tensomite.cr

A small pure-Crystal tensor maths runtime with a minimal native seam to BLAS and the GPU.

Tensomite gives Crystal lazy tensors with automatic gradients, for training models and running them. Its kernels are written in Crystal: they run on the CPU, through BLAS (Accelerate or OpenBLAS), or on an Apple GPU through Metal, and repeat their results exactly from run to run. It loads GPT-2 from Hugging Face's safetensors files.

require "tensomite"
include Tensomite

x = Tensor.from([1, 2, 3].map(&.to_f32), [3])
loss = (x * x).sum
Tensomite.grad(loss, [x]).first.to_a(Float32)   # => [2.0, 4.0, 6.0]

Using it

• User's guide: the API, with small examples.
• Tutorial: train a tiny network to multiply, then use it.
• examples/shakespeare.cr trains a character-level GPT; examples/gpt2_diagram.cr draws GPT-2's structure.

Working on it

• PLAN.md: goals, milestones and decisions.
• DESIGN.md: how it works, and its measured performance.
• DEVELOPMENT.md: setup, commands, conventions and workflow.
• HANDOFF.md: where things stand, for the next session.

Licensed under the MPL-2.0.