Wrinkle/09_standalone_model/benchmark_optimizers.py

#!/usr/bin/env python3
"""Benchmark: AdamW vs Adafactor vs Muon on WrinkleBrane.

All runs use FP32 baseline config (same as the 10K step run).
500 steps each, same data, same seed for model init.

Optimizers:
  1. AdamW   — current baseline (lr=3e-4, betas=(0.9, 0.95), wd=0.01)
  2. Adafactor — memory-efficient adaptive, no external scheduler (PyTorch 2.8)
  3. Muon   — Momentum + Newton-Schulz orthogonalization, lr=0.05, clip=2.0

DESIGNED FOR JUPYTERLAB TERMINAL — output is tee'd to both stdout and a
timestamped log file under logs/ (or --log_dir) so browser crashes don't
lose results. Run as:

    cd /data/WrinkleBrane-Research/09_standalone_model
    OMP_NUM_THREADS=8 MKL_NUM_THREADS=8 PYTHONPATH=src \\
        nohup python3 -u benchmark_optimizers.py &
    tail -f logs/benchmark_<timestamp>.log
"""

from __future__ import annotations

import copy
import gc
import math
import os
import subprocess
import sys
import time
from dataclasses import dataclass
from typing import Dict, List, Optional, Tuple

import argparse

import torch
from torch import nn, Tensor

sys.path.insert(0, os.path.join(os.path.dirname(__file__), "src"))

from wrinklebrane.standalone_model import WrinkleBraneConfig, WrinkleBraneModel
from wrinklebrane.data import load_train_val, VOCAB_SIZE


# ============================================================================
# Tee: write to stdout AND a log file simultaneously
# ============================================================================

class Tee:
    """Duplicate stdout writes to a log file.

    Replaces sys.stdout so all print() calls go to both the terminal
    and a persistent log file — survives browser/VSCode crashes.
    """
    def __init__(self, log_path: str):
        self.terminal = sys.__stdout__
        os.makedirs(os.path.dirname(log_path), exist_ok=True)
        self.log = open(log_path, "w", buffering=1)  # line-buffered
        print(f"  Logging to: {log_path}", file=self.terminal)

    def write(self, message):
        self.terminal.write(message)
        self.log.write(message)

    def flush(self):
        self.terminal.flush()
        self.log.flush()

    def close(self):
        self.log.close()
        sys.stdout = self.terminal


# ============================================================================
# Configuration
# ============================================================================

BENCHMARK_STEPS = 500
BATCH_SIZE = 16
SEQ_LEN = 128
WARMUP = 50
LOG_EVERY = 50
EVAL_EVERY = 100
SEED = 42


def make_config() -> WrinkleBraneConfig:
    """Same optimal config as the 10K training run."""
    return WrinkleBraneConfig(
        vocab_size=VOCAB_SIZE,
        d_model=128,
        n_layers=6,
        n_heads=4,
        L=16,
        K=SEQ_LEN,
        code_init="hadamard",
        learnable_codes=True,
        temperature=0.5,
        ffn_expansion=4,
        use_gated_ffn=True,
        max_seq_len=SEQ_LEN,
        dropout=0.1,
        ortho_lambda=0.01,
        persistence_lambda=0.99,
        weight_tying=True,
    )


# ============================================================================
# Muon Optimizer — Momentum with Newton-Schulz Orthogonalization
# ============================================================================

class Muon(torch.optim.Optimizer):
    """Muon: Momentum + Orthogonalized Updates via Newton-Schulz.

    For matrix-shaped parameters (2D), Muon orthogonalizes the momentum
    buffer using Newton-Schulz iterations before applying the update.
    This naturally preserves orthogonal structure in weight matrices.

    For non-matrix parameters (1D scalars, biases, embeddings), falls
    back to standard momentum SGD.

