File size: 11,631 Bytes

# Copyright (c) 2025 CMS Manhattan
# All rights reserved.
# Author: Konstantin Vladimirovich Grabko
# Email: grabko@cmsmanhattan.com
# Phone: +1(516)777-0945
#
# This program is free software: you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation, version 3 of the License.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU General Public License
# along with this program.  If not, see <https://www.gnu.org/licenses/>.
#
# Additional terms:
# Any commercial use or distribution of this software or derivative works
# requires explicit written permission from the copyright holder.

import os
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
from transformers import GPT2TokenizerFast
from tqdm import tqdm
import shutil
import math
from pathlib import Path
import re
from typing import Optional, List, Tuple

# from gpt_pytorch_33b import GPTPyTorch # REMOVED: Now loaded via JIT/TorchScript!

# ============================= SETTINGS =============================
TRAIN_SEQ_LEN = 256 # Context length
BATCH_SIZE = 12
EPOCHS = 50
LEARNING_RATE = 6e-6
WEIGHT_DECAY = 0.01
GRAD_CLIP = 1.0
KEEP_LAST_EPOCHS = 3
VAL_SPLIT_RATIO = 0.05

# === MODEL PATHS (ADAPTED FOR JIT) ===
# NOTE: Model must be saved by torch.jit.trace() or torch.jit.script()
BASE_MODEL_PATH = Path("models/JiRack_H12_L6_V50257_D768_MSL8192_FF768x4.script.pt")
LAST_TRAINED_PATH = Path("models/JiRack_last_H12_L6_V50257_D768_MSL8192_FF768x4.script.pt")
BACKUP_DIR = Path("models/backups")
BACKUP_DIR.mkdir(exist_ok=True)

# === AUTOCLEAN DATASET (Improved reliability) ===
RAW_PATH = Path("datasets/dialogues_text.txt")
CLEAN_PATH = Path("datasets/dialogues_text_clean.txt")

# Flag to control cleaning process
force_clean = False
if not CLEAN_PATH.exists():
    print("Cleaned dataset not found. Performing initial cleaning...")
    force_clean = True
else:
    try:
        if RAW_PATH.stat().st_mtime > CLEAN_PATH.stat().st_mtime:
            print("Detected changes in the raw dataset. Re-cleaning...")
            force_clean = True
        else:
            print(f"Using existing cleaned dataset → {CLEAN_PATH}")
    except FileNotFoundError:
        print("File system synchronization error. Performing re-cleaning for safety...")
        force_clean = True

if force_clean:
    if not RAW_PATH.exists():
        raise FileNotFoundError(f"ERROR: Source file {RAW_PATH} not found. Check the path.")

    print("Cleaning up the dataset from garbage (wrong separators, extra spaces)...")
    text = RAW_PATH.read_text(encoding="utf-8")

    # 3. General cleanup: remove multiple spaces and spaces around newlines
    text = re.sub(r' {2,}', ' ', text)
    text = text.replace(" \n", "\n").replace("\n ", "\n")

    CLEAN_PATH.write_text(text, encoding="utf-8")
    print(f"Dataset successfully cleaned and saved → {CLEAN_PATH}")

DATASET_PATH = CLEAN_PATH

OUTPUT_DIR = Path("build/fine_tuning_output")
MODEL_SAVE_NAME = "gpt_finetuned.script.pt" # Changed to JIT format

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")

# ============================= DATASET =============================
class TextDataset(Dataset):
    def __init__(self, text_file, seq_len=TRAIN_SEQ_LEN, tokenizer_name="gpt2", split_type='train', val_ratio=VAL_SPLIT_RATIO):
        self.seq_len = seq_len
        self.tokenizer = GPT2TokenizerFast.from_pretrained(tokenizer_name)
        self.tokenizer.pad_token = self.tokenizer.eos_token
        self.split_type = split_type

        print(f"Loading text from {text_file} for {split_type} split...")
        text = Path(text_file).read_text(encoding="utf-8")
        tokens = self.tokenizer.encode(text)

        if len(tokens) < seq_len * 2:
            raise ValueError("Text too short!")

