diff --git a/Makefile b/Makefile
index 865f6fa..8be333b 100644
--- a/Makefile
+++ b/Makefile
@@ -203,14 +203,14 @@ train_gpt2cu: train_gpt2.cu cudnn_att.o
 train_gpt2fp32cu: train_gpt2_fp32.cu
 	$(NVCC) $(NVCC_FLAGS) $< $(NVCC_LDFLAGS) $(NVCC_INCLUDES) $(NVCC_LDLIBS) $(CUDA_OUTPUT_FILE)
 
-test_gpt2cu: test_gpt2.cu
-	$(NVCC) $(NVCC_FLAGS) $(PFLAGS) $< $(NVCC_LDFLAGS) $(NVCC_INCLUDES) $(NVCC_LDLIBS) $(CUDA_OUTPUT_FILE)
+test_gpt2cu: test_gpt2.cu cudnn_att.o
+	$(NVCC) $(NVCC_FLAGS) $(PFLAGS) $< $(NVCC_LDFLAGS) $(NVCC_INCLUDES) $(NVCC_LDLIBS) $(CUDA_OUTPUT_FILE) cudnn_att.o
 
 test_gpt2fp32cu: test_gpt2_fp32.cu
 	$(NVCC) $(NVCC_FLAGS) $< $(NVCC_LDFLAGS) $(NVCC_INCLUDES) $(NVCC_LDLIBS) $(CUDA_OUTPUT_FILE)
 
-profile_gpt2cu: profile_gpt2.cu
-	$(NVCC) $(NVCC_FLAGS) $(PFLAGS) -lineinfo $< $(NVCC_LDFLAGS) $(NVCC_INCLUDES) $(NVCC_LDLIBS)  $(CUDA_OUTPUT_FILE)
+profile_gpt2cu: profile_gpt2.cu cudnn_att.o
+	$(NVCC) $(NVCC_FLAGS) $(PFLAGS) -lineinfo $< $(NVCC_LDFLAGS) $(NVCC_INCLUDES) $(NVCC_LDLIBS)  $(CUDA_OUTPUT_FILE) cudnn_att.o
 
 clean:
 	$(REMOVE_FILES) $(TARGETS)
diff --git a/profile_gpt2.cu b/profile_gpt2.cu
index 5b043e0..2ea7f98 100644
--- a/profile_gpt2.cu
+++ b/profile_gpt2.cu
@@ -49,9 +49,7 @@ int main() {
     cublasCheck(cublasSetMathMode(cublas_handle, cublas_math_mode));
     // setup the (global) cuBLASLt workspace
     cudaCheck(cudaMalloc(&cublaslt_workspace, cublaslt_workspace_size));
-    #ifdef ENABLE_CUDNN
-    checkCudnnErr(cudnnCreate(&cudnn_handle));
-    #endif
+    create_cudnn();
 
     // build the GPT-2 model from a checkpoint
     GPT2 model;
@@ -81,10 +79,7 @@ int main() {
 
     // free
     gpt2_free(&model);
-    #ifdef ENABLE_CUDNN
-    if (cudnn_workspace != NULL) { cudaCheck(cudaFree(cudnn_workspace)); }
-    checkCudnnErr(cudnnDestroy(cudnn_handle));
-    #endif
+    destroy_cudnn();
     cudaCheck(cudaFree(cublaslt_workspace));
     cublasCheck(cublasDestroy(cublas_handle));
     cublasCheck(cublasLtDestroy(cublaslt_handle));
diff --git a/test_gpt2.cu b/test_gpt2.cu
index 67afa50..1a7df63 100644
--- a/test_gpt2.cu
+++ b/test_gpt2.cu
@@ -106,10 +106,8 @@ int main(int argc, char *argv[]) {
     cublasMath_t cublas_math_mode = enable_tf32 ? CUBLAS_TF32_TENSOR_OP_MATH : CUBLAS_DEFAULT_MATH;
     cublasCheck(cublasSetMathMode(cublas_handle, cublas_math_mode));
     cudaCheck(cudaMalloc(&cublaslt_workspace, cublaslt_workspace_size));
-
-    #ifdef ENABLE_CUDNN
-    checkCudnnErr(cudnnCreate(&cudnn_handle));
-    #endif
+    // set up cuDNN (noop if not available)
+    create_cudnn();
 
     // build the GPT-2 model from a checkpoint
     GPT2 model;
@@ -326,10 +324,7 @@ int main(int argc, char *argv[]) {
     free(grads_memory_cpu);
     free(grads_memory_cpu_float);
     gpt2_free(&model);
-    #ifdef ENABLE_CUDNN
-    if (cudnn_workspace != NULL) { cudaCheck(cudaFree(cudnn_workspace)); }
-    checkCudnnErr(cudnnDestroy(cudnn_handle));
-    #endif
+    destroy_cudnn();
     cudaCheck(cudaFree(cublaslt_workspace));
     cublasCheck(cublasDestroy(cublas_handle));
     cublasCheck(cublasLtDestroy(cublaslt_handle));