PrototypeCommon
JustVugg/nanoeuler
GPT-2-style LLM built from scratch in C/CUDA with hand-written backprop, BPE tokenizer, FlashAttention, pretraining, and SFT.
- Owner
- JustVugg (personal account)
- Stars
- Stars: 127
- Forks
- Forks: 26
- Language
- Cuda
- License
- MIT
- Created
- 2026
- Topics
- bpe-tokenizerbyte-pair-encodingccublascudadeep-learningflashattentionfrom-scratchgpt2language-model
- Tier
- Prototype (100 to 249 ★) What this tier means
More repos by JustVugg
All 3 repos by JustVugg- JustVugg/colibriStars: 46,772CLTS
- JustVugg/lumabriStars: 211CPrototype
Nearby in Cuda
All Cuda repos- sjfeng1999/gpu-arch-microbenchmarkStars: 129CudaPrototype
- Avarok-Cybersecurity/dgx-vllmStars: 128CudaPrototype
- psmarter/CUDA-PracticeStars: 127CudaPrototype
- dingwentao/GPU-lossless-compressionStars: 126CudaPrototype
- lzhengchun/matrix-cudaStars: 125CudaPrototype
- dgSPARSE/dgSPARSE-LibStars: 125CudaPrototype