GPUTOK: GPU Accelerated Byte Level BPE Tokenization

2026-03-02

AI Summary

This work addresses the performance bottleneck caused by CPU-based tokenizers in large language models with million-token context windows, which severely underutilizes GPU resources. We present the first efficient GPU implementation of a byte-level BPE tokenizer adhering to GPT-2’s merge rules, ensuring bit-identical output to widely used CPU tokenizers such as tiktoken and HuggingFace. Our design leverages a BlockBPE-style kernel, static mappings from cuCollections, CUB reductions, and a pybind11 interface. Evaluated on the WikiText103 dataset, our tokenizer achieves a 1.7× speedup over tiktoken and a 7.6× speedup over the HuggingFace GPT-2 tokenizer for the longest input sequences, substantially alleviating the tokenization bottleneck in long-context processing.

Technology Category

• Natural Language Processing: (Large) Language Models
• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Large Vision Models

Application Category

• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Search and Retrieval-Augmented AI: Large language models for search

Abstract

As large language models move toward million-token context windows, CPU tokenizers become a major slowdown because they process text one step at a time while powerful GPUs sit unused. We built a GPU-based byte-level BPE tokenizer that follows GPT-2's merge rules. It includes a basic BlockBPE-style kernel and a faster, optimized version that uses cuCollections static map, CUB reductions, and a pybind11 interface for Python. On WikiText103 sequences up to 131k tokens, the optimized GPU tokenizer produces the same tokens as a CPU version and, for the longest inputs, is about 1.7x faster than tiktoken and about 7.6x faster than the HuggingFace GPT-2 tokenizer. Nsight profiling shows that 70-80% of CUDA API time goes to memory allocation, so adding memory pooling should give the biggest speed boost next. Tests on generation tasks using WikiText103 prompts show that our GPU tokenizer's outputs stay within about one percentage point of tiktoken and HuggingFace GPT-2 on similarity and overlap metrics, meaning it keeps output quality while making long-context inference more practical.

Problem

• GPU acceleration
• tokenization
• byte-level BPE
• large language models
• long-context inference

Innovation

• GPU tokenization
• byte-level BPE
• CUDA optimization
• long-context LLM
• memory-efficient tokenizer