr.py run end-to-end (HTTP + WebSocket paths), passing or failing on their own merits.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[AgainstEntropy]

/tag-and-rerun-ci

[AgainstEntropy]

/rerun-failed-ci

[vf827393]

Same issue encountered with v0.5.11