golem.py

from __future__ import annotations

print("\n 🎲 starting GPT-2-mini 💎 Sapphire β inference wrapper\n")
import os, sys, types, importlib.machinery as _machinery

# ---- HARD DISABLE TENSORFLOW / FLAX ----
#os.environ["TRANSFORMERS_NO_TF"]   = "1"
#os.environ["TRANSFORMERS_NO_FLAX"] = "1"
#os.environ["USE_TF"]               = "0"

"""
# Create a dummy tensorflow module *with a valid __spec__*
_tf_stub = types.ModuleType("tensorflow")
_tf_stub.__dict__["__getattr__"] = lambda name: None            # any attr → None
_tf_stub.__spec__ = _machinery.ModuleSpec("tensorflow", loader=None)
_tf_stub.__path__ = []                                          # pretend it's a pkg
sys.modules["tensorflow"] = _tf_stub

# Same for flax (rarely, but keeps everything clean)
_flax_stub = types.ModuleType("flax")
_flax_stub.__getattr__ = lambda name: None
_flax_stub.__spec__ = _machinery.ModuleSpec("flax", loader=None)
_flax_stub.__path__ = []
sys.modules["flax"] = _flax_stub
# ----------------------------------------

# now safe to import any HuggingFace / torch code
from transformers import GPT2Tokenizer, GPT2LMHeadModel
"""

import sys, os

# 🔥 Force local transformers override
#local_transformers = os.path.join(os.path.dirname(__file__), "transformers")
#if os.path.isdir(local_transformers):
#    sys.path.insert(0, local_transformers)
#    print("🔧 Using local transformers from:", local_transformers)
#else:
#    print("⚠️ WARNING: Local transformers directory not found!")

MAX_FORWARD_TOKENS = 75
import warnings
warnings.filterwarnings("ignore") 

# --- locate Sapphire repo root ------------------------------------------------
import pathlib
sys.path.insert(0, os.path.abspath(os.path.dirname(__file__)))
sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "cli")))
sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "core")))
sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "utils")))
print("💾 working directory : ", os.path.abspath(os.path.dirname(__file__)))
from itertools import chain
from collections.abc import Mapping
import warnings, os, re, json, glob, argparse, shutil, math, time
from dataclasses import dataclass
from datetime import datetime
from typing import List, Sequence, Tuple
import torch, torch.nn.functional as F
import numpy as np
import transformers
from difflib import SequenceMatcher
from collections import Counter
from transformers import (
    GPT2Tokenizer, GPT2LMHeadModel, Trainer, TrainingArguments, default_data_collator
)
from sentence_transformers import SentenceTransformer, util
import nltk 
nltk.download('punkt') # you can comment out dwnload after first shot gets i
import glob

from cli.settings_manager import handle_settings_command
### Hardware check
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")

###
from core.param_store import ParamStore

# ────────────────────────────────────────────────────────────────────────
# 1.   PROMPT CLASSIFIER
# ────────────────────────────────────────────────────────────────────────
class PromptClassifier:
    CAUSAL_RE  = re.compile(r"\bwhy\b|\bhow (?:do|does)\b|\bbecause\b", re.I)
    CONDIT_RE  = re.compile(r"\bif .* then\b", re.I)
    ARITH_RE   = re.compile(r"\d+\s*[\+\-\*/]\s*\d+", re.I)

    def classify(self, prompt: str) -> str:
        if self.ARITH_RE.search(prompt):  return "arithmetic"
        if self.CONDIT_RE.search(prompt): return "conditional"
        if self.CAUSAL_RE.search(prompt): return "causal"
        return "chat"

# ────────────────────────────────────────────────────────────────────────
# 2.   NHCE MEMORY + SOFT‑LOGIT Sampler             
# ────────────────────────────────────────────────────────────────────────

from core.sapphire_core_9_2 import (
    NHCE_Engine,
    ManualSampler,
    GPT2CustomTrainer,
    MemoryLoader,
    MemoryNode,
)
from ui.definitions_picker_9 import DefinitionsPicker
from ui.hyperparameter_dashboard_8 import HyperparameterDashboard
from ui.presets_panel_8 import PresetBankPanel
from core.system_banner import print_system_banner
from ui.hyperparams_launcher_panel import HyperparamsLauncherPanel
from ui.umb_wordcloud import UMBWordCloudFlower
# ────────────────────────────────────────────────────────────────────────
# 4.   CLI  – CHAT with a 'Presence'
# ────────────────────────────────────────────────────────────────────────