    Particularly suited for WrinkleBrane because:
    - Codebook parameters benefit from orthogonality preservation
    - Projection matrices (W_v, W_q, W_o) stay well-conditioned
    - The Newton-Schulz step acts like a natural preconditioner

    Reference: Keller Jordan et al., "Muon: An optimizer for hidden layers"

    Parameters
    ----------
    params : iterable
        Model parameters.
    lr : float
        Learning rate (default: 0.02, typically higher than Adam).
    momentum : float
        Momentum coefficient (default: 0.95).
    ns_steps : int
        Number of Newton-Schulz orthogonalization iterations (default: 5).
    weight_decay : float
        Decoupled weight decay (default: 0.0).
    """

    def __init__(
        self,
        params,
        lr: float = 0.02,
        momentum: float = 0.95,
        ns_steps: int = 5,
        weight_decay: float = 0.0,
    ):
        defaults = dict(
            lr=lr,
            momentum=momentum,
            ns_steps=ns_steps,
            weight_decay=weight_decay,
        )
        super().__init__(params, defaults)

    @staticmethod
    def newton_schulz_orthogonalize(M: Tensor, steps: int = 5) -> Tensor:
        """Orthogonalize matrix M using Newton-Schulz iteration.

        Computes the polar factor of M: the nearest orthogonal matrix.
        Uses the iteration: X_{k+1} = 1.5 * X_k - 0.5 * X_k @ X_k^T @ X_k

        For efficiency, operates on the smaller dimension.

        Parameters
        ----------
        M : Tensor [m, n]
        steps : int
            Number of NS iterations (5 is usually sufficient).

        Returns
        -------
        Tensor [m, n]
            Orthogonalized matrix.
        """
        m, n = M.shape
        transpose = False

        # Work with the smaller dimension for efficiency
        if m < n:
            M = M.T
            transpose = True

        # Normalize to unit spectral norm (approximate)
        X = M / (M.norm() + 1e-7)

        # Newton-Schulz iterations: X = 1.5*X - 0.5*X@X^T@X
        for _ in range(steps):
            A = X @ X.T
            X = 1.5 * X - 0.5 * A @ X

        if transpose:
            X = X.T

        return X

    @torch.no_grad()
    def step(self, closure=None):
        """Perform a single optimization step."""
        loss = None
        if closure is not None:
            with torch.enable_grad():
                loss = closure()

        for group in self.param_groups:
            lr = group["lr"]
            momentum = group["momentum"]
            ns_steps = group["ns_steps"]
            wd = group["weight_decay"]

            for p in group["params"]:
                if p.grad is None:
                    continue

                grad = p.grad

                # Decoupled weight decay
                if wd > 0:
                    p.mul_(1 - lr * wd)

                # Get or create momentum buffer
                state = self.state[p]
                if "momentum_buffer" not in state:
                    state["momentum_buffer"] = torch.zeros_like(p)

                buf = state["momentum_buffer"]
                buf.mul_(momentum).add_(grad)

                # For 2D parameters: apply Newton-Schulz orthogonalization
                if p.dim() == 2 and min(p.shape) > 1:
                    update = self.newton_schulz_orthogonalize(buf, steps=ns_steps)
                    # Scale update to match the gradient magnitude
                    update = update * (buf.norm() / (update.norm() + 1e-7))
                else:
                    # For 1D/scalar params: standard momentum
                    update = buf

                p.add_(update, alpha=-lr)

        return loss


# ============================================================================
# Training infrastructure
# ============================================================================

@dataclass
class BenchmarkResult:
    name: str
    steps: int
    total_time: float
    losses: List[float]
    eval_losses: List[Tuple[int, float]]
    tok_per_sec: List[float]
    grad_norms: List[float]

    @property
    def avg_tok_per_sec(self) -> float:
        return sum(self.tok_per_sec) / len(self.tok_per_sec) if self.tok_per_sec else 0

    @property
    def final_loss(self) -> float:
        return self.losses[-1] if self.losses else float("nan")

    @property
    def final_eval_loss(self) -> float:
        return self.eval_losses[-1][1] if self.eval_losses else float("nan")


def evaluate(model, val_corpus, seq_len, batch_size=16, n_batches=10, device=None):
    """Evaluate on validation data."""
    if device is None:
        device = next(model.parameters()).device
    model.eval()
    total_loss = 0.0
    total_tokens = 0
    with torch.no_grad():
        for _ in range(n_batches):
            inp, tgt = val_corpus.get_batch(batch_size, seq_len)
            inp, tgt = inp.to(device), tgt.to(device)
            logits = model(inp)
            B, T, V = logits.shape
            loss = nn.functional.cross_entropy(
                logits.reshape(B * T, V), tgt.reshape(B * T),
                reduction="sum",
            )
            total_loss += loss.item()
            total_tokens += B * T
    return total_loss / total_tokens


