chat_with_tars.py

import os
import torch
import torch.nn as nn
from transformers import BertModel, GPTNeoForCausalLM, AutoTokenizer

# ⚙️ Ensure temporary directory is writable
os.environ["TMPDIR"] = os.path.expanduser("~/tmp")
os.makedirs(os.environ["TMPDIR"], exist_ok=True)

# 💠 Optional modules (brain & heart, if available)
heart_module = None
brain_module = None

if os.path.isdir("heart"):
    try:
        from heart import heart
        heart_module = heart
    except Exception as e:
        print(f"[⚠️] Heart module error: {e}")

if os.path.isdir("brain"):
    try:
        from brain import brain
        brain_module = brain
    except Exception as e:
        print(f"[⚠️] Brain module error: {e}")

# TARSQuantumHybrid Class
class TARSQuantumHybrid(nn.Module):
    def __init__(self, bert_model="bert-base-uncased", gpt_model="EleutherAI/gpt-neo-125M"):
        super(TARSQuantumHybrid, self).__init__()
        self.bert = BertModel.from_pretrained(bert_model)
        self.gpt = GPTNeoForCausalLM.from_pretrained(gpt_model)

        gpt_hidden_dim = getattr(self.gpt.config, "hidden_size", None) or getattr(self.gpt.config, "n_embd", 768)
        self.embedding_proj = nn.Linear(self.bert.config.hidden_size, gpt_hidden_dim)

        self.tokenizer = AutoTokenizer.from_pretrained(gpt_model)

        # Ensure the tokenizer has a padding token
        if self.tokenizer.pad_token is None:
            self.tokenizer.add_special_tokens({'pad_token': '[PAD]'})
            self.gpt.resize_token_embeddings(len(self.tokenizer))
            print("✅ Padding token added and model resized.")

    def forward(self, input_ids, attention_mask=None, decoder_input_ids=None):
        bert_output = self.bert(input_ids=input_ids, attention_mask=attention_mask)
        cls_embedding = bert_output.last_hidden_state[:, 0, :]
        gpt_input = self.embedding_proj(cls_embedding).unsqueeze(1)
        outputs = self.gpt(inputs_embeds=gpt_input, decoder_input_ids=decoder_input_ids)
        return outputs

    def chat(self, text, max_length=128):
        # 🧠 Tokenize the input text
        cleaned_text = self.clean_input_text(text)
        if not cleaned_text.strip():
            return "🤖 Please provide a non-empty input."

        encoded_input = self.safe_tokenization(cleaned_text)
        
        # Extract input_ids and attention_mask
        input_ids = encoded_input["input_ids"]
        attention_mask = encoded_input["attention_mask"]

        # Debug: Check the token IDs and vocab size
        print(f"Input Text: {cleaned_text}")
        print(f"Input IDs: {input_ids}")
        print(f"Vocabulary Size: {self.tokenizer.vocab_size}")
        
        # Ensure token IDs are within bounds
        if input_ids.numel() > 0 and input_ids.max() >= self.tokenizer.vocab_size:
            raise ValueError(f"Token ID exceeds model's vocabulary size: {input_ids.max()}")

        decoder_input_ids = torch.tensor([[self.tokenizer.bos_token_id]])

        # 🧪 Generate output using the model
        with torch.no_grad():
            outputs = self.forward(
                input_ids=input_ids,
                attention_mask=attention_mask,
                decoder_input_ids=decoder_input_ids,
            )
            generated_ids = torch.argmax(outputs.logits, dim=-1)

        # Debug: Check the generated token IDs
        print(f"Generated Token IDs: {generated_ids}")

        raw_response = self.tokenizer.decode(generated_ids[0], skip_special_tokens=True)
        
        # 🧼 Clean model echo by removing the original input from the response
        cleaned = raw_response.replace(cleaned_text, "").strip()

        # 🧠 Add insights from optional modules (brain & heart)
        extra_thoughts = ""
        if brain_module and hasattr(brain_module, "get_brain_insight"):
            extra_thoughts += f"\n🧠 {brain_module.get_brain_insight()}"
        if heart_module and hasattr(heart_module, "get_heart_feeling"):
            extra_thoughts += f"\n❤️ {heart_module.get_heart_feeling()}"

        # 🪄 Return final response
        final_response = cleaned if cleaned else "🤖 ...processing quantum entanglement..."
        return final_response + extra_thoughts

    def clean_input_text(self, text):
        # Remove unwanted characters
        cleaned_text = ''.join(e for e in text if e.isalnum() or e.isspace())
        return cleaned_text

    def safe_tokenization(self, text):
        token_ids = self.tokenizer.encode(text, add_special_tokens=True)
        # Ensure that token ids are within vocabulary size
        token_ids = [min(i, self.tokenizer.vocab_size - 1) for i in token_ids]
        return {
            "input_ids": torch.tensor(token_ids).unsqueeze(0),  # Adding batch dimension
            "attention_mask": torch.ones((1, len(token_ids)), dtype=torch.long)
        }

# ✅ Torch-compatible loader
def load_tars(path="tars_v1.pt"):
    from torch.serialization import add_safe_globals
    add_safe_globals({"TARSQuantumHybrid": TARSQuantumHybrid})

    model = torch.load(path, weights_only=False)
    model.eval()
    return model

# ✅ Start chat loop
if __name__ == "__main__":
    print("🤖 TARS model loaded successfully. Ready to chat!")
    model = load_tars()

    while True:
        prompt = input("You: ")
        if prompt.strip().lower() in ["exit", "quit"]:
            print("TARS: Till we meet again in the quantum field. 🌌")
            break
        response = model.chat(prompt)
        print(f"TARS: {response}")