import random
from typing import List, Tuple, Sequence

# --------------------------------------------------


def main():
    global EOS_ID

   
    trainer = GPT2CustomTrainer()
    
    nhce  = NHCE_Engine(trainer.model, trainer.tok)
    gen   = ManualSampler(trainer.model, trainer.tok, nhce, _embedder = SentenceTransformer("all-MiniLM-L6-v2").to(DEVICE))
    clf   = PromptClassifier()
    loader = MemoryLoader()
    
    live_params = {
        "top_n": 13,
        "top_t": 7,
        "temp": 0.567,
        "top_p": 0.72,
        "top_k": 42,
        "repetition_penalty": 1.25,
        "max_forward_tokens": 77,
        "max_reply_sentences": 3,
        "bias_vec_gain": 0.1,
        "tau": 0.1456,
        "sigma": 0.1777,
        "n_sieve": 5,
        "lam": 0.5,
        "sieve_rank_mem": 2,
        "inference_mem": 1,
        "mem_cos_lex_blend": 0.5,
        "prompt_constr": "memory;prompt;memory;",
        "prompt_gain": 0.9,
        "mem_blend" : 0.6, 
        "mem_gain" : 0.8,
        "tail_blend" : 0.6,
        "tail_gain" : 0.5,
        "comp_precedence" : "memory",
        "comp_blend" : 0.3,
        "comp_gain" : 0.8,
        "master_gain" : 0.97,
        "memory_write": True,
    }

            
    def update_model_with_live_params(lp, gen):
        gen.top_n = lp.get("top_n")
        gen.top_t = lp.get("top_t")
        gen.temp = lp.get("temp")
        gen.top_p = lp.get("top_p")
        gen.top_k = lp.get("top_k")
        gen.repetition_penalty = lp.get("repetition_penalty")
        gen.max_forward_tokens = lp.get("max_forward_tokens")
        gen.max_reply_sentences = lp.get("max_reply_sentences")
        gen.bias_vec_gain = lp.get("bias_vec_gain")
        gen.tau = lp.get("tau")
        gen.sigma = lp.get("sigma")
        gen.n_sieve = lp.get("n_sieve")
        gen.lam = lp.get("lam")
        gen.sieve_rank_mem = lp.get("sieve_rank_mem")
        gen.inference_mem = lp.get("inference_mem")
        gen.mem.mem_cos_lex_blend = lp.get("mem_cos_lex_blend")
        gen.prompt_constr = lp.get("prompt_constr")
        gen.prompt_gain = lp.get("prompt_gain")
        gen.mem_blend = lp.get("mem_blend")
        gen.mem_gain = lp.get("mem_gain")
        gen.tail_blend = lp.get("tail_blend")
        gen.tail_gain = lp.get("tail_gain")
        gen.comp_precedence = lp.get("comp_precedence")
        gen.comp_blend = lp.get("comp_blend")
        gen.comp_gain = lp.get("comp_gain")
        gen.master_gain = lp.get("master_gain")
        gen.mem.memory_write = lp.get("memory_write")
        
    print("\n───────────────────────────────────────────────────────────────────────────────────")
    print("""                           ████                          
                          ░░███                          
          ███████  ██████  ░███   ██████  █████████████  
         ███░░███ ███░░███ ░███  ███░░███░░███░░███░░███ 
        ░███ ░███░███ ░███ ░███ ░███████  ░███ ░███ ░███ 
        ░███ ░███░███ ░███ ░███ ░███░░░   ░███ ░███ ░███ 
        ░░███████░░██████  █████░░██████  █████░███ █████
         ░░░░░███ ░░░░░░  ░░░░░  ░░░░░░  ░░░░░ ░░░ ░░░░░ 
         ███ ░███  * SAPPHIRE β CORE * GPT-2-mini *                        
        ░░██████                                          
         ░░░░░░ """)
    
    print("\n───────────────────────────────────────────────────────────────────────────────────")
    print("  🎲 GPT-2-mini 🗃️ microsoft/DialoGPT-small model 💎 Sapphire β inference engine ")
    print("───────────────────────────────────────────────────────────────────────────────────\n")
    