        all_inputs = []
        all_labels = []

        for i in range(0, len(tokens) - seq_len, seq_len):
            all_inputs.append(tokens[i:i + seq_len])
            all_labels.append(tokens[i + 1:i + seq_len + 1])

        total_sequences = len(all_inputs)
        val_size = int(total_sequences * val_ratio)
        train_size = total_sequences - val_size

        if self.split_type == 'train':
            self.inputs = all_inputs[:train_size]
            self.labels = all_labels[:train_size]
        elif self.split_type == 'val':
            self.inputs = all_inputs[train_size:]
            self.labels = all_labels[train_size:]
        else:
            raise ValueError("Invalid split_type. Must be 'train' or 'val'.")

        print(f"Created {len(self.inputs):,} sequences for {self.split_type} split.")

    def __len__(self):
        return len(self.inputs)

    def __getitem__(self, idx):
        return (torch.tensor(self.inputs[idx], dtype=torch.long),
                torch.tensor(self.labels[idx], dtype=torch.long))

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

# ============================= EVALUATION (VALIDATION) =============================
def get_logits_from_model(model, inputs):
    """
    Adapted model invocation handling a possible output of (logits, new_kv)
    or just logits for JIT models.
    """
    try:
        # Try to call as the original model (Logits, KV-cache)
        logits, _ = model(inputs)
    except Exception:
        # If JIT model returns only Logits (most likely)
        logits = model(inputs)
    return logits


def evaluate(model, dataloader, criterion, device):
    """Evaluates the model on the validation dataset."""
    model.eval()
    total_loss = 0.0

    with torch.no_grad():
        for inputs, targets in dataloader:
            inputs, targets = inputs.to(device), targets.to(device)

            logits = get_logits_from_model(model, inputs)
            loss = criterion(logits.view(-1, logits.size(-1)), targets.view(-1))
            total_loss += loss.item()

    avg_loss = total_loss / len(dataloader)
    model.train()
    return avg_loss

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

# ============================= CLEANUP OLD EPOCHS =============================
def cleanup_old_epochs(keep_last=KEEP_LAST_EPOCHS):
    epochs = sorted([p for p in OUTPUT_DIR.glob("epoch*") if p.is_dir()],
                    key=lambda x: int(x.name.replace("epoch", "")))
    for old in epochs[:-keep_last]:
        if old.exists():
            shutil.rmtree(old)
            print(f"Old epoch deleted: {old.name}")

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

# ============================= TRAINING =============================
def train():
    OUTPUT_DIR.mkdir(parents=True, exist_ok=True)

    print("Loading model...")
    model = None

    # === SMART JIT MODEL LOADING ===
    if LAST_TRAINED_PATH.exists():
        print(f"Continuing training from last JIT model: {LAST_TRAINED_PATH}")
        model = torch.jit.load(LAST_TRAINED_PATH, map_location=device)
    elif BASE_MODEL_PATH.exists():
        print(f"Starting from base JIT model: {BASE_MODEL_PATH}")
        model = torch.jit.load(BASE_MODEL_PATH, map_location=device)
    else:
        print("ERROR: JIT model not found. Cannot start training without source code or JIT file.")
        return

    model.train()

    # Create datasets and dataloaders (Train and Validation)
    train_dataset = TextDataset(DATASET_PATH, seq_len=TRAIN_SEQ_LEN, split_type='train', val_ratio=VAL_SPLIT_RATIO)
    val_dataset = TextDataset(DATASET_PATH, seq_len=TRAIN_SEQ_LEN, split_type='val', val_ratio=VAL_SPLIT_RATIO)

    train_dataloader = DataLoader(train_dataset, batch_size=BATCH_SIZE, shuffle=True, drop_last=True)
    val_dataloader = DataLoader(val_dataset, batch_size=BATCH_SIZE, shuffle=False, drop_last=True)

