chat.py

import os
import sys
import subprocess
import venv
import json
import math
import argparse
import shutil
import csv
import xml.etree.ElementTree as ET

SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
TEMP_DIR = os.path.join(SCRIPT_DIR, ".temp")
MODEL_DIR = os.path.join(SCRIPT_DIR, "banbtp-final")
DATASET_DIR = os.path.join(SCRIPT_DIR, "dataset")
BANNER_STATE_FILE = os.path.join(MODEL_DIR, "banner_state.pt")

os.makedirs(TEMP_DIR, exist_ok=True)
os.makedirs(DATASET_DIR, exist_ok=True)

os.environ["HF_HOME"] = os.path.join(TEMP_DIR, "hf_home")
os.environ["TOKENIZERS_PARALLELISM"] = "false"

def bootstrap_venv():
    venv_dir = os.path.join(TEMP_DIR, "venv")
    python_exe = os.path.join(venv_dir, "Scripts" if os.name == "nt" else "bin", "python")
    if os.path.realpath(sys.executable) != os.path.realpath(python_exe):
        if not os.path.exists(python_exe):
            print(">>> Creating isolated virtual environment...")
            venv.create(venv_dir, with_pip=True)
            subprocess.check_call([python_exe, "-m", "pip", "install", "--upgrade", "pip", "-q"], env=os.environ)
            subprocess.check_call([
                python_exe, "-m", "pip", "install",
                "torch", "transformers>=4.38.0", "safetensors", "sentencepiece", "tqdm",
                "pandas", "pyarrow", "accelerate", "-q"
            ], env=os.environ)
        os.execv(python_exe, [python_exe] + sys.argv)

bootstrap_venv()

import torch
import torch.nn.functional as F
from transformers import AutoModelForCausalLM, AutoTokenizer
from tqdm import tqdm
import pandas as pd
import pyarrow.parquet as pq
import pyarrow as pa

# ==============================================================================
# Banner Engine (Extra Parameters / Non-Parametric Memory)
# ==============================================================================
class BannerEngine:
    def __init__(self, model, tokenizer, device):
        self.model = model
        self.tokenizer = tokenizer
        self.device = device
        self.markov = {}
        self.rag_keys = []
        self.rag_texts = []
        self.gene_pool = torch.zeros((1000, model.config.hidden_size))
        self.fitness = torch.zeros(1000)
        self.generation = 0

    def extract_text(self, file_path):
        ext = os.path.splitext(file_path)[1].lower()
        try:
            if ext == '.txt':
                with open(file_path, 'r', encoding='utf-8', errors='ignore') as f: return f.read()
            elif ext == '.json':
                with open(file_path, 'r', encoding='utf-8', errors='ignore') as f: return json.dumps(json.load(f))
            elif ext == '.jsonl':
                with open(file_path, 'r', encoding='utf-8', errors='ignore') as f: return '\n'.join([json.dumps(json.loads(line)) for line in f if line.strip()])
            elif ext == '.csv':
                df = pd.read_csv(file_path, on_bad_lines='skip')
                return ' '.join(df.astype(str).agg(' '.join, axis=1).tolist())
            elif ext == '.tsv':
                df = pd.read_csv(file_path, sep='\t', on_bad_lines='skip')
                return ' '.join(df.astype(str).agg(' '.join, axis=1).tolist())
            elif ext == '.parquet':
                df = pq.read_table(file_path).to_pandas()
                return ' '.join(df.astype(str).agg(' '.join, axis=1).tolist())
            elif ext in ['.arrow', '.feather']:
                reader = pa.ipc.RecordBatchFileReader(pa.memory_map(file_path))
                df = reader.read_all().to_pandas()
                return ' '.join(df.astype(str).agg(' '.join, axis=1).tolist())
            elif ext == '.xml':
                tree = ET.parse(file_path)
                return ''.join(tree.getroot().itertext())
        except Exception as e:
            print(f"  [Warning] Could not parse {os.path.basename(file_path)}: {e}")
        return ""

    def ingest_text(self, text, chunk_size=256):
        if not text.strip(): return
        tokens = self.tokenizer.encode(text, add_special_tokens=False)
        
        for i in range(len(tokens)-1):
            prev, nxt = tokens[i], tokens[i+1]
            if prev not in self.markov: self.markov[prev] = {}
            self.markov[prev][nxt] = self.markov[prev].get(nxt, 0) + 1
            
        for i in range(0, len(tokens), chunk_size):
            chunk_tokens = tokens[i:i+chunk_size]
            if len(chunk_tokens) < 10: continue
            chunk_text = self.tokenizer.decode(chunk_tokens, skip_special_tokens=True)
            
            input_ids = torch.tensor([chunk_tokens], dtype=torch.long).to(self.device)
            with torch.no_grad():
                embeds = self.model.model.embed_tokens(input_ids)
            chunk_embed = embeds.mean(dim=1).squeeze(0).cpu()
            
