Commit c32d1dabe for llama.cpp
commit c32d1dabe819002ca8aa3a885aca4057f5968e3d
Author: Gaurav Garg <gaugarg@nvidia.com>
Date: Thu Sep 10 17:42:40 2026 +0530
tests : increase tolerance for Add fusion tests (#28691)
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index 503998883..0a6516e61 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -3328,6 +3328,14 @@ struct test_bin_bcast : public test_case {
return op == ggml_div;
}
+ double max_nmse_err() override {
+ if (op == ggml_add && type == GGML_TYPE_F16 && nf > 1) {
+ // Fused ADDs can keep FP32 intermediates while the CPU rounds each ADD to FP16.
+ return 1e-6;
+ }
+ return test_case::max_nmse_err();
+ }
+
double max_maa_err() override {
return op == ggml_add ? 1e-4 : 1e-3;
}