def run_benchmark(
    name: str,
    model: WrinkleBraneModel,
    config: WrinkleBraneConfig,
    optimizer: torch.optim.Optimizer,
    scheduler,
    train_corpus,
    val_corpus,
    steps: int = BENCHMARK_STEPS,
    grad_clip: float = 1.0,
    device: torch.device = None,
) -> BenchmarkResult:
    """Run a training benchmark for the given number of steps."""
    if device is None:
        device = next(model.parameters()).device

    print(f"\n{'='*70}")
    print(f"  BENCHMARK: {name}")
    print(f"  {steps} steps, batch_size={BATCH_SIZE}, seq_len={SEQ_LEN}")
    print(f"{'='*70}")

    param_count = sum(p.numel() for p in model.parameters())
    print(f"  Parameters: {param_count:,}")

    # Initial eval
    init_eval = evaluate(model, val_corpus, SEQ_LEN)
    print(f"  Initial eval loss: {init_eval:.4f} (PPL {math.exp(min(init_eval, 20)):.2f})")

    # Tracking
    losses = []
    eval_losses = [(0, init_eval)]
    tok_per_sec_list = []
    grad_norms_list = []

    running_loss = 0.0
    running_tokens = 0
    interval_start = time.time()
    total_start = time.time()

    print(f"\n  Training started at {time.strftime('%H:%M:%S')}")
    print(f"  {'─'*64}")

    for step in range(1, steps + 1):
        model.train()
        optimizer.zero_grad()

        inp, tgt = train_corpus.get_batch(BATCH_SIZE, SEQ_LEN)
        inp, tgt = inp.to(device), tgt.to(device)

        # Forward
        logits = model(inp)
        B, T, V = logits.shape
        task_loss = nn.functional.cross_entropy(
            logits.reshape(B * T, V), tgt.reshape(B * T),
        )

        # Ortho regularization
        ortho = config.ortho_lambda * model.ortho_loss()
        total_loss = task_loss + ortho

        # Backward
        total_loss.backward()
        grad_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), grad_clip)
        optimizer.step()
        if scheduler is not None:
            scheduler.step()

        running_loss += task_loss.item() * B * T
        running_tokens += B * T
        grad_norms_list.append(grad_norm.item())

        # Log
        if step % LOG_EVERY == 0:
            now = time.time()
            elapsed = now - interval_start
            avg_loss = running_loss / running_tokens
            tps = running_tokens / elapsed
            losses.append(avg_loss)
            tok_per_sec_list.append(tps)

            # Get current LR
            if scheduler is not None:
                lr = scheduler.get_last_lr()[0]
            else:
                lr = optimizer.param_groups[0]["lr"]

            print(f"  step {step:4d}/{steps} | "
                  f"loss={avg_loss:.4f} ppl={math.exp(min(avg_loss, 20)):7.2f} | "
                  f"lr={lr:.2e} gnorm={grad_norm:.2f} | "
                  f"{tps:6.0f} tok/s",
                  flush=True)

            running_loss = 0.0
            running_tokens = 0
            interval_start = time.time()

        # Eval
        if step % EVAL_EVERY == 0:
            val_loss = evaluate(model, val_corpus, SEQ_LEN)
            eval_losses.append((step, val_loss))
            print(f"  >>> EVAL step {step}: loss={val_loss:.4f}, "
                  f"ppl={math.exp(min(val_loss, 20)):.2f}", flush=True)

    total_time = time.time() - total_start

    # Final eval
    final_eval = evaluate(model, val_corpus, SEQ_LEN)
    eval_losses.append((steps, final_eval))

    print(f"\n  {'─'*64}")
    print(f"  DONE: {name}")
    print(f"  Total time: {total_time:.1f}s ({total_time/60:.1f} min)")
    print(f"  Final train loss: {losses[-1]:.4f}" if losses else "  No losses recorded")
    print(f"  Final eval loss: {final_eval:.4f} (PPL {math.exp(min(final_eval, 20)):.2f})")
    print(f"  Avg throughput: {sum(tok_per_sec_list)/len(tok_per_sec_list):.0f} tok/s" if tok_per_sec_list else "")

    return BenchmarkResult(
        name=name,
        steps=steps,
        total_time=total_time,
        losses=losses,
        eval_losses=eval_losses,
        tok_per_sec=tok_per_sec_list,
        grad_norms=grad_norms_list,
    )


def make_cosine_scheduler(optimizer, warmup, total_steps):
    """Cosine LR schedule with linear warmup."""
    def lr_schedule(step):
        if step < warmup:
            return step / warmup
        progress = (step - warmup) / max(total_steps - warmup, 1)
        return 0.5 * (1.0 + math.cos(math.pi * progress))
    return torch.optim.lr_scheduler.LambdaLR(optimizer, lr_schedule)