            self.rag_keys.append(chunk_embed)
            self.rag_texts.append(chunk_text)
            
        if tokens:
            input_ids = torch.tensor([tokens[:256]], dtype=torch.long).to(self.device)
            with torch.no_grad():
                embeds = self.model.model.embed_tokens(input_ids)
            self.adapt_genetic(embeds.mean(dim=1).squeeze(0).cpu(), 0.5)

    def adapt_genetic(self, hidden_state, error):
        weakest = torch.argmin(self.fitness)
        mutation = torch.randn_like(self.gene_pool[weakest]) * error
        self.gene_pool[weakest] = hidden_state + mutation
        self.fitness[weakest] = 1.0 / (error + 1e-5)
        self.generation += 1

    def retrieve(self, query_text, top_k=2):
        if not self.rag_keys: return [], 0.0
        tokens = self.tokenizer.encode(query_text, add_special_tokens=False)[:256]
        if not tokens: return [], 0.0
        input_ids = torch.tensor([tokens], dtype=torch.long).to(self.device)
        with torch.no_grad():
            q_embed = self.model.model.embed_tokens(input_ids).mean(dim=1).squeeze(0).cpu()
            
        keys_tensor = torch.stack(self.rag_keys)
        sims = F.cosine_similarity(q_embed.unsqueeze(0), keys_tensor, dim=1)
        top_sims, top_indices = torch.topk(sims, k=min(top_k, len(self.rag_texts)))
        
        max_sim = top_sims[0].item() if len(top_sims) > 0 else 0.0
        retrieved = [self.rag_texts[i] for i in top_indices.tolist()]
        return retrieved, max_sim

    def save_state(self):
        state = {
            "markov": self.markov,
            "rag_keys": self.rag_keys,
            "rag_texts": self.rag_texts,
            "gene_pool": self.gene_pool,
            "fitness": self.fitness,
            "generation": self.generation
        }
        torch.save(state, BANNER_STATE_FILE)

    def load_state(self):
        if not os.path.exists(BANNER_STATE_FILE): return False
        state = torch.load(BANNER_STATE_FILE, map_location="cpu", weights_only=False)
        self.markov = state.get("markov", {})
        self.rag_keys = state.get("rag_keys", [])
        self.rag_texts = state.get("rag_texts", [])
        self.gene_pool = state.get("gene_pool", torch.zeros((1000, self.model.config.hidden_size)))
        self.fitness = state.get("fitness", torch.zeros(1000))
        self.generation = state.get("generation", 0)
        return True

# ==============================================================================
# Finetune Mode
# ==============================================================================
def run_finetune(banner):
    print(f">>> Scanning {DATASET_DIR} for dataset files...")
    supported_ext = ('.txt', '.json', '.jsonl', '.csv', '.tsv', '.xml', '.parquet', '.arrow', '.feather')
    files = []
    for root, dirs, filenames in os.walk(DATASET_DIR):
        for f in filenames:
            if f.lower().endswith(supported_ext):
                files.append(os.path.join(root, f))
                
    if not files:
        print(">>> No supported files found in dataset folder.")
        return

    print(f">>> Found {len(files)} files. Ingesting into Banner extra parameters...")
    for fpath in tqdm(files, desc="Processing Files", unit="file"):
        text = banner.extract_text(fpath)
        banner.ingest_text(text)
        try:
            os.remove(fpath)
        except Exception as e:
            print(f"  [Warning] Could not delete {fpath}: {e}")
            
    for root, dirs, files in os.walk(DATASET_DIR, topdown=False):
        for name in dirs:
            dir_path = os.path.join(root, name)
            if dir_path != DATASET_DIR:
                try:
                    os.rmdir(dir_path)
                except OSError:
                    pass

    banner.save_state()
    print(f">>> Finetune complete! Banner state saved to {BANNER_STATE_FILE}")
    print(f"    Markov transitions: {sum(len(v) for v in banner.markov.values()):,}")
    print(f"    RAG memories: {len(banner.rag_texts):,}")
    print(f"    Genetic generation: {banner.generation}")
    print(f">>> Dataset folder cleared automatically.")

# ==============================================================================
# Chat Mode
# ==============================================================================
def run_chat(banner, model, tokenizer, device):
    temp = 0.7
    max_tokens = 64
    auto_temp = True
    
    print("=" * 50)
    print("  banbtp2.0v10 chat + Banner Engine")
    print("  type /help for commands")
    print("=" * 50)

    while True:
        try:
            user_input = input("\nyou> ").strip()
        except (EOFError, KeyboardInterrupt):
            print("\n>>> Saving banner state...")
            banner.save_state()
            break