    print(" 🧩  rendering device: ", DEVICE)
    print(" ❓   questions? type \"config help\"") 
    # ----------------------------
    # Corpora load handler
    # ----------------------------
    def select_txt_file(
        prompt      = "[config] Select corpora .txt file (number) ➜ ",
        search_dir  = "./corpora",
    ) -> str:
        """
        Lists *.txt files in *search_dir* and lets the user pick one,
        or enter any valid path.  Keeps asking until a readable .txt file
        is supplied and returns its absolute path.
        """
        txts = sorted(glob.glob(os.path.join(search_dir, "*.txt")))
        if txts:
            print("🗂  Available corpora:\n")
            for i, fp in enumerate(txts, 1):
                print(f" {i}) {os.path.basename(fp)}")
            
        while True:
            choice = input(prompt).strip()

            # numeric pick from list -------------------------------
            if choice.isdigit() and txts and 0 <= int(choice) <= len(txts):
                if choice == "0":
                    choice = input("↳ path: ").strip()
                else:
                    choice = txts[int(choice) - 1]

            # validation -------------------------------------------
            if os.path.isfile(choice) and choice.lower().endswith(".txt"):
                return os.path.abspath(choice)

            print("🚫 Not a valid .txt file; try again.")

#####
    
# ─────── MAIN CLI loop
    
    live_params, msg = handle_settings_command("config load master_default", live_params)
    update_model_with_live_params(live_params, gen)    

    print(" 📓 for chat history type \"tail\"\n---")
       
    while True:
        print('---')
        usr = input(" 🗣  > ")
        
        if len(usr) == 0: #nextusr
            continue
        
        if usr.lower() == "exit":
            print("---\n 🤖 > Good bye!!\n---\n")
            break
        
        if usr.lower().strip() == "cloud":
            UMBWordCloudFlower().show_from_memory(nhce.memory, max_terms=450)
            continue  # Skip standard generation

        if usr.lower().strip() == "tail":
            print(" 📓 chat history\n---")
            for chatlog in nhce.tail_memories(n=4):
                print(chatlog)
                print("--")
            print("-----")
            continue  # Skip standard generation
            
        if usr.lower().strip() == "chats":

            loader.choose_memory_file()
            nhce.memory_file = loader.memory_file #update cross system
            nhce.memory = loader.load_memory()
            print(" 📓 chat history \n")
            for chatlog in nhce.tail_memories(n=4):
                print(chatlog, "\n--")
            continue  # Skip standard generation            
        
        if usr.lower().strip() == "umb":
            print(">> 💾 ", nhce.memory_file)
            continue  # Skip standard generation        
        
        if usr.lower().strip() == "write on":
            print(">> 💽 - UMB memory session write: True")
            live_params["memory_write"] = True
            update_model_with_live_params(live_params, gen)
            continue  # Skip standard generation            
        
        if usr.lower().strip() == "write off":
            print(">> 💽 - UMB memory session write: False")
            live_params["memory_write"] = False
            update_model_with_live_params(live_params, gen)
            continue  # Skip standard generation            
        
        if usr.lower().strip() == "memory":
            print(">> 💽 - UMB memory session write: ", nhce.memory_write)
            continue  # Skip standard generation            
        
        if usr.lower().strip()[0:6] == "create": 
            
            anchors = usr.split()
            
            if len(anchors) == 3:
                prompt_anchor = anchors[1]
                llm_anchor = anchors[2]
                    
                print("🔢 initializing UMB", flush=True )
                # seed with root‑identity prompt
                now = datetime.utcnow().isoformat()
                root = MemoryNode(now, prompt_anchor, llm_anchor, "identity", 0.95, 1.0, 1.0)
                nhce.memory = []  #wiped
                nhce.memory.append(root)
                nhce.memory_file = "./memory/emergence_UMB.json"
                with open(nhce.memory_file, "w") as fh:
                    json.dump([root.__dict__ for m in nhce.memory], fh, indent=2)
                
                print("[config] UMB initialized") 
                continue # skip gen
            print("[config] ❓ your did not provide init root pair of 'prompt inference'.\n[config] please append to 'create' command ") 
            continue #nextusr
            
        if usr.lower().strip() == "reload":
            with open("./memory/emergence_UMB.json") as fh:
                raw = json.load(fh)
            nhce.memory = [MemoryNode(**m) for m in raw]
            nhce.memory_file = "./memory/emergence_UMB.json"
            print("\n>> 🚅 ", "Default UMB reloaded")
            continue  # Skip standard regeneration

        if usr.lower().startswith("saveumb"):
            # ------------------------------------------------------------
            # Expected syntax:  saveumb <filename>
            # ------------------------------------------------------------
            parts = usr.split(maxsplit=1)
            if len(parts) < 2:
                print("[config] error: please provide a filename – usage: saveumb <filename>")
                continue  #  Skip generation and wait for next input

            file_candidate = parts[1].strip()

