[Bugfix] Honor request seed in MOSS-TTS adapters (#6543)

Signed-off-by: JiataiWang <wangjiatai@proton.me>

tests/entrypoints/openai_api/test_tts_adapter.py

@@ -10,11 +10,14 @@ from types import SimpleNamespace
 from typing import Any
 
 import pytest
+from vllm.sampling_params import SamplingParams
 
+from vllm_omni.entrypoints.openai.protocol.audio import OpenAICreateSpeechRequest
 from vllm_omni.entrypoints.openai.tts_adapters import (
     TTS_ADAPTER_REGISTRY,
     ARTTSAdapter,
     DiffusionTTSAdapter,
+    SpeechServingContext,
     all_tts_model_types,
     detect_tts_model_type,
     resolve_adapter,
@@ -115,6 +118,45 @@ def test_moss_tts_applies_request_max_new_tokens(adapter_cls):
     assert stage_defaults[0].max_tokens == 4096
 
 
+def _build_moss_tts_request(adapter_cls, mocker, *, request_seed):
+    server = mocker.Mock()
+    server._build_moss_tts_params = mocker.AsyncMock(return_value={})
+    adapter = adapter_cls(SpeechServingContext(server=server))
+    request = OpenAICreateSpeechRequest(input="hello", seed=request_seed)
+
+    return asyncio.run(
+        adapter.build(
+            request,
+            [SamplingParams(seed=42)],
+            has_inline_ref_audio=False,
+        )
+    )
+
+
+# Full-family coverage pins the adapter contract; only Nano consumes this seed end to end today.
+@pytest.mark.parametrize("adapter_cls", [MossTTSAdapter, MossTTSNanoAdapter])
+@pytest.mark.parametrize("request_seed", [0, 1234])
+def test_moss_tts_request_seed_overrides_stage_default(adapter_cls, request_seed, mocker):
+    prepared = _build_moss_tts_request(
+        adapter_cls,
+        mocker,
+        request_seed=request_seed,
+    )
+
+    assert prepared.tts_params["seed"] == [request_seed]
+
+
+@pytest.mark.parametrize("adapter_cls", [MossTTSAdapter, MossTTSNanoAdapter])
+def test_moss_tts_seed_falls_back_to_stage_default(adapter_cls, mocker):
+    prepared = _build_moss_tts_request(
+        adapter_cls,
+        mocker,
+        request_seed=None,
+    )
+
+    assert prepared.tts_params["seed"] == [42]
+
+
 def test_qwen3_tts_metadata():
     assert Qwen3TTSAdapter.backend == "ar"
     assert issubclass(Qwen3TTSAdapter, ARTTSAdapter)

vllm_omni/entrypoints/openai/tts_adapters/moss_tts.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """MOSS-TTS serving adapters (Nano + full family).
 
 Both variants share the same build/validate flow (``_build_moss_tts_params``
@@ -97,9 +98,13 @@ class _MossTTSAdapterBase(ARTTSAdapter):
             if voice_lower in server.uploaded_speakers and not has_inline_ref_audio:
                 tts_params["voice_name"] = [voice_lower]
                 tts_params["voice_created_at"] = [server._voice_created_at(voice_lower)]
-        # MOSS reads the resolved seed at build time (it samples internally).
-        if sampling_params_list and getattr(sampling_params_list[0], "seed", None) is not None:
-            tts_params["seed"] = [sampling_params_list[0].seed]
+        # MOSS samples internally from additional_information. build() runs
+        # before the shared path applies request.seed to SamplingParams.
+        seed = request.seed
+        if seed is None and sampling_params_list:
+            seed = getattr(sampling_params_list[0], "seed", None)
+        if seed is not None:
+            tts_params["seed"] = [int(seed)]
         if isinstance(tts_params.get("prompt_token_ids"), list):
             prompt_token_ids = tts_params.pop("prompt_token_ids")
             prompt = tokens_input(prompt_token_ids=prompt_token_ids)