config.py

"""
Configuration file for easy parameter tuning
This file contains all adjustable parameters for the AI backend
"""

# ====== MODEL CONFIGURATION ======

# Model names and paths
MODELS = {
    "chat": {
        "name": "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
        "type": "causal-lm",
        "params": 1100_000_000
    },
    "code": {
        "name": "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
        "type": "causal-lm",
        "params": 1100_000_000
    },
    "summarization": {
        "name": "google/flan-t5-small",
        "type": "seq2seq",
        "params": 170_000_000
    }
}

# ====== GENERATION PARAMETERS ======

GENERATION_CONFIG = {
    "chat": {
        "max_tokens_limit": 200,      # Hard limit for CPU stability
        "default_max_tokens": 150,    # Default for user
        "min_temperature": 0.1,
        "max_temperature": 1.0,
        "default_temperature": 0.7,
        "top_p": 0.9,
        "top_k": 50,
    },
    "code": {
        "max_tokens_limit": 300,
        "default_max_tokens": 256,
        "min_temperature": 0.1,
        "max_temperature": 1.0,
        "default_temperature": 0.3,  # Lower temp for deterministic code
        "top_p": 0.95,
        "top_k": 40,
    },
    "summarization": {
        "max_length": 150,
        "min_length": 20,
        "max_input_length": 1000,    # Truncate input to prevent OOM
    },
    "image": {
        "max_width": 256,
        "max_height": 256,
        "default_width": 256,
        "default_height": 256,
        "min_width": 128,
        "min_height": 128,
    }
}

# ====== RESOURCE CONFIGURATION ======

RESOURCE_CONFIG = {
    "device": "cpu",
    "cpu_threads": 4,               # Number of CPU threads for torch
    "torch_dtype": "float32",       # Use float32 for CPU (not bfloat16)
    "low_cpu_mem_usage": True,      # Enable memory optimization
    "max_memory_percent": 85,       # Don't exceed 85% of available RAM
}

# ====== PERFORMANCE CONFIGURATION ======

PERFORMANCE_CONFIG = {
    # Queue settings
    "queue_max_size": 10,           # Max requests in queue
    "concurrency_limit": 2,         # Max concurrent requests
    
    # Timeouts (seconds)
    "model_load_timeout": 120,      # Timeout for loading models
    "request_timeout": 60,          # Timeout for generation
    
    # Memory management
    "gc_collect_freq": 1,           # Collect garbage after every N requests
    "model_unload_after": 300,      # Unload inactive models after N seconds
}

# ====== GRADIO CONFIGURATION ======

GRADIO_CONFIG = {
    "server_name": "0.0.0.0",
    "server_port": 7860,
    "share": False,
    "show_error": True,
    "analytics_enabled": False,
    "max_threads": 40,
}

# ====== UI CONFIGURATION ======

UI_CONFIG = {
    "title": "🤖 Lightweight AI Backend",
    "description": "Multi-Model API for Hugging Face Free CPU Tier",
    "theme": "soft",  # Options: default, soft, monochrome, glass
    "show_api_docs": True,
}

# ====== LOGGING CONFIGURATION ======

LOGGING_CONFIG = {
    "log_level": "INFO",            # DEBUG, INFO, WARNING, ERROR
    "log_to_file": False,
    "log_file": "app.log",
}

# ====== ADVANCED OPTIONS ======

ADVANCED_CONFIG = {
    # Model quantization (experimental)
    "use_quantization": False,      # Not recommended for CPU
    
    # Model compilation (experimental)
    "use_torch_compile": False,     # Not stable on CPU yet
    
    # Cache settings
    "cache_embeddings": False,      # Cache model outputs (memory tradeoff)
    "enable_attention_cache": True, # Enable KV cache
    
    # Error handling
    "max_retries": 2,
    "retry_delay": 1,              # Seconds between retries
}

# ====== ALTERNATIVE MODELS (For experimentation) ======

ALTERNATIVE_MODELS = {
    "chat_alternatives": [
        "microsoft/phi-1",           # Requires > 4GB RAM
        "stabilityai/stablelm-zephyr-3b",
        "TinyLlama/TinyLlama-1.1B-Chat-v1.0",  # RECOMMENDED
    ],
    "code_alternatives": [
        "Salesforce/codet5-small",
        "TinyLlama/TinyLlama-1.1B-Chat-v1.0",  # RECOMMENDED
    ],
    "summarization_alternatives": [
        "google/flan-t5-small",             # RECOMMENDED (170M)
        "google/flan-t5-base",              # Larger (250M)
        "facebook/bart-large-cnn",          # Heavier (406M)
    ],
}

# ====== FEATURE FLAGS ======

FEATURES = {
    "enable_chat": True,
    "enable_code": True,
    "enable_summarization": True,
    "enable_image_generation": True,
    
    # API exposure
    "expose_api": True,
    "api_auth_required": False,    # Require auth for API calls
    
    # UI tabs
    "show_about_tab": True,
    "show_api_docs": True,
}

# ====== DEPLOYMENT NOTES ======
"""
OPTIMIZATION TIPS FOR FREE CPU TIER:

1. Token Limits:
   - Keep max_tokens_limit at 200-300 for stability
   - Process smaller batches (1-2 requests at a time)

2. Memory Management:
   - Queue handling prevents concurrent large loads
   - Models lazy-load only when needed
   - Garbage collection runs after each request

3. Model Selection:
   - TinyLlama: 1.1B params, good quality/speed tradeoff
   - FLAN-T5-small: 170M params, fastest summarization
   - Procedural images: No ML overhead

4. If Space Crashes:
   - Reduce max_tokens limits further
   - Lower concurrency_limit to 1
   - Disable image generation tab

5. Expected Performance:
   - First request: 8-15s (includes model loading)
   - Subsequent requests: 2-8s per request
   - Total RAM: ~1.5-2GB
   - Idle memory: ~500MB-1GB
"""