Oracle/SmolLM.py

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

class SmolLM:
    def __init__(self, model_path="HuggingFaceTB/SmolLM2-1.7B-Instruct"):
        self.available = True
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        try:
            print(f"[INFO] Loading Oracle tokenizer from {model_path}")
            self.tokenizer = AutoTokenizer.from_pretrained(model_path)
            print(f"[INFO] Loading Oracle from {model_path} on {self.device}")
            self.model = AutoModelForCausalLM.from_pretrained(model_path).to(self.device)
            print("[INFO] Oracle loaded successfully")
        except Exception as e:
            print(f"[ERROR] Failed to load model '{model_path}': {e}")
            self.available = False

    def predict(self, prompt, max_new_tokens=200):
        if not self.available:
            print("[WARN] Oracle unavailable, returning default weight 0.5")
            return ""
        try:
            # Use chat template as per documentation
            messages = [{"role": "user", "content": prompt}]
            inputs = self.tokenizer.apply_chat_template(messages, return_tensors="pt").to(self.device)
            outputs = self.model.generate(
                inputs,
                max_new_tokens=max_new_tokens,
                temperature=0.2,
                top_p=0.9,
                do_sample=True
            )
            response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
            print(f"[INFO] Generated response: {response[:100]}...", flush=True)
            return response.split("<|assistant|>")[-1].strip()
        except Exception as e:
            print(f"[ERROR] Oracle has failed: {e}")
            return "0.5"