jax
83c292bd - [Mosaic GPU] Add conversion logic for `i4 -> f8e4m3fn`.

Commit
352 days ago
[Mosaic GPU] Add conversion logic for `i4 -> f8e4m3fn`. The inline PTX is able to upcast 4 values at a time, so we use a generator to pack several registers together when our registers don't hold enough packed values. This makes the generated PTX smaller, since the conversion routine needs to be called less often. PiperOrigin-RevId: 770640667
Author
Parents
Loading