# ============================================================================
# Main benchmark
# ============================================================================

def main():
    parser = argparse.ArgumentParser(description="WrinkleBrane optimizer benchmark")
    parser.add_argument(
        "--log_dir",
        default=os.path.join(os.path.dirname(__file__), "logs"),
        help="Directory for log files (default: ./logs/)",
    )
    parser.add_argument(
        "--steps", type=int, default=BENCHMARK_STEPS,
        help=f"Training steps per optimizer (default: {BENCHMARK_STEPS})",
    )
    parser.add_argument(
        "--device", type=str, default=None,
        help="Device: 'cuda', 'cuda:0', 'cpu'. Auto-detects if not set.",
    )
    args = parser.parse_args()

    # Redirect stdout to Tee (stdout + log file) — survives browser crashes
    timestamp = time.strftime("%Y%m%d_%H%M%S")
    log_path = os.path.join(args.log_dir, f"benchmark_optimizers_{timestamp}.log")
    tee = Tee(log_path)
    sys.stdout = tee

    # Device setup
    if args.device:
        device = torch.device(args.device)
    else:
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    print("=" * 70)
    print("  WrinkleBrane Optimizer Benchmark")
    print("  AdamW vs Adafactor vs Muon — 500 steps each (FP32) [v2: fixed configs]")
    print("=" * 70)
    print()

    # System info
    print("  Hardware:")
    cpu_model = subprocess.check_output(
        "grep 'model name' /proc/cpuinfo | head -1 | cut -d: -f2",
        shell=True, text=True,
    ).strip()
    print(f"    CPU: {cpu_model}")
    print(f"    Device: {device}")
    if device.type == "cuda":
        print(f"    GPU: {torch.cuda.get_device_name(device)}")
        print(f"    VRAM: {torch.cuda.get_device_properties(device).total_memory / 1e9:.1f} GB")

    # Set thread count properly (CPU only; on GPU, threads matter less)
    n_threads = int(subprocess.check_output("nproc", text=True).strip())
    torch.set_num_threads(n_threads)
    os.environ["OMP_NUM_THREADS"] = str(n_threads)
    os.environ["MKL_NUM_THREADS"] = str(n_threads)
    print(f"    Torch threads: {n_threads}")
    print(f"    PyTorch: {torch.__version__}")
    print()

    # Load data once
    print("  Loading data...")
    train_corpus, val_corpus = load_train_val(
        "/data/WrinkleBrane-Research/raw"
    )
    print()

    results = []

    # ────────────────────────────────────────────────────────────────────
    # Benchmark 1: AdamW (baseline)
    # ────────────────────────────────────────────────────────────────────
    print("  ┌─────────────────────────────────────┐")
    print("  │  1/3: AdamW (baseline)              │")
    print("  └─────────────────────────────────────┘")

    torch.manual_seed(SEED)
    config = make_config()
    model = WrinkleBraneModel(config).to(device)

    optimizer = torch.optim.AdamW(
        model.parameters(),
        lr=3e-4,
        weight_decay=0.01,
        betas=(0.9, 0.95),
    )
    scheduler = make_cosine_scheduler(optimizer, WARMUP, BENCHMARK_STEPS)

    result = run_benchmark(
        name="AdamW (lr=3e-4, β=(0.9,0.95), wd=0.01)",
        model=model,
        config=config,
        optimizer=optimizer,
        scheduler=scheduler,
        train_corpus=train_corpus,
        val_corpus=val_corpus,
        device=device,
    )
    results.append(result)
    del model, optimizer, scheduler
    gc.collect()

