[Maintenance] Refresh MOSS-TTS optimization on main

Signed-off-by: Sy03 <1370724210@qq.com>
This commit is contained in:
Sy03 committed 2026-09-09 12:46:57 +08:00
commit 28dd49a541
334 files changed
+24909 -6187

No files matched your search

+106 -28
View File
@@ -22,12 +22,14 @@ from vllm_omni.entrypoints.openai.tts_adapters import (
detect_tts_model_type,
resolve_adapter,
)
from vllm_omni.entrypoints.openai.tts_adapters.covo_audio import CovoAudioAdapter
from vllm_omni.entrypoints.openai.tts_adapters.higgs_audio_v2 import HiggsAudioV2Adapter
from vllm_omni.entrypoints.openai.tts_adapters.indextts2 import (
IndexTTS2Adapter,
IndexTTS25Adapter,
indextts2_conditioning_cache_salt,
)
from vllm_omni.entrypoints.openai.tts_adapters.ming_flash_omni_tts import MingFlashOmniTTSAdapter
from vllm_omni.entrypoints.openai.tts_adapters.moss_tts import (
MossTTSAdapter,
MossTTSNanoAdapter,
@@ -37,6 +39,8 @@ from vllm_omni.entrypoints.openai.tts_adapters.qwen3_tts import (
Qwen3TTSAdapter,
Qwen3TTSCodecLimitError,
)
from vllm_omni.entrypoints.openai.tts_adapters.step_audio2 import StepAudio2Adapter
from vllm_omni.entrypoints.openai.tts_adapters.voxtral import VoxtralTTSAdapter
from vllm_omni.model_executor.models.indextts2 import prompt_utils
from vllm_omni.model_executor.models.indextts2.tokenizer_v2_5 import (
INDEXTTS25_TOKENIZER_FILE,
@@ -125,8 +129,8 @@ def test_moss_tts_applies_request_max_new_tokens(adapter_cls):
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))
adapter._build_moss_tts_params = mocker.AsyncMock(return_value={})
request = OpenAICreateSpeechRequest(input="hello", seed=request_seed)
return asyncio.run(
@@ -162,46 +166,45 @@ def test_moss_tts_seed_falls_back_to_stage_default(adapter_cls, mocker):
assert prepared.tts_params["seed"] == [42]
@pytest.mark.parametrize("variant,ref_text", [("local", None), ("local", " "), ("tts", "Reference.")])
def test_moss_tts_generation_fallback(variant, ref_text, mocker):
server = mocker.Mock(_moss_variant=variant)
server._build_moss_tts_params = mocker.AsyncMock(return_value={"prompt_token_ids": [7]})
request = OpenAICreateSpeechRequest(input="Target.", ref_text=ref_text)
adapter = MossTTSAdapter(SpeechServingContext(server=server))
prepared = asyncio.run(adapter.build(request, [], has_inline_ref_audio=True))
server._build_moss_tts_params.assert_awaited_once_with(request, has_inline_ref_audio=True)
server._get_moss_processor.assert_not_called()
assert prepared.prompt["prompt_token_ids"] == [7]
def test_moss_local_reference_transcript_continuation(mocker):
@pytest.mark.parametrize(
"variant,ref_text,mode",
[
("local", " Reference. ", "continuation"),
("local", None, "generation"),
("local", " ", "generation"),
("tts", "Reference.", "generation"),
],
)
def test_moss_reference_transcript_mode(variant, ref_text, mode, mocker):
import torch
reference = [torch.ones((3, 12), dtype=torch.int64)]
unified = torch.arange(52, dtype=torch.int64).reshape(1, 4, 13)
processor = mocker.Mock(return_value={"input_ids": unified})
server = mocker.Mock(_moss_variant="local", uploaded_speakers={"speaker": {}})
server._build_moss_tts_params = mocker.AsyncMock(return_value={"prompt_token_ids": [7]})
server._get_moss_processor.return_value = processor
server._voice_created_at.return_value = 123
server._encode_moss_references = mocker.AsyncMock(return_value=(reference, {0: "reference-key"}))
request = OpenAICreateSpeechRequest(
input="Target.", ref_text=" Reference. ", language="English", voice="Speaker", seed=0, max_new_tokens=2048
input="Target.", ref_text=ref_text, language="English", voice="Speaker", seed=0, max_new_tokens=2048
)
adapter = MossTTSAdapter(SpeechServingContext(server=server))
adapter._moss_variant = variant
mocker.patch.object(adapter, "_get_moss_processor", return_value=processor)
mocker.patch.object(
adapter, "_encode_moss_references", new=mocker.AsyncMock(return_value=(reference, {0: "reference-key"}))
)
prepared = asyncio.run(adapter.build(request, [], has_inline_ref_audio=False))
processor.build_user_message.assert_called_once_with(text="Reference. Target.", language="English")
processor.build_assistant_message.assert_called_once_with(audio_codes_list=reference)