        if not user_input: continue

        if user_input.startswith("/"):
            cmd_parts = user_input.split()
            cmd = cmd_parts[0].lower()
            if cmd in ("/quit", "/exit", "/q"):
                print(">>> Saving banner state...")
                banner.save_state()
                break
            elif cmd == "/help":
                print("  /temp <0.1-2.0>   set manual temperature")
                print("  /auto             toggle auto-temperature (RAG adaptive)")
                print("  /tokens <n>       set max new tokens")
                print("  /stats            show banner memory stats")
                print("  /save             save banner state now")
                print("  /quit             save and exit")
            elif cmd == "/temp":
                if len(cmd_parts) > 1:
                    try:
                        temp = max(0.1, min(2.0, float(cmd_parts[1])))
                        auto_temp = False
                        print(f"  manual temperature = {temp}")
                    except ValueError: print("  usage: /temp <number>")
                else: print(f"  temperature = {temp}")
            elif cmd == "/auto":
                auto_temp = not auto_temp
                print(f"  auto-temperature {'ON' if auto_temp else 'OFF'}")
            elif cmd == "/tokens":
                if len(cmd_parts) > 1:
                    try:
                        max_tokens = max(1, int(cmd_parts[1]))
                        print(f"  max_new_tokens = {max_tokens}")
                    except ValueError: print("  usage: /tokens <number>")
                else: print(f"  max_new_tokens = {max_tokens}")
            elif cmd == "/save":
                banner.save_state()
                print("  banner state saved.")
            elif cmd == "/stats":
                print(f"  markov transitions: {sum(len(v) for v in banner.markov.values()):,}")
                print(f"  rag memories: {len(banner.rag_texts):,}")
                print(f"  genetic generation: {banner.generation}")
            else:
                print(f"  unknown command: {cmd}. type /help")
            continue

        retrieved, max_sim = banner.retrieve(user_input, top_k=2)
        if retrieved and max_sim > 0.5:
            context = "\n".join(retrieved)
            full_prompt = f"{context}\n{user_input}"
            current_temp = max(0.2, temp * 0.5) if auto_temp else temp
        else:
            full_prompt = user_input
            current_temp = temp * 1.2 if auto_temp else temp

        input_ids = tokenizer(full_prompt, return_tensors="pt", truncation=True, max_length=512)["input_ids"].to(device)
        if input_ids.shape[1] == 0:
            print("model> [Empty prompt]")
            continue
            
        generated = []
        
        with torch.no_grad():
            for _ in tqdm(range(max_tokens), desc="Thinking", bar_format='{l_bar}{bar}| {n_fmt}/{total_fmt}', leave=False):
                outputs = model(input_ids)
                next_token_logits = outputs.logits[:, -1, :] / current_temp
                
                if torch.isnan(next_token_logits).any():
                    next_token_logits = torch.nan_to_num(next_token_logits, nan=0.0)
                    
                probs = F.softmax(next_token_logits, dim=-1)
                next_token = torch.multinomial(probs, num_samples=1)
                
                generated.append(next_token.item())
                if next_token.item() == tokenizer.eos_token_id:
                    break
                    
                input_ids = torch.cat([input_ids, next_token], dim=-1)
                if input_ids.shape[1] > 1024:
                    input_ids = input_ids[:, -1024:]

        response = tokenizer.decode(generated, skip_special_tokens=True)
        
        out_tokens = tokenizer.encode(response, add_special_tokens=False)
        for i in range(len(out_tokens)-1):
            prev, nxt = out_tokens[i], out_tokens[i+1]
            if prev not in banner.markov: banner.markov[prev] = {}
            banner.markov[prev][nxt] = banner.markov[prev].get(nxt, 0) + 1
            
        if out_tokens:
            input_ids_embed = torch.tensor([out_tokens[:256]], dtype=torch.long).to(device)
            with torch.no_grad():
                embeds = model.model.embed_tokens(input_ids_embed)
            banner.adapt_genetic(embeds.mean(dim=1).squeeze(0).cpu(), 0.1)

        print(f"\nmodel> {response}")

# ==============================================================================
# Main
# ==============================================================================
def main():
    parser = argparse.ArgumentParser(description="banbtp2.0v10 chat interface")
    parser.add_argument("--chat", action="store_true", help="chat mode (default)")
    parser.add_argument("--finetune", action="store_true", help="ingest dataset folder into banner memory")
    args = parser.parse_args()

    if not os.path.exists(MODEL_DIR):
        print(f"ERROR: Model directory not found: {MODEL_DIR}")
        sys.exit(1)

    print(f">>> Loading banbtp2.0v10 from: {MODEL_DIR}")
    tokenizer = AutoTokenizer.from_pretrained(MODEL_DIR)
    
    # FIX: Removed device_map="auto" to prevent accelerate crash. 
    # Added trust_remote_code=True so it loads the custom arch silently.
    model = AutoModelForCausalLM.from_pretrained(
        MODEL_DIR,
        dtype=torch.float32,
        trust_remote_code=True
    )
    model.eval()
    device = next(model.parameters()).device

    banner = BannerEngine(model, tokenizer, device)
    if banner.load_state():
        print(">>> Loaded existing Banner state.")

    if args.finetune:
        run_finetune(banner)
    else:
        run_chat(banner, model, tokenizer, device)

if __name__ == "__main__":
    main()