    # ────────────────────────────────────────────────────────────────────
    # Benchmark 2: Adafactor
    # ────────────────────────────────────────────────────────────────────
    print("\n  ┌─────────────────────────────────────┐")
    print("  │  2/3: Adafactor                     │")
    print("  └─────────────────────────────────────┘")

    torch.manual_seed(SEED)
    config = make_config()
    model = WrinkleBraneModel(config).to(device)

    # Adafactor config (v2 — fixed):
    #   - lr=1e-3 constant, NO external scheduler
    #   - Adafactor manages its own adaptive second moment via beta2_decay
    #   - Layering a cosine schedule on top (v1) was double-scheduling it,
    #     fighting the internal adaptive rate and preventing convergence
    #   - eps=(None, 1e-3): no row-factor (None), column epsilon=1e-3
    #   - weight_decay=0.01 (match AdamW)
    optimizer = torch.optim.Adafactor(
        model.parameters(),
        lr=1e-3,
        beta2_decay=-0.8,
        eps=(None, 1e-3),
        weight_decay=0.01,
    )
    scheduler = None  # Adafactor manages its own schedule internally

    result = run_benchmark(
        name="Adafactor (lr=1e-3, no sched, β2d=-0.8, wd=0.01)",
        model=model,
        config=config,
        optimizer=optimizer,
        scheduler=scheduler,
        train_corpus=train_corpus,
        val_corpus=val_corpus,
        device=device,
    )
    results.append(result)
    del model, optimizer, scheduler
    gc.collect()

    # ────────────────────────────────────────────────────────────────────
    # Benchmark 3: Muon
    # ────────────────────────────────────────────────────────────────────
    print("\n  ┌─────────────────────────────────────┐")
    print("  │  3/3: Muon                          │")
    print("  └─────────────────────────────────────┘")

    torch.manual_seed(SEED)
    config = make_config()
    model = WrinkleBraneModel(config).to(device)

    # Muon config (v2 — fixed):
    #   - lr=0.05 (v1's 0.02 plateaued hard after step 150; orthogonalized
    #     updates are naturally bounded so Muon can push much harder)
    #   - momentum=0.95 (standard heavy momentum)
    #   - ns_steps=5 (Newton-Schulz iterations, standard)
    #   - weight_decay=0.01 (decoupled, match others)
    #   - grad_clip=2.0 (v1's avg grad norm was 0.474 — well under 1.0,
    #     so the clip was never helping; raise ceiling to let Muon breathe)
    optimizer = Muon(
        model.parameters(),
        lr=0.05,
        momentum=0.95,
        ns_steps=5,
        weight_decay=0.01,
    )
    scheduler = make_cosine_scheduler(optimizer, WARMUP, BENCHMARK_STEPS)

    result = run_benchmark(
        name="Muon (lr=0.05, mom=0.95, ns=5, wd=0.01, clip=2.0)",
        model=model,
        config=config,
        optimizer=optimizer,
        scheduler=scheduler,
        train_corpus=train_corpus,
        val_corpus=val_corpus,
        grad_clip=2.0,
        device=device,
    )
    results.append(result)
    del model, optimizer, scheduler
    gc.collect()

    # ────────────────────────────────────────────────────────────────────
    # Results comparison
    # ────────────────────────────────────────────────────────────────────

    print("\n\n")
    print("=" * 70)
    print("  OPTIMIZER BENCHMARK RESULTS")
    print("=" * 70)
    print()

    # Summary table
    header = f"  {'Metric':<28}"
    sep = f"  {'─'*28}"
    for r in results:
        short_name = r.name.split("(")[0].strip()
        header += f" │ {short_name:<26}"
        sep += f" │{'─'*26}"
    print(header)
    print(sep)

    # Total time
    row = f"  {'Total time':<28}"
    for r in results:
        row += f" │ {r.total_time:>7.1f}s ({r.total_time/60:>4.1f}m)       "
    print(row)

    # Time per step
    row = f"  {'Time per step':<28}"
    for r in results:
        row += f" │ {r.total_time/r.steps*1000:>7.1f} ms/step         "
    print(row)

    # Throughput
    row = f"  {'Avg throughput':<28}"
    for r in results:
        row += f" │ {r.avg_tok_per_sec:>7.0f} tok/s            "
    print(row)

    # Final train loss
    row = f"  {'Final train loss':<28}"
    for r in results:
        row += f" │ {r.final_loss:>7.4f}                  "
    print(row)

