Commit 64c46fc

JohannesGaessler

committed

CUDA: faster softmax via shared memory + fp16 math

1 parent 540938f commit 64c46fcCopy full SHA for 64c46fc

2 files changed

+285

-24

lines changed

ggml-cuda.cu
tests
- test-backend-ops.cpp

2 files changed

+285

-24

lines changed

Comments

(0)