Fix missing arg in test/profile

karpathy · Jul 26, 2024 · 91fec3d · 91fec3d
1 parent ec04185
commit 91fec3d
Show file tree

Hide file tree

Showing 2 changed files with 2 additions and 2 deletions.
diff --git 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
@@ -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