cuda: add a fused candidate-scoring kernel for CLM and Kev - #69

Draft

xiaoyu-xyz wants to merge 7 commits into

ThinkFlowLab:mainfrom

xiaoyu-xyz:cuda-scoring-kernel

Commits on Sep 28, 2026

cuda: fused candidate scoring, verified on an RTX 4090
xiaoyu-xyz
committed
19f2d5c

Commits on Oct 3, 2026

cuda: the scoring kernel is compiled, and its batch path and rejections run
xiaoyu-xyz
committed
b96b940
cuda: quote #9 and #27 for the readout instead of a phrase neither contains
xiaoyu-xyz
committed
54d37df
cuda: say in the manifest what was actually run
xiaoyu-xyz
committed
78ddfe6
cuda: measure the fusion, and fix what the measurement found
xiaoyu-xyz
committed
1a85317
cuda: time the fused kernel through a CUDA Graph too
xiaoyu-xyz
committed
251d691

Commits on Oct 4, 2026

cuda: give the scoring library an ABI macro the manifest can be checked against
xiaoyu-xyz
committed
a0d3d4b