From 5a505cbc62dddf99d0bf588044953559e30e0f37 Mon Sep 17 00:00:00 2001
From: Concedo <39025047+LostRuins@users.noreply.github.com>
Date: Sat, 3 Jan 2026 22:45:06 +0800
Subject: [PATCH] disable blackwell mma for now

---
 ggml/src/ggml-cuda/common.cuh | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
index 3c78e8fa9..a6e21092a 100644
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -256,7 +256,7 @@ static const char * cu_get_error_str(CUresult err) {
 #endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 
 #if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_BLACKWELL && __CUDA_ARCH__ < GGML_CUDA_CC_RUBIN
-#    define BLACKWELL_MMA_AVAILABLE
+// #    define BLACKWELL_MMA_AVAILABLE // kcpp: no blackwell mma for now, requires too high cuda
 #endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_BLACKWELL
 
 #if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
@@ -331,6 +331,7 @@ static bool cp_async_available(const int cc) {
 }
 
 static bool blackwell_mma_available(const int cc) {
+    return false; //kcpp: no blackwell mma for now, requires too high cuda
     return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_BLACKWELL &&
            ggml_cuda_highest_compiled_arch(cc) < GGML_CUDA_CC_RUBIN;
 }