o c iG=@sdZddlZddlZddlZddlZddlZddlZddl Z ddl m Z de de de fddZ d e d ejd efd d ZddZddZddZddZddZddZddZedkr`edSdS)zCLI for FasterQwen3TTS.N)FasterQwen3TTSmodel_iddevicedtypecCs:|dkrtj}n |dkrtj}ntj}tj|||dddS)Nbf16fp16sdpa)rrattn_implementation max_seq_len)torchbfloat16float16float32rfrom_pretrained)rrr torch_dtyperJ/home/ubuntu/vllm_env/lib/python3.10/site-packages/faster_qwen3_tts/cli.py _load_modelsrout_pathaudiosrcCs,tjtj|p dddt|||dS)N.T)exist_ok)osmakedirspathdirnamesfwrite)rrrrrr _write_audiosr cCsHg}d}|D] \}}}||q|stjdtjddfSt||fS)N)ri.)appendnpzerosr concatenate)genchunksr audio_chunk_rrr_stream_to_audio$s r*c CsRt|j|j|j}|jrOt}|j|j|j |j |j |j |j |j|j|j |j|j|jd }t|\}}t|}|rBt||nd}|dkrL||nd}n@t}|j|j|j |j |j |j |j|j|j |j|j|jd \} }| d}t|}|rt||nd}|dkr||nd}t|j||td|jd|dd|dd dS) N textlanguage ref_audioref_text chunk_sizemax_new_tokens temperaturetop_k do_samplerepetition_penalty xvec_onlynon_streaming_moder r,r-r.r/r1r2r3r4r5r6r7Wrote  (dur .2fs, RTF ))rmodelrr streamingtime perf_countergenerate_voice_clone_streamingr,r-r.r/r0r1r2r3greedyr5r6r7r*lengenerate_voice_cloner outputprint argsr?startr&rr total_time audio_durrtf audio_listrrr cmd_clone.sR    &rPc Cst|j|j|j}|jr|jpg}td|dS|js(tdt d|j rjt }|j|j|j|j|j|j|j|j|j|j |jd }t|\}}t |}|r]t||nd}|dkrg||nd} n)rr?rr list_speakersget_supported_speakersrHjoinrTsysexitr@rArBgenerate_custom_voice_streamingr,r-rUr0r1r2r3rDr5r*rEgenerate_custom_voicer rG) rJr?speakersrKr&rrrLrMrNrOrrr cmd_custom]sX    &r_c Cs:t|j|j|j}|jrIt}|j|j|j |j |j |j |j |j|j |jd }t|\}}t|}|r)rr?rrr@rArBgenerate_voice_design_streamingr,rUr-r0r1r2r3rDr5r*rEgenerate_voice_designr rGrHrIrrr cmd_designsF   &rdcCst|j|j|j}|jdkr|jr|jstdt d|jdkr.|j s.tdt d|jdkr?|j s?tdt dtdd }tj D])}| }|sRqH|d vr[dStj|jd |d d }|d 7}t}|jdkr|jr|j||j|j|j|j|j|j|j|j |jd|jd }t|\}} n|j ||j|j|j|j|j|j|j |jd|jd \} } | d}n|jdkr|jr|j!||j |j|j |j|j|j|j|j |jd }t|\}} n[|j"||j |j|j |j|j|j|j |jd \} } | d}n=|jr%|j#||j |j|j|j|j|j|j |jd }t|\}} n|j$||j |j|j|j|j|j |jd\} } | d}t%||| t|} | rUt&|| nd} | dkr`| | nd} td|d| dd| ddqHdS)Nclonez=ERROR: --ref-audio and --ref-text are required for clone moderRcustomz,ERROR: --speaker is required for custom modedesignz-ERROR: --instruct is required for design modezCServer started. Enter text per line. Type 'exit' or 'quit' to stop.r!)r[quitstopout_04dz.wavFr+r9rrSrVr`rar8r:r;r<r=r>)'rr?rrmoder.r/rHrZr[rTrUstdinstriplowerrrrY output_dirrArBr@rCr-r0r1r2r3rDr5r7r*rFr\r]rbrcr rE)rJr?idxliner,rrKr&rrrOrLrMrNrrr cmd_serves               "rscCsdtjddd}|jdddd|jdd gd d d |jd dd}dd}|jddd}|||jdddd|jdddd|jdddd|jtd|jdd d}|||jd!d"d|jd#d$d%d|jd&dd'd|jtd|jd(d)d}|||jd#dd*d|jtd|jd+d,d}|jd-dgd.d/|jd0dd1d|jd2d3d4d|jdd5d|jdd6d|jd!d7d|jd#d$d8d|jd9dd:d| }|jd;d|jd?d|jddB|jdCt dDdEdF|jdGt dHdI|jdJt dKdI|jdLt dMdI|jdNt dOdI|jdPddQd|jdRdSdTd|jt d|S)UNzfaster-qwen3-ttszFasterQwen3TTS CLI)prog descriptionz--devicecudazDevice (cuda or cpu)defaulthelpz--dtyper)rrfp32z Model dtype)rxchoicesrycommandT)destrequiredcSs|jdddd|jdddd|jd dd d|jd dd d|jd tdd|jdtdd|jdtdd|jdtdd|jdddd|jdddd|}|jddddd|jd dd!d"d|jdd#|jd$td%d&d'dS)(Nz--textTzText to synthesizer~ry --languageAuto%Language (Auto, English, French, ...)rwz--outputzOutput wav path--modelModel id or local path--max-new-tokensr typerx --temperature?--top-k2--repetition-penalty?--greedy store_trueDisable samplingactionry --streamingUse streaming generation--non-streaming-moder7+Prefill full text for non-streaming qualityr}rry--no-non-streaming-mode store_false?Disable full-text prefill (match upstream non-streaming layout)r7 --chunk-sizeStreaming chunk sizerrxry) add_argumentintfloatadd_mutually_exclusive_group set_defaults)sp nsm_grouprrr add_common:s2 z build_parser..add_commonrezVoice cloning (reference audio))ryz --ref-audiozReference audio pathrz --ref-textzReference transcriptz --xvec-onlyrzUse speaker embedding onlyr)fnrfzCustomVoice model (speaker IDs)z --speakerz Speaker IDz --instructzOptional instructionz--list-speakerszList available speaker IDsrgz%VoiceDesign model (instruction-based)zVoice/style instructionservez8Keep model hot and generate multiple requests from stdinz--mode)rerfrg)r~r{rrrrrzReference audio path (clone)zReference transcript (clone)zSpeaker ID (custom)zInstruction (custom/design)rrrr7rrrrrrrrrrrr rrrrrrrrrz --output-diroutputszDirectory for output wavs) argparseArgumentParserradd_subparsers add_parserrrPr_rdrrrrs)psubrrrrrr build_parser4sj     rcCst}|}||dS)N)r parse_argsr)parserrJrrrmainsr__main__)__doc__rrrZrAnumpyr# soundfilerr faster_qwen3_ttsrstrrndarrayrr r*rPr_rdrsrr__name__rrrrs*  /4)zW