diff --git a/ggml/src/ggml-cuda/conv2d-implicit.cu b/ggml/src/ggml-cuda/conv2d-implicit.cu index bd67ac2b86..d2b5ee6d33 100644 --- a/ggml/src/ggml-cuda/conv2d-implicit.cu +++ b/ggml/src/ggml-cuda/conv2d-implicit.cu @@ -949,6 +949,8 @@ static __global__ void conv2d_implicit_kernel(const half * __restrict__ input, s = 0; r++; } + A_block_smem = shmem; + B_block_smem = &shmem[BM * BK]; } // iter r // reuse smem