Selecting the vocabulary size, often denoted as $|V|$ , for your subword tokenizer is a significant hyperparameter choice with direct consequences for model performance, memory consumption, and computational cost. Unlike simpler tokenization methods where vocabulary might grow organically, subword algorithms like BPE or WordPiece require you to pre-define a target vocabulary size. This decision involves balancing several competing factors.

Impact on Sequence Length and Information Granularity

A smaller vocabulary size forces the tokenizer to break words down into smaller subword units more frequently.

Advantage: It handles rare and out-of-vocabulary (OOV) words gracefully by composing them from common subwords. This improves robustness to typos and neologisms.
Disadvantage: It leads to longer sequences of tokens for the same input text. Since the computational cost of attention mechanisms in Transformers scales quadratically with sequence length ( $O(L^2)$ ), longer sequences significantly increase training and inference time and memory usage.

Conversely, a larger vocabulary size allows the tokenizer to represent more common words or frequent subword sequences as single tokens.

Advantage: It results in shorter token sequences, reducing the computational burden of the attention mechanism. More semantic meaning might be captured within a single token embedding if common words are kept intact.
Disadvantage: It increases the risk of OOV words if the vocabulary isn't large or diverse enough for the target domain. It also increases the size of the model's embedding layer.

Consider the word "tokenization".

With a small vocabulary (e.g., $|V|=10,000$ ), it might be tokenized as ['tok', 'en', 'ization'] (3 tokens).
With a larger vocabulary (e.g., $|V|=50,000$ ) that includes "tokenization" during training, it might become ['tokenization'] (1 token).
With a medium vocabulary (e.g., $|V|=30,000$ ), perhaps ['token', 'ization'] (2 tokens).

Longer sequences directly impact the memory needed for activations during training, especially the attention score matrix ( $L \times L$ ).

Larger vocabularies generally lead to shorter average sequence lengths for a fixed corpus, reducing attention computation costs.

Impact on Model Size and Memory

The vocabulary size directly determines the size of the input embedding matrix and the output projection layer (often tied). The embedding matrix has dimensions $|V| \times d_{model}$ , where $d_{model}$ is the hidden dimension of the model.

\text{Embedding Matrix Size} = |V| \times d_{model} \times \text{bytes\_per\_parameter}

A larger $|V|$ leads to a proportionally larger embedding matrix. For large models with $d_{model}$ in the thousands, increasing $|V|$ from 30,000 to 100,000 can add billions of parameters and gigabytes to the model's memory footprint, solely in the embedding layer.

import torch
import torch.nn as nn

# Example embedding layer sizes
d_model = 4096  # Example hidden dimension

vocab_size_small = 32000
vocab_size_large = 128000

embedding_small = nn.Embedding(vocab_size_small, d_model)
embedding_large = nn.Embedding(vocab_size_large, d_model)

params_small = sum(p.numel() for p in embedding_small.parameters())
params_large = sum(p.numel() for p in embedding_large.parameters())

# Memory assuming float32 (4 bytes per parameter)
mem_small_gb = params_small * 4 / (1024**3)
mem_large_gb = params_large * 4 / (1024**3)

print(
    f"Vocab Size: {vocab_size_small}, "
    f"Embedding Params: {params_small:,}, "
    f"Memory: {mem_small_gb:.2f} GB"
)
print(
    f"Vocab Size: {vocab_size_large}, "
    f"Embedding Params: {params_large:,}, "
    f"Memory: {mem_large_gb:.2f} GB"
)

# --- Output ---
# Vocab Size: 32000, Embedding Params: 131,072,000, Memory: 0.49 GB
# Vocab Size: 128000, Embedding Params: 524,288,000, Memory: 1.95 GB

This calculation only covers the embedding layer. The final output layer, which projects the hidden state back to the vocabulary dimension for predicting the next token, often has a similar size ( $d_{model} \times |V|$ ), further amplifying the impact of $|V|$ on model size.

Impact on Computation (Softmax)

During training and inference, the model typically calculates probabilities over the entire vocabulary using a softmax function applied to the output logits. The computational cost of this final softmax layer scales linearly with $|V|$ .

\text{Softmax Cost} \propto L \times d_{model} \times |V|

While the attention mechanism ( $O(L^2 \times d_{model})$ ) often dominates for long sequences, the softmax computation ( $O(L \times d_{model} \times |V|)$ ) can become a significant bottleneck, especially with very large vocabularies or during inference where batch sizes might be small and latency is important. A larger $|V|$ directly increases this cost.

Finding a Balance: Common Practices

There's no single "best" vocabulary size; it depends on the specific task, language(s), dataset size, model architecture, and available computational resources. However, some common observations and practices exist:

Typical Ranges: For many English language models, vocabulary sizes often range from 30,000 to 100,000. Sizes around 32,000, 50,000, and 64,000 are frequently seen, often chosen for computational convenience (e.g., divisibility by powers of 2 or hardware vector units). Multilingual models might require larger vocabularies (e.g., 100,000 to 256,000) to adequately cover multiple character sets and morphologies.
Diminishing Returns: Increasing $|V|$ yields diminishing returns in sequence length reduction and perplexity improvement beyond a certain point. The gain from 30k to 50k might be noticeable, while the gain from 100k to 150k might be marginal but incur significant memory costs.
Data Characteristics: The optimal size is influenced by the diversity and size of the training data. More diverse data (multiple languages, domains, code) might benefit from a larger vocabulary.
Model Size: For smaller models, the proportional increase in size due to a larger vocabulary is more significant. For extremely large models (hundreds of billions of parameters), the embedding layer might represent a smaller fraction of the total parameters, potentially making larger vocabularies more acceptable if they offer performance benefits.
Computational Budget: Ultimately, hardware limitations (GPU memory) and training time constraints often impose practical upper bounds on $|V|$ .

Trade-offs associated with selecting a smaller versus a larger vocabulary size ( $|V|$ ).

In practice, selecting $|V|$ often involves some empirical testing or adopting sizes reported in literature for similar model scales and datasets. You need to weigh the benefits of shorter sequences and potentially better representation of common terms (larger $|V|$ ) against the costs of increased memory usage, slower softmax computation, and potentially worse handling of rare terms (smaller $|V|$ ).