            # ---- validation -------------------------------------------------------
            if any(ch in file_candidate for ch in ".\\/"):
                print("[config] error: filename must not contain '.', '/' or '\\'.")
                continue

            # Optionally append an extension so on‑disk files stay consistent
            file_target = f"emergence_UMB_{file_candidate}.json"

            # Update NHCE engine so future loads know which file we just wrote
            nhce.memory_file = 'memory/' + file_target

            # Persist the current memory bank
            nhce._persist(nhce.memory)

            print(f"\n>> 🚅  UMB saved as : {file_target}")
            continue  # Skip standard regeneration

        if usr.lower().startswith("makedef"):
            # ------------------------------------------------------------
            # Expected syntax:  saveumb <filename>
            # ------------------------------------------------------------


            # Optionally append an extension so on‑disk files stay consistent
            file_target = "emergence_UMB.json"

            # Update NHCE engine so future loads know which file we just wrote
            nhce.memory_file = 'memory/' + file_target

            # Persist the current memory bank
            nhce._persist(nhce.memory)

            print(f"\n>> 🚅  UMB saved as : {file_target}")
            continue  # Skip standard regeneration

        if usr.lower().strip() == "train":
            corpus = select_txt_file()
            epochs = int(input("[trainer] number of epochs : "))
            trainer.ckpt_name = input("[trainer] output model name : ")
            print("\n")
            trainer.finetune(corpus, epochs)
            continue  # Skip standard regeneration

        if usr.lower().strip() == "model":
            trainer.maybe_load_latest()
            continue  # Skip standard regeneration

        if usr.lower().strip() == "hyper":
            from ui.hyperparams_isolated import run_hyperparams_isolated

            updated = run_hyperparams_isolated(
                live_params,
                module_name="hyperparameter_dashboard_8"  # where HyperparameterDashboard lives
                # timeout=900,  # optional: seconds
            )

            if updated:
                # merge and apply to gen.* via your existing callback
                live_params.update(updated)
                update_model_with_live_params(live_params, gen)
                print("[hyper] applied updated parameters.")
            else:
                print("[hyper] no changes returned (window closed without Apply).")

            continue  # Skip standard regeneration




        if usr.lower().strip() == "turns":
            print("🗣 ♟️🤖 chat turns: ", len(nhce.memory), "\n---")
            
            continue  # Skip standard regeneration

        if usr.lower().strip() == "sys":
            print("---")
            print_system_banner(model=trainer.model, umb_path=nhce.memory_file)
            print("---")
            continue  # Skip standard regeneration


        if usr.lower().strip()[:6] == "preset":
            name = PresetBankPanel(params_path="presets/sapphire_params.json").run()
            if name:
                print("[presets] selected:", name)

                live_params, msg = handle_settings_command("config load " + name, live_params)
                print(msg)
                update_model_with_live_params(live_params, gen)
                continue
            else:
                print("[presets] canceled")
            continue  # Skip standard regeneration

            


        if usr.lower().strip() == "defs":
 
            chosen = DefinitionsPicker(path="presets").run()
            if not chosen:
                print("[defs] canceled or no selection")
                continue # return to CLI loop
            
            print("[defs 📚] selected:", chosen)
            reply = gen.generate(chosen, write_memory=nhce.memory_write)    
            print("\n 🤖📚  > ", reply.strip(), "\n--")

            continue #skip std gen
        
        
        if usr.lower().startswith("config"):
                live_params, msg = handle_settings_command(usr, live_params)
                print(msg)
                update_model_with_live_params(live_params, gen)
                continue 

        else:
            reply = gen.generate(usr, write_memory=nhce.memory_write)

        print("\n 🤖  > ", reply.strip(), "\n--")

if __name__ == "__main__":
    main()