processor.assert_called_once_with(
conversations=[[processor.build_user_message.return_value, processor.build_assistant_message.return_value]],
mode="continuation",
)
server._encode_moss_references.assert_awaited_once_with(request, has_inline_ref_audio=False, two_speaker=False)
server._build_moss_tts_params.assert_not_called()
if mode == "continuation":
processor.build_user_message.assert_called_once_with(text="Reference. Target.", language="English")
processor.build_assistant_message.assert_called_once_with(audio_codes_list=reference)
conversation = [processor.build_user_message.return_value, processor.build_assistant_message.return_value]
else:
processor.build_user_message.assert_called_once_with(text="Target.", language="English", reference=reference)
processor.build_assistant_message.assert_not_called()
conversation = [processor.build_user_message.return_value]
processor.assert_called_once_with(conversations=[conversation], mode=mode)
assert prepared.prompt["prompt_token_ids"] == unified[0, :, 0].tolist()
assert torch.equal(prepared.tts_params["codes"]["ref"], unified[0, :, 1:])
assert prepared.tts_params["max_new_frames"] == [2048]
@@ -218,6 +221,81 @@ def test_qwen3_tts_metadata():
assert issubclass(Qwen3TTSAdapter, ARTTSAdapter)
def test_qwen3_tts_build_constructs_prepared_request():
server = SimpleNamespace(_tts_executor=None, _tts_tokenizer=None, uploaded_speakers={})
engine_client = SimpleNamespace(model_config=SimpleNamespace())
adapter = Qwen3TTSAdapter(SimpleNamespace(server=server, engine_client=engine_client))
async def estimate_prompt_len(_tts_params):
return 3
adapter._estimate_prompt_len_async = estimate_prompt_len
request = OpenAICreateSpeechRequest(input="hello")
prepared = asyncio.run(adapter.build(request, [], False))
assert prepared.prompt["prompt_token_ids"] == [1, 1, 1]
assert prepared.prompt["additional_information"] is prepared.tts_params
assert prepared.tts_params["text"] == ["hello"]
assert prepared.tts_params["speaker"] == ["Vivian"]
assert prepared.model_type == "CustomVoice"
def test_covo_audio_build_constructs_tokenized_prompt():
class FakeTokenizer:
def apply_chat_template(self, messages, *, tokenize, add_generation_prompt):
assert messages[-1] == {"role": "user", "content": "hello"}
assert tokenize is True
assert add_generation_prompt is True
return [11, 12, 13]
adapter = CovoAudioAdapter(SimpleNamespace(server=SimpleNamespace(), engine_client=SimpleNamespace()))
adapter._tokenizer = FakeTokenizer()
prepared = asyncio.run(adapter.build(OpenAICreateSpeechRequest(input="hello"), [], False))
assert prepared.prompt == {"prompt_token_ids": [11, 12, 13]}
assert prepared.tts_params == {}
assert prepared.model_type == "covo_audio"
def test_ming_flash_omni_build_constructs_prepared_request():
server = SimpleNamespace(uploaded_speakers={})
adapter = MingFlashOmniTTSAdapter(SimpleNamespace(server=server, engine_client=SimpleNamespace()))
request = OpenAICreateSpeechRequest(input="hello", voice="test")
prepared = asyncio.run(adapter.build(request, [], False))
assert prepared.prompt["prompt_token_ids"] == [0]
info = prepared.prompt["additional_information"]
assert info["ming_task"] == "instruct"
assert info["text"] == "hello"
assert info["voice_name"] == "test"
assert prepared.model_type == "ming_flash_omni_tts"
def test_step_audio2_build_constructs_open_assistant_turn():
adapter = StepAudio2Adapter(SimpleNamespace(server=SimpleNamespace(), engine_client=SimpleNamespace()))
request = OpenAICreateSpeechRequest(input="hello", instructions="Read warmly.")
prepared = asyncio.run(adapter.build(request, [], False))
assert "<|im_start|>system\nRead warmly.<|im_end|>" in prepared.prompt["prompt"]
assert prepared.prompt["prompt"].endswith("<|im_start|>assistant\n<tts_start>")
assert prepared.tts_params == {}
assert prepared.model_type == "step_audio2"
def test_voxtral_prompt_builder_is_sync():
ctx = SimpleNamespace(
server=SimpleNamespace(_tts_executor=None),
engine_client=SimpleNamespace(),
)
adapter = VoxtralTTSAdapter(ctx)
assert not asyncio.iscoroutinefunction(adapter._build_prompt)
@pytest.mark.parametrize(
("task_type", "text_tokens", "request_cap", "expected_cap"),
[