isa-model 0.4.0

Unified AI model serving framework

isa_model_sdk - Unified AI Model Serving Framework

A comprehensive Python framework for working with multiple AI providers and models through a unified interface. Support for OpenAI, Replicate, Ollama, and more, with advanced training and evaluation capabilities.

Installation

pip install isa_model_sdk

Quick Start

The isa_model_sdk package supports three main usage patterns:

1. Pass API Keys Directly (Recommended)

This is the most flexible approach - no environment variables needed:

from isa_model.inference.ai_factory import AIFactory

# Create factory instance
factory = AIFactory.get_instance()

# Use OpenAI with API key
llm = factory.get_llm(
    model_name="gpt-4o-mini", 
    provider="openai", 
    api_key="your-openai-api-key-here"
)

# Use Replicate for image generation
image_gen = factory.get_vision_model(
    model_name="stability-ai/sdxl", 
    provider="replicate", 
    api_key="your-replicate-token-here"
)

2. Use Environment Variables

Set your API keys as environment variables:

export OPENAI_API_KEY="your-openai-api-key"
export REPLICATE_API_TOKEN="your-replicate-token"

Then use without passing keys:

from isa_model.inference.ai_factory import AIFactory

factory = AIFactory.get_instance()

# Will automatically use OPENAI_API_KEY from environment
llm = factory.get_llm(model_name="gpt-4o-mini", provider="openai")

# Will automatically use REPLICATE_API_TOKEN from environment  
image_gen = factory.get_vision_model(model_name="stability-ai/sdxl", provider="replicate")

3. Use Local Models (No API Key Needed)

For local models like Ollama, no API keys are required:

from isa_model.inference.ai_factory import AIFactory

factory = AIFactory.get_instance()

# Use local Ollama model (no API key needed)
llm = factory.get_llm(model_name="llama3.1", provider="ollama")

🎯 Training & Evaluation Framework

NEW in v0.0.1: Comprehensive training and evaluation capabilities for LLMs, Stable Diffusion, and ML models.

Quick Training Example

from isa_model.training import TrainingFactory, train_gemma
from isa_model.eval import EvaluationFactory

# Quick Gemma training
model_path = train_gemma(
    dataset_path="tatsu-lab/alpaca",
    model_size="4b",
    num_epochs=3,
    use_lora=True
)

# Comprehensive evaluation
evaluator = EvaluationFactory(use_wandb=True)
results = evaluator.evaluate_llm(
    model_path=model_path,
    dataset_path="test_data.json",
    metrics=["perplexity", "bleu", "rouge"]
)

# Run benchmarks
mmlu_results = evaluator.run_benchmark(
    model_path=model_path,
    benchmark="mmlu"
)

Advanced Training Configuration

from isa_model.training import TrainingFactory

factory = TrainingFactory()

# Advanced LLM training
model_path = factory.train_model(
    model_name="google/gemma-2-4b-it",
    dataset_path="custom_dataset.json",
    use_lora=True,
    batch_size=4,
    num_epochs=3,
    learning_rate=2e-5,
    lora_rank=8,
    lora_alpha=16
)

# Upload to HuggingFace
hf_url = factory.upload_to_huggingface(
    model_path=model_path,
    hf_model_name="your-username/gemma-4b-custom",
    hf_token="your-hf-token"
)

Cloud Training on RunPod

# Train on RunPod cloud infrastructure
result = factory.train_on_runpod(
    model_name="google/gemma-2-4b-it",
    dataset_path="tatsu-lab/alpaca",
    runpod_api_key="your-runpod-key",
    template_id="your-template-id",
    gpu_type="NVIDIA RTX A6000"
)

Function Calling with bind_tools

Enhanced in v0.0.1: LangChain-compatible function calling interface for all LLM services.

Basic Function Calling

import asyncio
from isa_model.inference.ai_factory import AIFactory

# Define your tool functions
def get_weather(location: str) -> str:
    """Get weather information for a location"""
    weather_data = {
        "paris": "Sunny, 22°C",
        "london": "Cloudy, 18°C",
        "tokyo": "Clear, 25°C"
    }
    return weather_data.get(location.lower(), f"Weather data not available for {location}")

def calculate_math(expression: str) -> str:
    """Calculate a mathematical expression"""
    try:
        result = eval(expression)  # Use safely in production
        return f"The result of {expression} is {result}"
    except:
        return f"Error calculating {expression}"

async def main():
    factory = AIFactory.get_instance()
    
    # Create LLM with any provider
    llm = factory.get_llm("gpt-4o-mini", "openai", api_key="your-key")
    # or: llm = factory.get_llm("llama3.1", "ollama")  # Local model
    
    # Bind tools to the service (LangChain-style interface)
    llm_with_tools = llm.bind_tools([get_weather, calculate_math])
    
    # Use the service with tools
    response = await llm_with_tools.achat([
        {"role": "user", "content": "What's the weather in Paris? Also calculate 15 * 8"}
    ])
    
    print(response)  # Model will use tools automatically
    await llm.close()

asyncio.run(main())

Supported Services

Language Models (LLM)

# OpenAI models
llm = factory.get_llm("gpt-4o-mini", "openai", api_key="your-key")
llm = factory.get_llm("gpt-4o", "openai", api_key="your-key") 

