o ۷io` @szUddlZddlZddlZddlZddlZddlZddlZddlZddlm Z ddl m Z ddl m Z ddlZddlZddlmZddlmZmZddlmZddlmZdd lmZdd lmZdd lm Z m!Z!dd l"m#Z#ee$Z%d Z&dZ'e(de(de(de(de(de(de(de(dgZ)dhZ*e+e,e-d<hdZ.e+e,e-d<dZ/dZ0dZ1GdddeeZ2dS) N)Any)urlparse)urlopen)Request)ResponseStreamingResponse) OpenAIServing) init_logger) random_uuid) AudioMixin) CreateAudioOpenAICreateSpeechRequest)OmniRequestOutputi z 127.0.0.0/8z 10.0.0.0/8z 172.16.0.0/12z192.168.0.0/16z169.254.0.0/16z::1/128zfc00::/7z fe80::/10 qwen3_tts_TTS_MODEL_STAGES> AutoFrenchGermanKoreanChineseEnglishItalianRussianSpanishJapanese Portuguese_TTS_LANGUAGESiics0eZdZfddZdedBfddZddZdefd d Zde e fd d Z d e dedBfddZ dee efdefddZdefddZdede dBfddZede deeeeffddZde fddZedeedBe dBffddZdedee effd d!Z d%ded"edBfd#d$ZZS)&OmniOpenAIServingSpeechcsrtj|i|||_|jdu|_||_||_t dt |jdt |jd|_ ||_dS)NzLoaded z supported speakers: )super__init___find_tts_stage _tts_stage_is_tts _compute_max_instructions_length_max_instructions_length_load_supported_speakerssupported_speakersloggerinfolensorted_tts_tokenizer_load_codec_frame_rate_codec_frame_rate)selfargskwargs __class__a/home/ubuntu/vllm_env/lib/python3.10/site-packages/vllm_omni/entrypoints/openai/serving_speech.pyr!@s    "z OmniOpenAIServingSpeech.__init__returnNc CsBz[|jjj}tj|dd}tj|rZt| }t |}Wdn1s(wY| d}| d}|rZ|rZ|dkrZt |t |}t d|dd |d |d |WSWntyu}zt d |WYd}~nd}~wwz |jjj} t| d d}|durt d|dt |WSWdStyYdSw)zPLoad codec frame rate from speech tokenizer config for prompt length estimation.speech_tokenizerz config.jsonNoutput_sample_rateencode_downsample_raterzLoaded codec frame rate: z.1fz Hz (output_sample_rate=z, encode_downsample_rate=)z>Failed to load codec frame rate from speech tokenizer config: codec_frame_rate_hzz'Using codec frame rate from hf_config: z Hz) engine_client model_configmodelospathjoinexistsopenjsonloadgetfloatr)r* Exceptionwarning hf_configgetattr) r0 model_pathst_config_pathf st_config output_sr downsamplerateerKr5r5r6r.RsH           z.OmniOpenAIServingSpeech._load_codec_frame_ratecCs@t|jdd}|dur dS|D]}t|ddtvr|SqdS)zHFind and return the TTS stage from the stage list, or None if not found. stage_listN model_stagerLr=r)r0rUstager5r5r6r"qsz'OmniOpenAIServingSpeech._find_tts_stagecCsFt|jdd}|dur |S|jdur!t|jdi}d|vr!|dStS)zCompute max instructions length with precedence: CLI > stage config > default. Called once during initialization; result is cached in self._max_instructions_length. tts_max_instructions_lengthNtts_argsmax_instructions_length)rLr=r#_TTS_MAX_INSTRUCTIONS_LENGTH)r0 cli_overriderZr5r5r6r%{s z8OmniOpenAIServingSpeech._compute_max_instructions_lengthc Csz-|jjjj}dD]}t||d}|r$t|tr$dd|DWSq t dWt St yI}zt d|WYd}~t Sd}~ww)zHLoad supported speakers (case-insensitive) from the model configuration.)spk_id speaker_idNcSsh|]}|qSr5)lower).0speakerr5r5r6 szCOmniOpenAIServingSpeech._load_supported_speakers..zBNo speakers found in talker_config (checked spk_id and speaker_id)z+Could not load speakers from model config: ) r=r>rK talker_configrL isinstancedictkeysr)rJrIset)r0rd attr_name speakers_dictrTr5r5r6r's   z0OmniOpenAIServingSpeech._load_supported_speakers ref_audiocCs.