[Bugfix][Qwen3-Omni] Restore the talker sampling seed (#7474)

Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
Signed-off-by: Yueqian Lin <70319226+linyueqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Yueqian Lin <70319226+linyueqian@users.noreply.github.com>

tests/config/test_config_factory.py

@@ -2857,6 +2857,28 @@ class TestPlatformOverrides:
             config = deploy.stages[0].compilation_config or {}
             assert "+rotary_embedding" not in config.get("custom_ops", [])
 
+    def test_qwen3_omni_talker_sampling_is_seeded(self):
+        """The only stochastic Qwen3-Omni stage must stay reproducible.
+
+        Dropping this seed (as #4986 did) leaves the talker sampling codec
+        tokens unseeded at temperature 0.9, which reopened the audio-vs-text
+        nightly failures in #6090. Pin it so a cleanup cannot remove it again
+        without failing here.
+        """
+        deploy_path = Path(get_deploy_config_path("qwen3_omni_moe.yaml"))
+        pipeline = resolve_pipeline_config(
+            "qwen3_omni_moe",
+            Q3_OMNI_ALL_STAGES_HF_CONFIG,
+        )
+        assert isinstance(pipeline, PipelineConfig)
+
+        for platform in ("cpu", "cuda", "musa", "npu", "rocm", "xpu"):
+            deploy = _apply_platform_overrides(load_deploy_config(deploy_path), platform=platform)
+            stages = merge_pipeline_deploy(pipeline, deploy)
+            talker_sampling = stages[1].yaml_extras["default_sampling_params"]
+            assert talker_sampling["temperature"] > 0.0, "talker is expected to sample, not decode greedily"
+            assert talker_sampling.get("seed") == 42, f"talker sampling lost its seed on {platform}"
+
     def test_minicpmo_4_5_cuda_caps_talker_kv_cache(self):
         pipeline = resolve_pipeline_config("minicpmo_4_5")
         assert isinstance(pipeline, PipelineConfig)

vllm_omni/deploy/qwen3_omni_moe.yaml

@@ -46,6 +46,16 @@ stages:
     input_connectors:
       from_stage_0: connector_of_shared_memory
     default_sampling_params:
+      # The talker is the only stochastic stage (the thinker and code2wav both
+      # decode greedily at temperature 0). Without a seed, every request draws
+      # fresh codec tokens, and a small fraction of draws derail into fluent
+      # nonsense part-way through an utterance while the thinker text stays
+      # correct -- the audio-vs-text nightly failures in #6090. This seed was
+      # present from the original deploy schema and was dropped as collateral
+      # in #4986; keep it so Qwen3-Omni speech is reproducible for a given
+      # request. It bounds sampling only: it does not make the talker immune
+      # to the underlying degeneration, which is tracked separately.
+      seed: 42
       temperature: 0.9
       top_k: 50
       max_tokens: 4096