Tensorwire

Two things I've been building in private: running big open-weight models across the machines you already have, and training on them too.

We published the SQ4 whitepaper and opened a PR to bring percentile-calibrated 4-bit quantization to llama.cpp.

Inference +28%. Training +54% at dim≥1024. The tokenizer was O(n²) and nobody noticed until we fed it 48MB.

The 1.54x number has an asterisk. I ran PyTorch with default settings. Here's what I still owe.

I got tired of Python's toolchain for fine-tuning. So I wrote my own — and accidentally built something faster.