TurboGPT - Train a tiny GPT in under a minute on CUDA
Show HN: TurboGPT: train 22KiB transformer in 13s
TurboGPT is a lightweight C++/CUDA framework that lets you train a 22KiB transformer in just 13 seconds. It replaces CLI hyperparameters with a simple experiment.h config, supports resumable PyTorch-compatible checkpoints, and logs per-step training loss to TensorBoard. Runs end with a short generated sample, making it perfect for rapid experimentation, education, and prototyping tiny language models without the overhead of large frameworks. If you have an NVIDIA GPU, you can go from clone to trained model in under a minute.
Train a tiny GPT in under a minute (CUDA only).
- _345
Trying to be a little less negative than the guy that got flagged, I too don't understand the motivation behind these projects. I've seen a hundred of them at this point- and each of them is probably worse and has less learning value than the one Andrej Karpathy made to teach people the building blocks involved in a GPT
So, why do people keep making these? Asking genuinely
- vjsrinivas
Interesting trend of labeling models after how big they are on disk vs how many learnable parameters they have.
- jey
At that scale it seems like you could just solve the KKT conditions directly. Exaggerating but only a little