What is the purpose of torch.manual_seed() and how do you ensure reproducibility in PyTorch?

PyTorch uses pseudo-random number generators for weight initialisation, dropout masks, data shuffling, and more. Setting seeds explicitly ensures experiments are reproducible — critical for debugging, comparing model variants fairly, and scientific rigor.

import torch
import numpy as np
import random
import os

def set_seed(seed: int = 42):
    """Set all relevant seeds for full reproducibility."""
    random.seed(seed)                       # Python's random module
    np.random.seed(seed)                    # NumPy
    torch.manual_seed(seed)                 # PyTorch CPU
    torch.cuda.manual_seed_all(seed)        # PyTorch all GPUs

    # Force deterministic algorithms (may be slower!)
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False   # disable auto-tuner (non-deterministic)

    os.environ["PYTHONHASHSEED"] = str(seed)

set_seed(42)

# Verify reproducibility
model1 = torch.nn.Linear(10, 5)
set_seed(42)
model2 = torch.nn.Linear(10, 5)
print(torch.equal(model1.weight, model2.weight))  # True — identical init

# DataLoader reproducibility — also needs a worker_init_fn for num_workers > 0
def seed_worker(worker_id):
    worker_seed = torch.initial_seed() % 2**32
    np.random.seed(worker_seed)
    random.seed(worker_seed)

generator = torch.Generator()
generator.manual_seed(42)

from torch.utils.data import DataLoader
loader = DataLoader(
    dataset,
    batch_size=32,
    shuffle=True,
    num_workers=4,
    worker_init_fn=seed_worker,   # seeds each worker process
    generator=generator,           # seeds the shuffling order
)