When I start fine-tuning yolo11n.pt (or any other version of yolo), the python.exe program will takes over 8GB, so I try everything to reduce the memory usage . But anyway,the following is my model configuration ,I'm wonder is this normal or not? And does it influence the accuracy or is there a way I can improve it without causing OOM eeror?(my dataset only contains 2000 image)

 if torch.cuda.is_available():
    torch.cuda.empty_cache()
    mp.set_start_method('spawn', force=True)  # Ensures proper process handling on Windows
    mp.freeze_support()  # Needed for Windows multiprocessing
    model = YOLO('yolo11n.pt')
    device=torch.device('cuda' if torch.cuda.is_available() 
    else 'cpu')
    results = model.train(
    data=data_yaml_path,
    epochs=150,              # Increased epochs
    batch=2,                # Reduced batch size
    imgsz=512,             # Reduced image size
    amp=True,
    # Stabilization parameters
    lr0=0.001,            # Lower learning rate
    lrf=0.001,  
    cache='disk',
    # Other parameters
    project='fall_detection_optimized',  # Save results in a new folder
    name='branch_first_phase',  
    device=device,
    optimizer='AdamW',
    workers=1,
    patience=25,
    weight_decay=0.0005,
    exist_ok=True,
    plots=True,
    save_period=50
)
# Save the best model
model.save('Fall_detection_optimized_best.pt')

So I lower everything and using cache to disk but it still takes 5GB, and I'm wonder does it impact the accuracy or how can I improve the accuracy