diff --git a/test_gpt2.cu b/test_gpt2.cu
index 7613b6b..654e35d 100644
--- a/test_gpt2.cu
+++ b/test_gpt2.cu
@@ -247,7 +247,7 @@ int main(int argc, char *argv[]) {
             // In that case it's ok to extend the tolerance by a bit, after a manual review.
             allok = allok & check_tensor(tensors1[0], tensors2[0], V * C, "wte", 8e-1f);
             allok = allok & check_tensor(tensors1[1], tensors2[1], maxT * C, "wpe", 1e-2f);
-            allok = allok & check_tensor(tensors1[2], tensors2[2], L * 3*C * C, "qkvw", 1.1e-1); // hmm a bit high
+            allok = allok & check_tensor(tensors1[2], tensors2[2], L * 3*C * C, "qkvw", 1.4e-1); // hmm a bit high
             allok = allok & check_tensor(tensors1[3], tensors2[3], L * 3*C, "qkvb", 4e-2f);
             allok = allok & check_tensor(tensors1[4], tensors2[4], L * C * C, "attprojw", 3e-2f);
             allok = allok & check_tensor(tensors1[5], tensors2[5], L * C, "attprojb", 3e-2f);