    # NOTE: .parameters() should work for JIT models if saved with parameters.
    optimizer = optim.AdamW(model.parameters(), lr=LEARNING_RATE, weight_decay=WEIGHT_DECAY)
    criterion = nn.CrossEntropyLoss()

    total_steps = len(train_dataloader) * EPOCHS
    print(f"\n=== BEGINNING LONG-TERM TRAINING ===")
    print(f"Epochs: {EPOCHS} | Steps (Train): {total_steps} | Examples (Train): {len(train_dataset)}")

    global_step = 0
    for epoch in range(1, EPOCHS + 1):
        print(f"\n--- Epoch {epoch}/{EPOCHS} ---")
        epoch_loss = 0.0

        # ====================== TRAINING LOOP ======================
        with tqdm(train_dataloader, desc=f"Epoch {epoch} [TRAIN]", leave=False) as pbar:
            for inputs, targets in pbar:
                inputs, targets = inputs.to(device), targets.to(device)

                optimizer.zero_grad()

                # ADAPTED MODEL CALL
                logits = get_logits_from_model(model, inputs)

                loss = criterion(logits.view(-1, logits.size(-1)), targets.view(-1))
                loss.backward()
                torch.nn.utils.clip_grad_norm_(model.parameters(), GRAD_CLIP)
                optimizer.step()

                loss_val = loss.item()
                epoch_loss += loss_val
                global_step += 1

                pbar.set_postfix({
                    "loss": f"{loss_val:.3f}",
                    "ppl": f"{math.exp(min(loss_val, 10)):.1f}",
                    "step": f"{global_step}/{total_steps}"
                })

        avg_train_loss = epoch_loss / len(train_dataloader)
        print(f"  [TRAIN] Average loss: {avg_train_loss:.3f} | PPL: {math.exp(avg_train_loss):.1f}")

        # ====================== VALIDATION LOOP ======================
        print("  [VALIDATION] Starting evaluation...")
        val_loss = evaluate(model, val_dataloader, criterion, device)
        print(f"  [VALIDATION] Average loss: {val_loss:.3f} | PPL: {math.exp(val_loss):.1f}")

        # ====================== SAVING MODEL (ADAPTED) ======================
        epoch_dir = OUTPUT_DIR / f"epoch{epoch}"
        epoch_dir.mkdir(exist_ok=True)
        # Save JIT model: use .save() instead of torch.save(state_dict)
        model.save(epoch_dir / MODEL_SAVE_NAME)
        print(f"Model saved: {epoch_dir / MODEL_SAVE_NAME}")
        cleanup_old_epochs()

    # === FINAL SAVE ===
    final_dir = OUTPUT_DIR / "final"
    final_dir.mkdir(exist_ok=True)
    model.save(final_dir / MODEL_SAVE_NAME) # Save final JIT model
    train_dataset.tokenizer.save_pretrained(final_dir)

    # === AUTO-SAVE LAST MODEL + BACKUP (ADAPTED) ===
    if LAST_TRAINED_PATH.exists():
        backup_path = BACKUP_DIR / f"gpt_last_trained_backup_{int(os.path.getmtime(LAST_TRAINED_PATH))}.script.pt"
        shutil.copy(LAST_TRAINED_PATH, backup_path)
        print(f"Backup of previous model created → {backup_path.name}")

    shutil.copy(final_dir / MODEL_SAVE_NAME, LAST_TRAINED_PATH)
    print(f"Last trained model saved → {LAST_TRAINED_PATH}")

    print(f"\nTRAINING COMPLETED! Model ready:")
    print(f"   • For chat: {final_dir / MODEL_SAVE_NAME}")
    print(f"   • For further fine-tuning: {LAST_TRAINED_PATH}")

if __name__ == "__main__":
    if not RAW_PATH.exists():
        print(f"ERROR: No file {RAW_PATH}")
        print("Put your text into datasets/dialogues_text.txt")
    else:
        train()