bpe-tokenizer
nanoeuler
GPT-2-style LLM built from scratch in C/CUDA with hand-written backprop, BPE tokenizer, FlashAttention, pretraining, and SFT.
- 127
- Cuda
- MIT
- 5 mo ago
bpe-tokenizer
Byte-Pair Encoding tokenizer for training large language models on huge datasets
- 6
- Python
- GPL-3.0
- 2 y ago
Biptoken
Fast BPE tokenizer with guaranteed perfect text reconstruction. 2x faster than tiktoken at encoding/decoding while preserving all whitespace and formatting.
- 2
- Python
- MIT
- 18 mo ago
PythonGPT
A 54M parameter GPT transformer implemented from scratch in raw PyTorch (no HuggingFace). Trained on 1GB of filtered Python code from bigcode/the-stack-dedup, fine-tuned on 863 curated algorithm implementations. Features BPE tokenization, top-p nucleus sampling, constrained decoding via Python's tokenize module, and a Gemini-inspired Streamlit UI.
- 1
- Python
- MIT
- 5 mo ago
Build-and-Finetune-LLM-From-Scratch-Deploy-via-vLLM-AWS-GCP
A complete end-to-end learning repo covering everything from building Large Language Models (LLMs) from scratch to mastering practical deep learning with PyTorch. Includes tokenizer coding, transformers, attention, training loops, model finetuning, and hands-on PyTorch projects.
- 0
- Jupyter Notebook
- MIT
- 7 mo ago

