Niklas K.

Support offloading encode, for generate() with much less VRAM

6

generate() from Transformers can take encoder outputs as kwargs instead of running the encoder. This PR extends this to "super conditioning" sampling. It also enables providing only one "null sequence"...

Parallelizing "super conditioning" sampling speeds up inference by 20-40%

2

The current implementation of `DalleBart._sample()` performs "super conditioning" by running `DalleBart.decode()` twice in a row: https://github.com/borisdayma/dalle-mini/blob/a72705f46ce29a45d1c56c40b39e12476bfa6584/src/dalle_mini/model/modeling.py#L1838-L1856 Allowing JAX to schedule these in parallel makes inference *much* faster, especially but not...

support parameterized raw expressions in inserts/updates/criteria values

1

This PR enables using `raw()` *expressions* with parameters in insert queries, update queries and inside (right-hand side) value expressions of (where) criteria. Tests updated accordingly. ### Insert ```php QB::table('people')->getQuery('insert', [...

Shadows broken at high aspect ratios

2

Hi, as a multi-monitor user I'm somewhat used to glitches like these, so I don't consider this a high-priority issue, but I figured I'd flag it anyway. At extreme aspect...