handler.py

import os
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from typing import Dict, List, Any # <-- ADD THIS LINE

class EndpointHandler():
    def __init__(self, model_id: str):
        """
        Initializes the handler by loading the model and tokenizer.

        Args:
            model_id (str): The Hugging Face model ID (e.g., "MoritzLaurer/DeBERTa-v3-base-mnli")
                            This is automatically passed by the Inference Endpoint infrastructure.
        """
        print(f"Loading model '{model_id}'...")
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        print(f"Using device: {self.device}")

        self.tokenizer = AutoTokenizer.from_pretrained(model_id)
        self.model = AutoModelForSequenceClassification.from_pretrained(model_id)

        # Move model to the determined device
        self.model.to(self.device)
        # Set model to evaluation mode for consistent inference
        self.model.eval()
        print("Model and tokenizer loaded successfully.")

        # --- Determine Label Order ---
        # Preferred: Dynamically get labels from model config
        try:
            # Sort by ID to ensure consistent order if dict isn't ordered
            sorted_labels = sorted(self.model.config.id2label.items())
            self.label_names = [label for _, label in sorted_labels]
            print(f"Using label names from model config: {self.label_names}")
            # Basic validation for NLI task
            if len(self.label_names) != 3:
                 print(f"Warning: Expected 3 labels for NLI, but model config has {len(self.label_names)}. Proceeding with model's labels.")
            if not any("entail" in l.lower() for l in self.label_names) or \
               not any("neutral" in l.lower() for l in self.label_names) or \
               not any("contra" in l.lower() for l in self.label_names):
                 print(f"Warning: Model labels {self.label_names} might not match standard NLI labels ('entailment', 'neutral', 'contradiction').")

        except AttributeError:
            # Fallback: Use the explicitly requested labels if config is missing/malformed
            self.label_names = ["entailment", "neutral", "contradiction"]
            print(f"Warning: Could not read labels from model config. Falling back to default: {self.label_names}")
            print("Ensure this order matches the actual output order of the model!")

        print(f"Configured label order for output: {self.label_names}")


    # Corrected type hints in the signature below
    def __call__(self, data: Dict[str, Any]) -> Dict[str, Any] | List[Dict[str, Any]]:
        """
        Handles inference requests.

        Args:
            data (Dict[str, Any]): The input data payload from the request.
                                   Expected keys: "premise" (str) and "hypothesis" (str).
                                   Can optionally be nested under "inputs".

        Returns:
            Dict[str, Any] | List[Dict[str, Any]]: A dictionary containing error info,
                                                    or a list of dictionaries, each mapping
                                                    a label name to its probability score.
        """
        # --- Input Parsing ---
        inputs = data.get("inputs", data) # Allow for optional "inputs" nesting
        premise = inputs.get("premise")
        hypothesis = inputs.get("hypothesis")

        # Basic input validation
        if not premise or not isinstance(premise, str):
            return {"error": "Missing or invalid 'premise' key in input. Expected a string."}
        if not hypothesis or not isinstance(hypothesis, str):
            return {"error": "Missing or invalid 'hypothesis' key in input. Expected a string."}

        # --- Tokenization ---
        # Tokenize the premise-hypothesis pair
        try:
            tokenized_inputs = self.tokenizer(
                premise,
                hypothesis,
                return_tensors="pt", # Return PyTorch tensors
                truncation=True,     # Truncate if longer than max length
                padding=True,        # Pad to the longest sequence in the batch (or max_length)
                max_length=self.tokenizer.model_max_length # Use model's max length
            )
        except Exception as e:
             print(f"Error during tokenization: {e}")
             return {"error": f"Failed to tokenize input: {e}"}


        # Move tokenized inputs to the same device as the model
        tokenized_inputs = {k: v.to(self.device) for k, v in tokenized_inputs.items()}

        # --- Inference ---
        try:
            with torch.no_grad(): # Disable gradient calculations for efficiency
                outputs = self.model(**tokenized_inputs)
                logits = outputs.logits

            # Apply Softmax to get probabilities
            probabilities = torch.softmax(logits, dim=-1)

            # Move probabilities to CPU and convert to list
            # Squeeze or index [0] if processing single pairs (typical for endpoints)
            scores = probabilities.cpu().numpy()[0].tolist()

            # --- Format Output ---
            # Pair labels with their corresponding scores
            result = [{"label": label, "score": score} for label, score in zip(self.label_names, scores)]

            return result

        except Exception as e:
            print(f"Error during model inference: {e}")
            # Consider logging the full traceback here in a real deployment
            # import traceback
            # traceback.print_exc()
            return {"error": f"Model inference failed: {e}"}