From 91fec3d670aaf35347d6e5309348dc7660d9f531 Mon Sep 17 00:00:00 2001 From: Aleksa Gordic Date: Fri, 26 Jul 2024 17:08:43 +0200 Subject: [PATCH] Fix missing arg in test/profile --- profile_gpt2.cu | 2 +- test_gpt2.cu | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/profile_gpt2.cu b/profile_gpt2.cu index 6d1b69850..54811a78d 100644 --- a/profile_gpt2.cu +++ b/profile_gpt2.cu @@ -56,7 +56,7 @@ int main(int argc, char *argv[]) { // override number of layers to 1 because all layers repeat the same kernels, only profile once model.config.num_layers = 1; - set_zero_configs(&multi_gpu_config, 0, model.num_parameters); + set_zero_configs(&multi_gpu_config, 0, model.num_parameters, model.high_perf_mode); gpt2_allocate_state(&model, B, T); // do a training step diff --git a/test_gpt2.cu b/test_gpt2.cu index cec435568..e344288ea 100644 --- a/test_gpt2.cu +++ b/test_gpt2.cu @@ -142,7 +142,7 @@ int main(int argc, char *argv[]) { printf("batch_size: %d\n", B); printf("seq_len: %d\n", T); - set_zero_configs(&multi_gpu_config, 0, model.num_parameters); + set_zero_configs(&multi_gpu_config, 0, model.num_parameters, model.high_perf_mode); // read reference information from the file saved from Python/PyTorch side // 1) input x and y