Commit 8f900ab

JohannesGaessler

authored

CUDA: faster softmax via shared memory + fp16 math (#4742)

1 parent 1fc2f26 commit 8f900abCopy full SHA for 8f900ab

2 files changed

+318

-26

lines changed

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

2 files changed

+318

-26

lines changed

Comments

(0)