    # Final eval loss
    row = f"  {'Final eval loss':<28}"
    for r in results:
        row += f" │ {r.final_eval_loss:>7.4f}                  "
    print(row)

    # Final eval PPL
    row = f"  {'Final eval PPL':<28}"
    for r in results:
        ppl = math.exp(min(r.final_eval_loss, 20))
        row += f" │ {ppl:>7.2f}                  "
    print(row)

    # Avg gradient norm
    row = f"  {'Avg gradient norm':<28}"
    for r in results:
        avg_gn = sum(r.grad_norms) / len(r.grad_norms) if r.grad_norms else 0
        row += f" │ {avg_gn:>7.3f}                  "
    print(row)

    # Peak gradient norm
    row = f"  {'Peak gradient norm':<28}"
    for r in results:
        peak_gn = max(r.grad_norms) if r.grad_norms else 0
        row += f" │ {peak_gn:>7.3f}                  "
    print(row)

    # Best eval loss (with step)
    row = f"  {'Best eval loss (step)':<28}"
    for r in results:
        best_step, best_loss = min(r.eval_losses, key=lambda x: x[1])
        row += f" │ {best_loss:.4f} @ {best_step:<4}           "
    print(row)

    # Relative comparisons
    print(f"\n{sep}")
    base_eval = results[0].final_eval_loss
    base_time = results[0].total_time

    row = f"  {'Eval loss vs AdamW':<28}"
    for r in results:
        delta = r.final_eval_loss - base_eval
        pct = (delta / base_eval) * 100
        sign = "+" if delta >= 0 else ""
        row += f" │ {sign}{delta:>6.4f} ({sign}{pct:.1f}%)       "
    print(row)

    row = f"  {'Time vs AdamW':<28}"
    for r in results:
        ratio = r.total_time / base_time
        row += f" │ {ratio:>6.2f}x                    "
    print(row)

    # ────────────────────────────────────────────────────────────────────
    # Loss curves
    # ────────────────────────────────────────────────────────────────────
    print(f"\n\n  EVAL LOSS CURVES")
    header = f"  {'Step':<8}"
    sep2 = f"  {'─'*8}"
    for r in results:
        short_name = r.name.split("(")[0].strip()
        header += f" │ {short_name:<26}"
        sep2 += f" │{'─'*26}"
    print(header)
    print(sep2)

    all_steps = sorted(set(s for r in results for s, _ in r.eval_losses))
    for step in all_steps:
        row = f"  {step:<8}"
        for r in results:
            val = next((l for s, l in r.eval_losses if s == step), None)
            if val is not None:
                ppl = math.exp(min(val, 20))
                row += f" │ {val:>6.4f} (PPL {ppl:>6.2f})  "
            else:
                row += f" │ {'—':>24}  "
        print(row)

    # ────────────────────────────────────────────────────────────────────
    # Training loss curves (per LOG_EVERY)
    # ────────────────────────────────────────────────────────────────────
    print(f"\n\n  TRAINING LOSS CURVES")
    header = f"  {'Step':<8}"
    for r in results:
        short_name = r.name.split("(")[0].strip()
        header += f" │ {short_name:<26}"
    print(header)
    print(sep2)

    max_entries = max(len(r.losses) for r in results)
    for i in range(max_entries):
        step = (i + 1) * LOG_EVERY
        row = f"  {step:<8}"
        for r in results:
            if i < len(r.losses):
                ppl = math.exp(min(r.losses[i], 20))
                row += f" │ {r.losses[i]:>6.4f} (PPL {ppl:>6.2f})  "
            else:
                row += f" │ {'—':>24}  "
        print(row)

    # Winner announcement
    print(f"\n\n{'='*70}")
    best_result = min(results, key=lambda r: r.final_eval_loss)
    best_ppl = math.exp(min(best_result.final_eval_loss, 20))
    print(f"  WINNER: {best_result.name}")
    print(f"  Final eval loss: {best_result.final_eval_loss:.4f} (PPL {best_ppl:.2f})")
    improvement = (base_eval - best_result.final_eval_loss) / base_eval * 100
    if best_result.name != results[0].name:
        print(f"  Improvement over AdamW: {improvement:.2f}%")
    print(f"{'='*70}")
    print(f"\n  Log saved to: {log_path}", flush=True)
    tee.close()


if __name__ == "__main__":
    main()