# Ollama models (local)
llm = factory.get_llm("llama3.1", "ollama")
llm = factory.get_llm("codellama", "ollama")

# Replicate models
llm = factory.get_llm("meta/llama-3-70b-instruct", "replicate", api_key="your-token")

# All LLM services support bind_tools() for function calling
llm_with_tools = llm.bind_tools([your_functions])

Vision Models

# OpenAI vision
vision = factory.get_vision_model("gpt-4o", "openai", api_key="your-key")

# Replicate image generation
image_gen = factory.get_vision_model("stability-ai/sdxl", "replicate", api_key="your-token")

# Ollama vision (local)
vision = factory.get_vision_model("llava", "ollama")

Embedding Models

# OpenAI embeddings
embedder = factory.get_embedding("text-embedding-3-small", "openai", {"api_key": "your-key"})

# Ollama embeddings (local)
embedder = factory.get_embedding("bge-m3", "ollama")

Training Framework Features

Multi-Modal Training Support

• LLM Training: Gemma, Llama, GPT-style models with LoRA/QLoRA
• Stable Diffusion: Image generation model training
• ML Models: XGBoost, Random Forest, traditional ML
• Computer Vision: CNN, Vision Transformers

Training Modes

• Local Training: On your machine with CPU/GPU
• Cloud Training: RunPod, AWS, GCP integration
• Distributed Training: Multi-GPU support

Data Pipeline

• Annotation Service: Human-in-the-loop data annotation
• Dataset Management: HuggingFace, local, cloud storage
• Quality Control: Data validation and filtering

Evaluation Framework Features

Comprehensive Evaluation

• LLM Metrics: Perplexity, BLEU, ROUGE, BERTScore
• Benchmark Tests: MMLU, HellaSwag, ARC, GSM8K
• Image Metrics: FID, IS, LPIPS for generative models
• Custom Metrics: Domain-specific evaluations

Experiment Tracking

• Weights & Biases: Experiment tracking and visualization
• MLflow: Model registry and experiment management
• Model Comparison: Side-by-side performance analysis

Usage Examples

Chat Completion

import asyncio
from isa_model.inference.ai_factory import AIFactory

async def chat_example():
    factory = AIFactory.get_instance()
    llm = factory.get_llm("gpt-4o-mini", "openai", api_key="your-key")
    
    messages = [
        {"role": "user", "content": "Hello, how are you?"}
    ]
    
    response = await llm.achat(messages)
    print(response)

# Run the async function
asyncio.run(chat_example())

Image Generation

import asyncio
from isa_model.inference.ai_factory import AIFactory

async def image_gen_example():
    factory = AIFactory.get_instance()
    image_gen = factory.get_vision_model(
        "stability-ai/sdxl", 
        "replicate", 
        api_key="your-replicate-token"
    )
    
    result = await image_gen.generate_image(
        prompt="A beautiful sunset over mountains",
        width=1024,
        height=1024
    )
    
    # Save the generated image
    with open("generated_image.png", "wb") as f:
        f.write(result["image_data"])

asyncio.run(image_gen_example())

Complete Training and Evaluation Workflow

from isa_model.training import TrainingFactory
from isa_model.eval import EvaluationFactory

# Initialize factories
trainer = TrainingFactory()
evaluator = EvaluationFactory(use_wandb=True, wandb_project="my-project")

# Train model
model_path = trainer.train_model(
    model_name="google/gemma-2-4b-it",
    dataset_path="training_data.json",
    use_lora=True,
    num_epochs=3
)

# Evaluate model
results = evaluator.evaluate_llm(
    model_path=model_path,
    dataset_path="test_data.json",
    metrics=["bleu", "rouge", "accuracy"]
)

# Run benchmarks
benchmark_results = evaluator.run_benchmark(
    model_path=model_path,
    benchmark="mmlu"
)

# Compare with base model
comparison = evaluator.compare_models([
    "google/gemma-2-4b-it",  # Base model
    model_path  # Fine-tuned model
], benchmark="arc")

print(f"Training completed: {model_path}")
print(f"Evaluation results: {results}")

What's New in v0.0.1

🎯 Training Framework

• Multi-modal training: LLM, Stable Diffusion, ML models
• Cloud integration: RunPod training support
• LoRA/QLoRA: Memory-efficient fine-tuning
• HuggingFace integration: Direct dataset loading and model uploading

📊 Evaluation Framework

• Comprehensive metrics: BLEU, ROUGE, perplexity, and more
• Standard benchmarks: MMLU, HellaSwag, ARC, GSM8K
• Experiment tracking: Weights & Biases and MLflow integration
• Model comparison: Side-by-side performance analysis

🔧 Enhanced Inference

• Improved function calling: Better tool binding and execution
• Better error handling: More informative error messages
• Performance optimizations: Faster model loading and inference

Development

Installing for Development

git clone <repository-url>
cd isA_Model
pip install -e .

Running Tests

# Set environment variables
export OPENAI_API_KEY="your-key"
export REPLICATE_API_TOKEN="your-token"

# Run inference tests
python tests/units/inference/test_all_services.py

# Run training tests
python tests/test_training_setup.py

Building and Publishing

# Build the package
python -m build

# Upload to PyPI (requires PYPI_API_TOKEN in .env.local)
source .venv/bin/activate
source .env.local
python -m twine upload dist/isa_model_sdk-0.0.1* --username __token__ --password "$PYPI_API_TOKEN"

License

MIT License - see LICENSE file for details.

Contributing

Contributions are welcome! Please read our contributing guidelines and submit pull requests to our GitHub repository.

Support

For questions and support, please open an issue on our GitHub repository.