[Bugfix][Qwen3-Omni] Default the talker voice to Ethan (#8568)

Signed-off-by: Yueqian Lin <linyueqian@outlook.com>

tests/model_executor/models/qwen3_omni/test_qwen3_omni_default_speaker.py

@@ -0,0 +1,42 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+"""Regression tests for the Qwen3-Omni default talker voice.
+
+The released checkpoint lists ``talker_config.speaker_id`` as
+``{"chelsie": 2301, "ethan": 2302, "aiden": 2303}``. The default voice used to
+be the first key, so a request that named no speaker was spoken by Chelsie,
+while the reference ``generate(speaker="Ethan")`` (and other servers) use Ethan.
+Benchmarks that omit ``speaker`` then compared two different voices.
+"""
+
+import pytest
+
+from vllm_omni.model_executor.models.qwen3_omni.qwen3_omni import (
+    DEFAULT_SPEAKER,
+    Qwen3OmniMoeForConditionalGeneration,
+    resolve_default_speaker,
+)
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
+
+# Order as shipped in Qwen/Qwen3-Omni-30B-A3B-Instruct config.json.
+CHECKPOINT_SPEAKER_IDS = {"chelsie": 2301, "ethan": 2302, "aiden": 2303}
+
+
+def test_default_speaker_is_ethan_even_when_listed_second() -> None:
+    assert DEFAULT_SPEAKER == "ethan"
+    assert resolve_default_speaker(CHECKPOINT_SPEAKER_IDS) == "ethan"
+
+
+def test_default_speaker_falls_back_to_first_key_without_ethan() -> None:
+    assert resolve_default_speaker({"chelsie": 1, "aiden": 2}) == "chelsie"
+
+
+def test_unknown_or_missing_speaker_uses_ethan_token() -> None:
+    model = object.__new__(Qwen3OmniMoeForConditionalGeneration)
+    model.tts_text_spk_token_ids = dict(CHECKPOINT_SPEAKER_IDS)
+    model.default_tts_text_spk_type = resolve_default_speaker(model.tts_text_spk_token_ids)
+
+    assert model._get_text_spk_token_id("not-a-speaker") == CHECKPOINT_SPEAKER_IDS["ethan"]
+    assert model._get_text_spk_token_id(model.default_tts_text_spk_type) == CHECKPOINT_SPEAKER_IDS["ethan"]
+    assert model._get_text_spk_token_id("chelsie") == CHECKPOINT_SPEAKER_IDS["chelsie"]

vllm_omni/model_executor/models/qwen3_omni/qwen3_omni.py

@@ -67,6 +67,20 @@ from vllm_omni.model_executor.stage_input_processors.chunk_size_utils import par
 from vllm_omni.platforms import current_omni_platform
 from vllm_omni.utils.device_copy import to_device_nonblocking
 
+# Voice used when a request names no speaker. Matches the reference
+# ``Qwen3OmniMoeForConditionalGeneration.generate(speaker="Ethan")``; the
+# checkpoint's ``talker_config.speaker_id`` lists "chelsie" first, so taking the
+# first key would silently switch the default to a different (female) voice.
+DEFAULT_SPEAKER = "ethan"
+
+
+def resolve_default_speaker(speaker_token_ids: dict[str, int]) -> str:
+    """Return the speaker key used for requests that do not name one."""
+    if DEFAULT_SPEAKER in speaker_token_ids:
+        return DEFAULT_SPEAKER
+    return next(iter(speaker_token_ids))
+
+
 # Special token IDs for Qwen3 Omni MoE
 # Reference: https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct/blob/main/tokenizer_config.json
 
@@ -552,7 +566,7 @@ class Qwen3OmniMoeForConditionalGeneration(
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
         generate_audio: bool = True,
-        voice_type: str = "ethan",
+        voice_type: str = DEFAULT_SPEAKER,
         codec: torch.Tensor | None = None,
         sampling_metadata: SamplingMetadata | None = None,
         logits_index: int | None = None,
@@ -963,7 +977,7 @@ class Qwen3OmniMoeForConditionalGeneration(
                 "prefix_caching": talker_hf_config.audio_start_token_id,
             }
 
-        self.default_tts_text_spk_type = list(self.tts_text_spk_token_ids.keys())[0]
+        self.default_tts_text_spk_type = resolve_default_speaker(self.tts_text_spk_token_ids)
 
         return set(["thinker_embedding.weight", "talker_embedding.weight"])