diff --git a/tests/unit/test_qk_direct.cu b/tests/unit/test_qk_direct.cu index edc27718..b205e370 100644 --- a/tests/unit/test_qk_direct.cu +++ b/tests/unit/test_qk_direct.cu @@ -92,7 +92,7 @@ test_qk_direct_kernel(float* __restrict__ out_s, __syncthreads(); const bool my_warp_active = (wid == 0); - const int my_row = lane; + const int my_row = threadIdx.x % 32; const bool my_row_active = my_row < T; constexpr int NUM_READS = SK / 8;