o Mi&@sdZddlmZddlZddlZejejejeZ ej j dd dd dd d d d d ddddddddddddddddddd d!d"d#jeje d$d%d%d&d'd(jeje d$d%d)d*d'd(jeje d$d%d+d,d'd(jeje d$d-d.d'd(d/Zz ejjd0d'd1ZWn eydZYnwejd2ed3Zejd4erd5einiejd6gdd7d8d9d:d'd; ejdGd?d@d@Zejd4erd5einiejd6gdd7d8d9d:dAedBdCZ edDdEZ!ejd4erd5einid8dFdGdHZ"e#dIdJZ$dS)Ka Modal entrypoint for Veena3 TTS service. Deploys a GPU-backed ASGI app with: - True streaming TTS - Autoscaling (0 to N containers) - Per-container concurrency (vLLM batching) - Memory snapshots for fast cold starts Usage: # Deploy to Modal modal deploy veena3modal/app.py # Serve locally (for testing) modal serve veena3modal/app.py # Run a single function modal run veena3modal/app.py::tts_api ) annotationsNz3.10)python_versionffmpeggitz torch>=2.0.0ztorchaudio>=2.0.0ztransformers>=4.35.0z vllm>=0.4.0zfastapi[standard]>=0.100.0zuvicorn[standard]>=0.23.0z numpy>=1.24.0z scipy>=1.10.0zlibrosa>=0.10.0zsoundfile>=0.12.0zpydantic>=2.0.0zprometheus_client>=0.17.0zsupabase>=2.0.0z einops>=0.6.0z einx>=0.2.0zomegaconf>=2.3.0zsafetensors>=0.5.0z soxr>=0.5.0zmatplotlib>=3.8.0znatsort>=8.0.0z joblib>=1.0.0z0pip install --no-cache-dir nvidia-ml-py3 || truetruez3/root/external/sparktts:/root/external/AP-BWE:/rootz2025-12-25-sr-chunking-support)AUTH_BYPASS_MODE PYTHONPATH_IMAGE_BUILD_VERSIONz0mkdir -p /root/external/sparktts/sparktts/modelsz/mkdir -p /root/external/sparktts/sparktts/utilsexternalsparktts /root/external/sparktts/sparkttsT) local_path remote_pathcopymodels'/root/external/sparktts/sparktts/modelsutilsz&/root/external/sparktts/sparktts/utilszAP-BWEz/root/external/AP-BWE veena3modalz veena3-models)create_if_missing veena3-tts)nameimageL40S/modelszveena3-secretsi,iXi) gpuvolumessecretsmin_containersbuffer_containersscaledown_windowtimeoutstartup_timeoutenable_memory_snapshot) max_inputs target_inputsc@s0eZdZdZeddZeddZdS) TTSServicez TTS service class with model lifecycle management. Uses @modal.enter for model loading (once per container). c Csddl}ddl}|j|jd|t}|d|jdd}|jdd}|j |s<| d || d dSz)dd l m }|||j |rM|ndd d |j |d|_|d|jjWdSty}z|d|ddl}|WYd}~dSd}~ww)zm Load TTS model when container starts. Called once per container lifecycle. rN)levelu"🚀 Starting TTS model loading... MODEL_PATHz/models/spark_tts_4speakerAP_BWE_CHECKPOINT_DIRz/models/ap_bwe/16kto48kzModel path not found: z,TTS will return 503 until model is available)initialize_runtimecudag333333?) model_pathsr_checkpoint_dirdevicegpu_memory_utilization enable_sru✅ TTS model loaded: u❌ Model loading failed: )oslogging basicConfigINFO getLogger__name__infoenvirongetpathexistswarning veena3modal.services.tts_runtimer,runtime model_version Exceptionerror traceback print_exc) selfr3r4loggerr.sr_pathr,erDrJQ/home/ubuntu/.cursor/worktrees/spark__SSH__216.81.248.184_/byy/veena3modal/app.py load_modelsB      zTTSService.load_modelcCsddlm}|S)z? Return FastAPI ASGI app for serving requests. r create_appveena3modal.api.fastapi_apprN)rFrNrJrJrKserves zTTSService.serveN) r8 __module__ __qualname____doc__modalenterrLasgi_apprQrJrJrJrKr(s  1r()rrrrrr r!r"cCsddlm}|S)z Standalone ASGI app function. Note: Model will be loaded on first request (higher TTFB). For production, use TTSService class with @modal.enter. rrMrOrMrJrJrKtts_apis rXcCs ddddS)zF Simple health check that can be called from Modal dashboard. okrz0.1.0)statusserviceversionrJrJrJrJrK health_checksr])rrr!c Csddl}ddl}|jdd|jddiid}gd}|D];}|j|rTz||}|dd|d |<WqtyS}zd ||d |<WYd}~qd}~wwd |d |<qgd }|D])}z t|d |d|<Wqaty}zd||d|<WYd}~qad}~ww|S)z{ Debug function to test import paths and file structure. Run with: modal run veena3modal/app.py::debug_imports rNrNOT_SET ) pythonpathsys_pathfiles import_tests)z/rootz/root/externalz/root/external/sparkttsr rz/root/veena3modalrrbzERROR: DOES_NOT_EXIST)r zsparktts.modelszsparktts.models.audio_tokenizerrzveena3modal.corezveena3modal.processingzveena3modal.audioOKrczFAILED: ) r3sysr:r;r<r=listdirrB __import__) r3rgresults dirs_to_checkdrbrIrcmodulerJrJrK debug_imports s8      rncCsFddl}tdtdtdt}tdt|j|dddS) zT Local entrypoint for testing. Usage: modal run veena3modal/app.py rNzVeena3 TTS Modal Appz(========================================u/ 🔍 Running debug_imports on GPU container...u 📊 Debug Results:)indent)jsonprintrnremotedumps)rqresultrJrJrKmainIsrv)%rT __future__rr3rUr<dirnameabspath__file__ REPO_ROOTImage debian_slim apt_install pip_install run_commandsenv add_local_dirjoinadd_local_python_sourcerVolume from_name model_volumerBAppappclsSecret concurrentr(functionrWrXr]rnlocal_entrypointrvrJrJrJrKs  $@EKQX    E     ;