CUDA: add log line when mxfp4 acceleration is used (llama/18483)

* CUDA: add log line when mxfp4 acceleration is used

* add in backend_get_features

diff --git a/src/ggml-cuda/ggml-cuda.cu b/src/ggml-cuda/ggml-cuda.cu
index 40ffe92c575b7cb1918b60b1ae49a0ace1a24592..55e1c20c963f1ebabb9aea05b6239abd5ab2f9e6 100644 (file)
--- a/src/ggml-cuda/ggml-cuda.cu
+++ b/src/ggml-cuda/ggml-cuda.cu
@@ -4785,6 +4785,16 @@ static ggml_backend_feature * ggml_backend_cuda_get_features(ggml_backend_reg_t
         features.push_back({ "FA_ALL_QUANTS", "1" });
     #endif
 
+    {
+        const auto & info = ggml_cuda_info();
+        for (int id = 0; id < info.device_count; ++id) {
+            if (blackwell_mma_available(info.devices[id].cc)) {
+                features.push_back({ "BLACKWELL_NATIVE_FP4", "1"});
+                break;
+            }
+        }
+    }
+
     #undef _STRINGIFY
     #undef STRINGIFY