I have the following script, through which I would like to compare the memory usages of each size of the model on my hardware:

import torch
from sentence_transformers import SentenceTransformer
import gc

def print_memory_usage(stage: str):
    if torch.cuda.is_available():
        allocated_memory: int = torch.cuda.memory_allocated() / (1024 ** 2)
        print(f"Memory usage after {stage}: {allocated_memory:.2f} MB")
    else:
        print(f"CUDA is not available, memory tracking skipped after {stage}.")

def test_model(truncate_dim: int, use_half: bool):
    print(f"Testing with truncate_dim={truncate_dim}, use_half={use_half}")

    model = SentenceTransformer("aari1995/German_Semantic_V3", trust_remote_code=True, truncate_dim=truncate_dim)

    print_memory_usage("loading model")

    # model.max_seq_length = 16  # Optionally limit sequence length
    
    sentences: list[str] = [
        'Eine Flagge weht.',
        'Die Flagge bewegte sich in der Luft.',
        'Zwei Personen beobachten das Wasser.',
    ]
    
    # to make memory usage more visible
    sentences = sentences * 1000

    embeddings = model.encode(sentences, convert_to_tensor=True)
    
    if use_half:
        embeddings = embeddings.half()
    
    assert embeddings.shape[0] == len(sentences)
    print(f"Embeddings shape: {embeddings.shape}")
    
    print_memory_usage("encoding sentences")
    
    # Get the similarity scores for the embeddings
    similarities = model.similarity(embeddings, embeddings)
    print_memory_usage("calculating similarities")
    
    del model
    gc.collect()
    print_memory_usage("after deleting model")


truncate_dims: list[int] = [64, 128, 256, 512]
use_half_options: list[bool] = [True, False]

for truncate_dim in truncate_dims:
    for use_half in use_half_options:
        test_model(truncate_dim, use_half)
        print("-" * 50)

But I get roughly the same output for each model size, either with half precision or not:

Testing with truncate_dim=64, use_half=True
Memory usage after loading model: 5375.36 MB
Embeddings shape: torch.Size([3000, 64])
Memory usage after encoding sentences: 5383.85 MB
Memory usage after calculating similarities: 5401.02 MB
Memory usage after after deleting model: 25.66 MB
--------------------------------------------------
Testing with truncate_dim=64, use_half=False
Memory usage after loading model: 5383.49 MB
Embeddings shape: torch.Size([3000, 64])
Memory usage after encoding sentences: 5384.22 MB
Memory usage after calculating similarities: 5418.55 MB
Memory usage after after deleting model: 43.19 MB
--------------------------------------------------
Testing with truncate_dim=128, use_half=True
Memory usage after loading model: 5383.49 MB
Embeddings shape: torch.Size([3000, 128])
Memory usage after encoding sentences: 5384.22 MB
Memory usage after calculating similarities: 5401.39 MB
Memory usage after after deleting model: 26.02 MB
--------------------------------------------------
Testing with truncate_dim=128, use_half=False
Memory usage after loading model: 5383.49 MB
Embeddings shape: torch.Size([3000, 128])
Memory usage after encoding sentences: 5384.99 MB
Memory usage after calculating similarities: 5419.32 MB
Memory usage after after deleting model: 43.96 MB
--------------------------------------------------
Testing with truncate_dim=256, use_half=True
Memory usage after loading model: 5383.49 MB
Embeddings shape: torch.Size([3000, 256])
Memory usage after encoding sentences: 5384.99 MB
Memory usage after calculating similarities: 5402.15 MB
Memory usage after after deleting model: 26.79 MB
--------------------------------------------------
Testing with truncate_dim=256, use_half=False
Memory usage after loading model: 5383.49 MB
Embeddings shape: torch.Size([3000, 256])
Memory usage after encoding sentences: 5386.42 MB
Memory usage after calculating similarities: 5420.75 MB
Memory usage after after deleting model: 45.39 MB
--------------------------------------------------
Testing with truncate_dim=512, use_half=True
Memory usage after loading model: 5383.49 MB
Embeddings shape: torch.Size([3000, 512])
Memory usage after encoding sentences: 5386.42 MB
Memory usage after calculating similarities: 5403.58 MB
Memory usage after after deleting model: 28.22 MB
--------------------------------------------------
Testing with truncate_dim=512, use_half=False
Memory usage after loading model: 5383.49 MB
Embeddings shape: torch.Size([3000, 512])
Memory usage after encoding sentences: 5389.35 MB
Memory usage after calculating similarities: 5423.68 MB
Memory usage after after deleting model: 48.32 MB
--------------------------------------------------

What exactly is going on here, is there something wrong with the methodology? If so how can this script be fixed to show the correct memory usages?