Skip to content

QK-norm + RoPE fusion

QK-norm + RoPE fusion combines Q/K normalization and rotary-position work on the attention Q/K path. This targets repeated small operators that otherwise launch separately and move intermediate tensors through memory.

Role in Sol-Engine

This is part of the DiT attention-path fusion design space and is especially relevant when attention remains active after sparse attention choices are fixed.

Search knobs

  • which attention blocks use fused Q/K preprocessing.
  • tensor layout expected by the downstream attention backend.
  • precision format used before attention.

Validation

Confirm compatibility with the selected attention backend and with any quantization path applied to QKV projections.

References