PSE: Non-Bijunctive Attention Collapse for LLM Inference
A POWER8-specific research prototype that replaces some full-matrix attention computation with a hardware AES instruction used as a fast prefilter, so only a subset of attention paths get the full computation.
PSE vcipher collapse is a research prototype for POWER8 hardware that explores selective attention-path collapse in LLM inference. Standard transformer attention computes every query-key dot product before softmax selects the important ones; this prototype instead uses a hardware AES instruction as a near-constant-time prefilter to select candidate attention paths before computing the full dot product only for the top quarter of them, and uses a vector permute instruction to prune weak paths and duplicate strong ones.
The documentation is explicit that this is a POWER8-specific exploration, not a general GPU attention replacement, and that a later documentation pass summarized existing claims without rerunning benchmarks or validating new llama.cpp throughput. It targets an IBM POWER8 system and is written in C, licensed under AGPL-3.0.