models/tinygpt2_model.py

"""
TinyGPT2 Model Wrapper for easy integration (CPU-friendly)
"""
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
import os
HUGGINGFACE_TOKEN = os.getenv("HUGGINGFACE_TOKEN")
class TinyGPT2Model:
    """
    Wrapper for sshleifer/tiny-gpt2 model with caching and optimization
    Suitable for CPU-only Hugging Face Spaces
    """
    _instance = None
    _model = None
    _tokenizer = None

    def __new__(cls):
        if cls._instance is None:
            cls._instance = super().__new__(cls)
        return cls._instance

    def __init__(self):
        if TinyGPT2Model._model is None:
            self._initialize_model()

    def _initialize_model(self):
        """Initialize Tiny-GPT2 model"""
        print("Loading TinyGPT2 model...")

        model_id = "sshleifer/tiny-gpt2"

        # Load tokenizer (no need for token argument, model is public)
        TinyGPT2Model._tokenizer = AutoTokenizer.from_pretrained(model_id,token=HUGGINGFACE_TOKEN)

        # Load model (no quantization, pure CPU)
        TinyGPT2Model._model = AutoModelForCausalLM.from_pretrained(
            model_id,token=HUGGINGFACE_TOKEN,
            torch_dtype=torch.float32  # Safe for CPU only
        )

        print("TinyGPT2 model loaded successfully!")

    def generate(
        self, 
        prompt: str, 
        max_length: int = 64,
        temperature: float = 0.7,
        top_p: float = 0.95
    ) -> str:
        """Generate response from TinyGPT2"""

        # For TinyGPT2, no special prompt formatting needed
        formatted_prompt = prompt

        # Tokenize
        inputs = TinyGPT2Model._tokenizer(
            formatted_prompt,
            return_tensors="pt",
            truncation=True,
            max_length=256
        )

        # Move to CPU (optional, for explicitness)
        inputs = {k: v.cpu() for k, v in inputs.items()}

        # Generate on CPU
        with torch.no_grad():
            outputs = TinyGPT2Model._model.generate(
                **inputs,
                max_new_tokens=max_length,
                temperature=temperature,
                top_p=top_p,
                do_sample=True,
                pad_token_id=TinyGPT2Model._tokenizer.eos_token_id
            )

        # Decode only the newly generated tokens (after the prompt)
        response = TinyGPT2Model._tokenizer.decode(
            outputs[0][inputs['input_ids'].shape[1]:],
            skip_special_tokens=True
        )

        return response.strip()

    def generate_embedding(self, text: str) -> torch.Tensor:
        """Generate embeddings for text using last hidden state"""
        inputs = TinyGPT2Model._tokenizer(
            text,
            return_tensors="pt",
            truncation=True,
            max_length=256
        )
        inputs = {k: v.cpu() for k, v in inputs.items()}

        with torch.no_grad():
            outputs = TinyGPT2Model._model(**inputs, output_hidden_states=True)
            embeddings = outputs.hidden_states[-1].mean(dim=1)

        return embeddings