config.py

from functools import lru_cache
from typing import List

from pydantic_settings import BaseSettings


class Settings(BaseSettings):
    """Application settings optimized for CPU deployment."""
    
    # Application metadata
    app_name: str = "Vision"
    app_version: str = "1.0.0"
    
    # Model settings - Using official ONNX model with Q4 for speed
    model_id: str = "LiquidAI/LFM2.5-VL-1.6B-ONNX"
    encoder_variant: str = "q4"  
    decoder_variant: str = "q4"
    
    # Server settings (same as lfm-text for consistency)
    host: str = "0.0.0.0"
    port: int = 7860
    
    # CORS settings
    cors_origins: List[str] = ["*"]
    
    # Vision processing settings
    min_image_tokens: int = 64
    max_image_tokens: int = 256
    do_image_splitting: bool = True
    
    # Supported image formats
    supported_formats: List[str] = ["jpeg", "jpg", "png", "gif", "webp", "bmp"]
    max_image_size_mb: int = 10
    
    # Generation defaults (from LiquidAI recommendations)
    temperature: float = 0.1
    top_k: int = 50
    top_p: float = 0.5
    min_p: float = 0.15
    max_tokens: int = 2024
    repetition_penalty: float = 1.05
    
    # CPU optimization
    num_threads: int = 2  # Optimized for 2 vCPU
    
    # Logging
    log_level: str = "info"
    
    class Config:
        env_prefix = "LFMVL_"


@lru_cache()
def get_settings() -> Settings:
    """Get cached settings."""
    return Settings()


settings = get_settings()