ZigTensor

A deep-learning library in Zig. Explicit allocators, error unions, CPU ops you can read, CUDA kernels you wrote. Not a magic box, and not ready. The notes below are the sessions that moved the design, newest first.

  1. Strides, almost always hidden
  2. A hundred epochs, then fill again
  3. MNIST as a check, not a framework
  4. Zig 0.16
  5. Add that allocates
  6. The tensor has to own its shape
  7. This probably runs
  8. Nine months. Then the build.
  9. A placeholder for a log
  10. Fill on init
  11. Gone generic
  12. CPU and GPU are not the same init
  13. A tensor that exists
  14. The library has to compile