|jdurdSz|}t|tr,|r,t|dkr!t|dttfr!n |d}t|tr,|st|trsz>OmniOpenAIServingSpeech._estimate_prompt_len..codec_language_idspk_is_dialect)additional_informationrtokenize_promptrrestimate_ref_code_lenz=Failed to estimate TTS prompt length, using fallback 2048: %s):vllm_omni.model_executor.models.qwen3_tts.qwen3_tts_talkerr~r- transformersrr=r>r?from_pretrainedrKrdrG/estimate_prompt_len_from_additional_informationrLr|rIr)rJ) r0r}r~r model_namerKrdrrTr5rr6_estimate_prompt_lens4         z,OmniOpenAIServingSpeech._estimate_prompt_lencCs:t|jdd}|r|D]}t|dd}|tvrdSq dS)z4Check if the current model is a supported TTS model.rUNrVTFrW)r0rUrXrVr5r5r6 _is_tts_models z%OmniOpenAIServingSpeech._is_tts_modelrequestcCsh|jpd}|jdur|j|_|jr|jsdS|jdur2|jtvr2d|jddttS|dkrS|jdurS|j rS|j|j vrSd|jddt|j S|dkrl|j dur^d S|j d sl|j d sld S|dkr~|j durwd S|j dur~dS|dkr|jsdS|jrt|j|jkrd|jdS|jdur|jtkrdtS|jtkrdtSdS)z?Validate TTS request parameters. Returns error message or None.rNzInput text cannot be emptyzInvalid language 'z'. Supported: z, zInvalid speaker 'Basez0Base task requires 'ref_audio' for voice cloning)zhttp://zhttps://data:zBref_audio must be a URL (http/https) or base64 data URL (data:...)z&'ref_text' is only valid for Base taskz0'x_vector_only_mode' is only valid for Base task VoiceDesignz>VoiceDesign task requires 'instructions' to describe the voicezInstructions too long (max z characters)z max_new_tokens must be at least zmax_new_tokens cannot exceed )rvoicer`inputstriplanguagerrBr,r(rk startswithref_textx_vector_only_mode instructionsr+r&max_new_tokens_TTS_MAX_NEW_TOKENS_MIN_TTS_MAX_NEW_TOKENS_MAX)r0rrr5r5r6_validate_tts_requests<           z-OmniOpenAIServingSpeech._validate_tts_request ref_audio_strcsftdtddfdddttjtfffdd }t}|d|IdH\}}| |fS)z;Resolve ref_audio URL/base64 to (wav_samples, sample_rate).urlr7Ncsvt|j}|s tdt|dD]'}t|ddddd}t|t fddt Dr8tdqdS) Nz%ref_audio URL must include a hostnamer%rc3s|]}|vVqdSNr5)ranetaddrr5r6 -szROmniOpenAIServingSpeech._resolve_ref_audio.._check_ssrf..z+ref_audio URL resolves to blocked address: ) rhostname ValueErrorsocket getaddrinfostrsplit ipaddress ip_addressany_REF_AUDIO_BLOCKED_NETWORKS)rhostr*ip_strr5rr6 _check_ssrf&s  z?OmniOpenAIServingSpeech._resolve_ref_audio.._check_ssrfcsjdvr:ttd}|td}t|tkr%tdtdWdn1s/wYt|}n  drV}d|vrM| ddd}tt |}ntdt j|d d d \}}t|tjrv|jdkrvtj|d d }tj|tjdt|fS)N)httphttps)timeoutrzref_audio URL exceeds z bytesr,z4ref_audio must be an http(s) URL or data: base64 URIfloat32F)dtype always_2dro)axis)r)schemer_REF_AUDIO_TIMEOUT_Sread_REF_AUDIO_MAX_BYTESr+rioBytesIOrrbase64 b64decodesfrenpndarrayrtmeanasarrayrrq)respdatabufb64audioryrparsedrr5r6 _fetch_sync0s&    z?OmniOpenAIServingSpeech._resolve_ref_audio.._fetch_sync) rrrrrrrqasyncioget_running_looprun_in_executortolist)rrloopwav_npryr5rr6_resolve_ref_audio!s  z*OmniOpenAIServingSpeech._resolve_ref_audio request_idc Csxd}d}z|2z3dHW}||\}}|durq|d}|dur=t|tr.|r.|dn|} t| dr9| nt| }||} t| trQ| |d} t| }n| dur]| g} |d7}ng} | D]-} t| drr|  n| } | j dkr}| } t| |d d d d d }||jVqaq6WdStjytd|ty}ztd||WYd}~dSd}~ww)aGenerate PCM audio chunks for streaming response. Handles two audio output modes from the engine: - Cumulative mode (list): Engine returns growing list of chunks; we emit only the new tail on each iteration. - Per-step mode (tensor): Engine returns single tensor per iteration; we emit it directly. Args: generator: Async generator from the engine request_id: Request identifier for logging Yields: Raw PCM bytes for each audio chunk r]NryrorwrrHpcm?rF audio_tensor sample_rateresponse_formatspeed stream_format base64_encodez(Streaming request %s cancelled by clientz-Streaming speech generation failed for %s: %s)_extract_audio_outputrGrerprsrwrqr+rHdetachcpunumpyrtsqueezer