.env.example

# =============================================================================
# Uzbek TTS API Configuration
# Copy this file to .env and adjust the values as needed
# =============================================================================

# -----------------------------------------------------------------------------
# Application Settings
# -----------------------------------------------------------------------------
APP_NAME="Uzbek TTS API"
APP_VERSION="1.0.0"
DEBUG=false

# -----------------------------------------------------------------------------
# Server Configuration
# -----------------------------------------------------------------------------
HOST=0.0.0.0
PORT=8000
WORKERS=1

# -----------------------------------------------------------------------------
# Model Paths
# IMPORTANT: Download the model first using: python scripts/download_model.py
# -----------------------------------------------------------------------------
MODEL_PATH=./models/UZ.safetensors
VOCAB_PATH=./config/uz_vocab.txt
CONFIG_PATH=./config/UZTTS_conf.yaml

# -----------------------------------------------------------------------------
# GPU Configuration (NVIDIA A40 Optimized)
# -----------------------------------------------------------------------------
# Device: auto, cuda, cpu (auto will detect GPU automatically)
DEVICE=auto

# Number of parallel workers (6-8 optimal for A40 with 48GB VRAM)
NUM_WORKERS=6

# Enable TF32 for Ampere GPUs (A40, A100, RTX 30xx/40xx)
# Provides 20-30% speedup with minimal quality loss
ENABLE_TF32=true

# Enable torch.compile() for additional 15-25% speedup
# First request will be slower due to compilation
ENABLE_COMPILE=true

# CUDA device ID (0 = first GPU, 1 = second GPU, etc.)
CUDA_VISIBLE_DEVICES=0

# -----------------------------------------------------------------------------
# Inference Configuration
# Balance between speed and quality
# -----------------------------------------------------------------------------
# NFE Steps: Number of inference steps
# - 32: Original (best quality, slower)
# - 18-20: Optimized (good quality, 2x faster) ✅ RECOMMENDED for A40
# - 12-16: Fast (acceptable quality, 3x faster)
NFE_STEPS=18

# CFG Strength: Classifier-free guidance strength
# - 2.0: Original (best quality, slower)
# - 1.5: Optimized (good quality, faster) ✅ RECOMMENDED for A40
# - 1.0: Fast (lower quality)
CFG_STRENGTH=1.5

# Target RMS for audio normalization (0.0-1.0)
TARGET_RMS=0.1

# Default speech speed (0.5-2.0)
DEFAULT_SPEED=1.0

# -----------------------------------------------------------------------------
# Cache Configuration
# -----------------------------------------------------------------------------
# LRU cache size for reference audio embeddings
# Larger cache = more memory, but faster for repeated references
# A40 with 48GB can handle 1000-2000 easily
CACHE_SIZE=1000

# Cache directory
CACHE_DIR=./cache

# Enable caching
ENABLE_CACHE=true

# -----------------------------------------------------------------------------
# Request Queue Configuration
# -----------------------------------------------------------------------------
# Maximum number of requests in queue
MAX_QUEUE_SIZE=50

# Queue timeout in seconds
QUEUE_TIMEOUT=300

# Redis URL (optional, for distributed queue)
# REDIS_URL=redis://localhost:6379/0
USE_REDIS=false

# -----------------------------------------------------------------------------
# API Limits
# -----------------------------------------------------------------------------
# Maximum text length in characters
MAX_TEXT_LENGTH=5000

# Maximum audio duration in seconds
MAX_AUDIO_DURATION=300

# Maximum batch size for batch processing
MAX_BATCH_SIZE=8

# -----------------------------------------------------------------------------
# CORS Configuration
# -----------------------------------------------------------------------------
# Allowed origins (comma-separated or * for all)
# Example: http://localhost:3000,https://yourdomain.com
CORS_ORIGINS=["*"]

# -----------------------------------------------------------------------------
# Monitoring
# -----------------------------------------------------------------------------
# Enable Prometheus metrics
ENABLE_METRICS=true

# Metrics port
METRICS_PORT=9090

# -----------------------------------------------------------------------------
# Logging
# -----------------------------------------------------------------------------
# Log level: DEBUG, INFO, WARNING, ERROR, CRITICAL
LOG_LEVEL=INFO

# Log directory
LOG_DIR=./logs

# -----------------------------------------------------------------------------
# File Upload
# -----------------------------------------------------------------------------
# Upload directory for reference audio
UPLOAD_DIR=./uploads

# Maximum upload size in bytes (10MB default)
MAX_UPLOAD_SIZE=10485760

# =============================================================================
# Performance Tuning Notes for NVIDIA A40
# =============================================================================
#
# A40 Specifications:
# - 48GB VRAM
# - 10,752 CUDA cores
# - Ampere architecture
# - 300W TDP
#
# Recommended Settings for Different Use Cases:
#
# 1. MAXIMUM QUALITY (slower)
#    NFE_STEPS=32
#    CFG_STRENGTH=2.0
#    NUM_WORKERS=4
#    Expected latency: 2-3s per 10s audio
#
# 2. BALANCED (recommended) ✅
#    NFE_STEPS=18
#    CFG_STRENGTH=1.5
#    NUM_WORKERS=6
#    Expected latency: 1-1.5s per 10s audio
#
# 3. MAXIMUM SPEED
#    NFE_STEPS=12
#    CFG_STRENGTH=1.0
#    NUM_WORKERS=8
#    Expected latency: 0.5-1s per 10s audio
#
# GPU Memory Usage:
# - Per worker: ~1.5-2GB VRAM
# - 6 workers: ~10-14GB VRAM (30% of A40)
# - Reserve: ~34GB for buffers and cache
#
# =============================================================================