← Back to events
ActiveAIRelease

viable/strict/1789770346: [inductor] Optimize full slice_scatter chains (#195631)

Photo: PyTorch Releases

What happened

The following summary was prepared with assistance from Codex and reviewed by the author. Rewrite slice_scatter chains that fill an entire tensor into peak memory efficient cat or copy_. Functionalizing chunking results in a chain of slice_scatter nodes, holding full-sized intermediate by the end of the chain. The main goal of this pass is to make the "chunking" region the most peak memory efficient, as this is the…

Summary assembled by rule from the sources below

Why it's spreading

Sources