create_audio audio_datarCancelledErrorr)r*rI exception)r0 generatorr prev_countsample_rate_valres audio_output audio_keysr_rawsr_val audio_val new_chunks chunk_tensorchunk_np audio_objrTr5r5r6_generate_pcm_chunksHsV       & z,OmniOpenAIServingSpeech._generate_pcm_chunkscCs\t|dd}|st|dd}|rt|ddnd}|sdSd|vr"dnd|vr(dnd}||fS)zReturn (audio_output dict, audio key) or (None, None). Returns the raw dict so callers can apply their own extraction strategy: streaming needs per-chunk delta slicing; non-streaming needs full concatenation. multimodal_outputNrequest_output)NNr model_outputs)rL)rmmrokeyr5r5r6rs  z-OmniOpenAIServingSpeech._extract_audio_outputcCsi}|jg|d<|jdur|jg|d<ndg|d<|jdur%|jg|d<ndg|d<|jdur6|jg|d<n |dddkrCd g|d<|jdurO|jg|d <nd g|d <|jdur_|jg|d <|jdurj|jg|d <|jdurv|jg|d<ndg|d<|dddkrdg|d<|S)zBuild TTS parameters from request. Processes each parameter if present, skips if not. Values are wrapped in lists as required by the model. textNrrrrrbrVivianinstructrrrrrTnon_streaming_mode)rrrrrrrr)r0rparamsr5r5r6_build_tts_paramss2                z)OmniOpenAIServingSpeech._build_tts_params raw_requestc s||IdH}|durtd||S|jjr|jjdt}z|jr_||}|r5| |WS| |}|j durQ| |j IdH\}}||gg|d<| |} dg| |d} ni}d|ji} td|t|jd krz|jdd d n|j|d d gd |jj} |jj| || dgd} |jrt|| |ddWSd} | 2z3dHW}|} q6| dur| dWS|| \}}|dur| dWS||}|dd}t|tr|r|dn|}t|dr|nt|}t|trd dl}|j|dd}t|dr| !"}|j#dkr|$}t%|||j&p#d|j'p(d|j(dd}|)|}t*|j+|j,dWSt-j.yK| dYSt/yb}z | |WYd}~Sd}~wt0y}zt1d || d!|WYd}~Sd}~ww)"a Create Speech API similar to OpenAI's API. See https://platform.openai.com/docs/api-reference/audio/createSpeech for the API specification. This API mimics the OpenAI Create Speech API. For Qwen3-TTS models, additional parameters are supported: - task_type: "CustomVoice", "VoiceDesign", or "Base" - language: Language code (e.g., "Chinese", "English", "Auto") - voice: Speaker name (e.g., "Vivian", "Ryan") for CustomVoice - instructions: Voice style/emotion instructions - ref_audio: Reference audio for voice cloning (Base task) - ref_text: Transcript of reference audio (Base task) - x_vector_only_mode: Use speaker embedding only (Base task) Streaming is supported via stream=True with response_format='pcm'. Each Code2Wav chunk is yielded as raw PCM bytes as soon as it is decoded. NzError with model %szspeech-rkr)prompt_token_idsrpromptz,TTS speech request %s: text=%r, task_type=%s2z...runknownrr)rrsampling_params_listoutput_modalitiesz audio/pcm) media_typez#No output generated from the model.z'TTS model did not produce audio output.ryrrorw)dimrHrxrFr)contentrzClient disconnectedzSpeech generation failed: %szSpeech generation failed: )2 _check_modelr)errorr=errored dead_errorr r$rcreate_error_responserrkrrrr*r+rGdefault_sampling_params_listgeneratestreamrrrrerprsrwrqtorchcatrHrrrrtrr rrrrrrrrrrrIr)r0rrerror_check_retrvalidation_errorr}wav_listryph_lenrrr final_outputrrrrrrrr"raudio_responserTr5r5r6 create_speechs        $           z%OmniOpenAIServingSpeech.create_speechr)__name__ __module__ __qualname__r!rHr.r"rqr%rhrr'objectr|rfrrboolrr r staticmethodrrrprrrrrr* __classcell__r5r5r3r6r?s,  # 5 &? 9r)3rrrrrErur@rtypingr urllib.parserurllib.requestrrr soundfilerfastapirfastapi.responsesrr&vllm.entrypoints.openai.engine.servingr vllm.loggerr vllm.utilsr .vllm_omni.entrypoints.openai.audio_utils_mixinr +vllm_omni.entrypoints.openai.protocol.audior r vllm_omni.outputsrr+r)rr ip_networkrrrhr__annotations__rr\rrrr5r5r5r6sN