Repository navigation
Commit 6debe24
Write the GPT as plain SQL again, without the shared-CTE workaround
The model is a view again, built by functions that each return a few of
its CTEs as text, and the loss is a query that reads it. The DataFusion
host no longer runs each CTE on its own, and _relation.py is gone: the
demo shows the SQL a reader would write, and the work to make it fast
on DataFusion belongs upstream. ddx computes each CTE once on its own,
so training steps are as fast as before; DataFusion's forward loss takes
2.8 s rather than 0.04 s, and a model of more than one layer doesn't
plan on DataFusion until it computes a shared CTE once.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01NyN7idACeDMh94qrnFMu6h1 parent 714ba63 commit 6debe24
5 files changed
Lines changed: 149 additions & 342 deletions
0 commit comments