HNHacker News
TopNewBestAskShowJobs

emergentorder

3 karma · joined September 17, 2018

submissionscomments
emergentorder··on Scala 3.0
Yes, it uses ONNX Runtime / MLAS (their native BLAS lib) under the hood. And yes, there is copy overhead but you can eliminate it internally to a single function/graph by compiling it down to a single ONNX model. The end result is within ~15% run time of PyTorch w/ MKL when training a reasonably-sized MLP. And ORT also provides support for CUDA and a number of other "execution providers".
emergentorder··on Scala 3.0
Check out my project NDScala (https://github.com/SciScala/NDScala).

N-dimensional arrays in Scala 3. Think NumPy ndarray, but with compile-time type-checking/inference over shapes, ndarray/axis labels & numeric data types.