[Cleanup] Use tokens_input() for TTS prompt construction (#3227)

Signed-off-by: Nick Cao <ncao@redhat.com>
Co-authored-by: Claude <noreply@anthropic.com>
This commit is contained in:
Nick CaoandClaude authored and GitHub committed 2026-05-04 10:17:25 -04:00
1 parent 9a53703676
commit 5fc0bfe01e
3 files changed
+23 -19

No files matched your search

@@ -109,6 +109,7 @@ class TestVoxCPMServing:
assert call.kwargs["prompt"] == { assert call.kwargs["prompt"] == {
"prompt_token_ids": [1], "prompt_token_ids": [1],
"additional_information": tts_params, "additional_information": tts_params,
"type": "token",
} }
assert call.kwargs["output_modalities"] == ["audio"] assert call.kwargs["output_modalities"] == ["audio"]
@@ -140,4 +141,5 @@ class TestVoxCPMServing:
assert call.kwargs["prompt"] == { assert call.kwargs["prompt"] == {
"prompt_token_ids": [1], "prompt_token_ids": [1],
"additional_information": tts_params, "additional_information": tts_params,
"type": "token",
} }
+17 -18
View File
@@ -24,6 +24,7 @@ from vllm.entrypoints.openai.engine.protocol import (
RequestResponseMetadata, RequestResponseMetadata,
) )
from vllm.entrypoints.openai.engine.serving import OpenAIServing from vllm.entrypoints.openai.engine.serving import OpenAIServing
from vllm.inputs import tokens_input
from vllm.logger import init_logger from vllm.logger import init_logger
from vllm.multimodal.media import MediaConnector from vllm.multimodal.media import MediaConnector
from vllm.utils import random_uuid from vllm.utils import random_uuid
@@ -1537,10 +1538,9 @@ class OmniOpenAIServingSpeech(OpenAIServing, AudioMixin):
self._tts_tokenizer = mistral_tokenizer.instruct self._tts_tokenizer = mistral_tokenizer.instruct
if voice is not None: if voice is not None:
tokens = self._tts_tokenizer.encode_speech_request(SpeechRequest(input=text, voice=voice)).tokens tokens = self._tts_tokenizer.encode_speech_request(SpeechRequest(input=text, voice=voice)).tokens
return { prompt = tokens_input(prompt_token_ids=tokens)
"prompt_token_ids": tokens, prompt["additional_information"] = {"voice": [voice]}
"additional_information": {"voice": [voice]}, return prompt
}
else: else:
tokenized = self._tts_tokenizer.encode_speech_request(SpeechRequest(input=text, ref_audio=ref_audio)) tokenized = self._tts_tokenizer.encode_speech_request(SpeechRequest(input=text, ref_audio=ref_audio))
audio = tokenized.audios[0] audio = tokenized.audios[0]
@@ -1586,10 +1586,9 @@ class OmniOpenAIServingSpeech(OpenAIServing, AudioMixin):
} }
if request.max_new_tokens is not None: if request.max_new_tokens is not None:
additional_information["max_new_tokens"] = [request.max_new_tokens] additional_information["max_new_tokens"] = [request.max_new_tokens]
return { prompt = tokens_input(prompt_token_ids=prompt_ids)
"prompt_token_ids": prompt_ids, prompt["additional_information"] = additional_information
"additional_information": additional_information, return prompt
}
wav_samples, sr = ref_audio_data wav_samples, sr = ref_audio_data
normalized_text, normalized_ref_text = normalize_fish_voice_clone_texts(request.input, request.ref_text) normalized_text, normalized_ref_text = normalize_fish_voice_clone_texts(request.input, request.ref_text)
@@ -1612,10 +1611,9 @@ class OmniOpenAIServingSpeech(OpenAIServing, AudioMixin):
additional_information["voice_created_at"] = self.uploaded_speakers[voice_lower].get("created_at", 0) additional_information["voice_created_at"] = self.uploaded_speakers[voice_lower].get("created_at", 0)
if request.max_new_tokens is not None: if request.max_new_tokens is not None:
additional_information["max_new_tokens"] = request.max_new_tokens additional_information["max_new_tokens"] = request.max_new_tokens
return { prompt = tokens_input(prompt_token_ids=[1] * ph_len)
"prompt_token_ids": [1] * ph_len, prompt["additional_information"] = additional_information
"additional_information": additional_information, return prompt
}
# ---- CosyVoice3 helpers ---- # ---- CosyVoice3 helpers ----
@@ -1685,10 +1683,9 @@ class OmniOpenAIServingSpeech(OpenAIServing, AudioMixin):
if has_spk_emb: if has_spk_emb:
# Passed as plain float list # Passed as plain float list
additional_information["spk_emb"] = list(request.speaker_embedding) additional_information["spk_emb"] = list(request.speaker_embedding)
return { prompt = tokens_input(prompt_token_ids=[0])
"prompt_token_ids": [0], prompt["additional_information"] = additional_information
"additional_information": additional_information, return prompt
}
# ---- Common speech generation helpers ---- # ---- Common speech generation helpers ----
@@ -1764,7 +1761,8 @@ class OmniOpenAIServingSpeech(OpenAIServing, AudioMixin):
tts_params = {} tts_params = {}
elif self._tts_model_type == "moss_tts_nano": elif self._tts_model_type == "moss_tts_nano":
tts_params = await self._build_moss_tts_params(request) tts_params = await self._build_moss_tts_params(request)
prompt = {"prompt_token_ids": [1], "additional_information": tts_params} prompt = tokens_input(prompt_token_ids=[1])
prompt["additional_information"] = tts_params
else: else:
tts_params = self._build_tts_params(request) tts_params = self._build_tts_params(request)
# Resolve ref_audio (explicit or auto-set for uploaded voices) # Resolve ref_audio (explicit or auto-set for uploaded voices)
@@ -1778,7 +1776,8 @@ class OmniOpenAIServingSpeech(OpenAIServing, AudioMixin):
tts_params["ref_audio"] = [[wav_list, sr]] tts_params["ref_audio"] = [[wav_list, sr]]
ph_len = await self._estimate_prompt_len_async(tts_params) ph_len = await self._estimate_prompt_len_async(tts_params)
prompt = {"prompt_token_ids": [1] * ph_len, "additional_information": tts_params} prompt = tokens_input(prompt_token_ids=[1] * ph_len)
prompt["additional_information"] = tts_params
else: else:
# Qwen omni models (Qwen3-Omni, Qwen2.5-Omni) use a "talker" # Qwen omni models (Qwen3-Omni, Qwen2.5-Omni) use a "talker"
# stage whose preprocess requires chat-templated tokens. The # stage whose preprocess requires chat-templated tokens. The
@@ -23,6 +23,7 @@ import torch
import torch.nn as nn import torch.nn as nn
from vllm.config import VllmConfig from vllm.config import VllmConfig
from vllm.forward_context import get_forward_context, override_forward_context from vllm.forward_context import get_forward_context, override_forward_context
from vllm.inputs import tokens_input
from vllm.logger import init_logger from vllm.logger import init_logger
from vllm.model_executor.models.utils import ( from vllm.model_executor.models.utils import (
AutoWeightsLoader, AutoWeightsLoader,
@@ -121,7 +122,9 @@ def build_voxcpm2_prompt(
else: else:
additional["reference_audio"] = [[ref_audio, ref_sr]] additional["reference_audio"] = [[ref_audio, ref_sr]]
prefill_len += ref_len + 2 # ref_start / ref_end prefill_len += ref_len + 2 # ref_start / ref_end
return {"prompt_token_ids": [1] * prefill_len, "additional_information": additional} prompt = tokens_input(prompt_token_ids=[1] * prefill_len)
prompt["additional_information"] = additional
return prompt
def _encode_raw_audio( def _encode_raw_audio(