Joel Berkeley
Joel Berkeley
See [Stefan Hoeck's thread](https://discord.com/channels/827106007712661524/827106088343175180/1157021237801910394) in Discord Unfortunately, the current implementation of ST does not provide the performance benefits we'd want from it, because the resulting monadic code is not further...
The Compiler module should only be used to provide implementation for `Tensor` functionality. It shouldn't be importable by client code.
Strictly orthogonal `Broadcastable` constructors may make implementations easier
The implementations for uniform and normal sampling in XLA use the same XLA API we use in Idris, so there's little performance gain from using those APIs. IIRC they also...
We are seeing a lot of tests for floating point ops fail frequently. Common culprits are `cosh`, `sinh`, `uniform`, `normal`. We'd like two things Step 1: CI to reliably pass...