Add option (again)

Author: pt13762104

ggml/CMakeLists.txt

@@ -158,6 +158,8 @@ option(GGML_CUDA                            "ggml: use CUDA"
 option(GGML_MUSA                            "ggml: use MUSA"                                  OFF)
 option(GGML_CUDA_FORCE_MMQ                  "ggml: use mmq kernels instead of cuBLAS"         OFF)
 option(GGML_CUDA_FORCE_CUBLAS               "ggml: always use cuBLAS instead of mmq kernels"  OFF)
+option(GGML_CUDA_NO_TURING_MMA              "ggml: disable the use of mma in mmq kernels"     OFF)
+
 set   (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
                                             "ggml: max. batch size for using peer access")
 option(GGML_CUDA_NO_PEER_COPY               "ggml: do not use peer to peer copies"            OFF)

ggml/src/ggml-cuda/CMakeLists.txt

@@ -77,6 +77,11 @@ if (CUDAToolkit_FOUND)
         add_compile_definitions(GGML_CUDA_FORCE_CUBLAS)
     endif()
 
+    if (GGML_CUDA_NO_TURING_MMA)
+        add_compile_definitions(GGML_CUDA_NO_TURING_MMA)
+        add_compile_definitions(GGML_CUDA_FORCE_MMQ)
+    endif()
+
     if (GGML_CUDA_NO_VMM)
         add_compile_definitions(GGML_CUDA_NO_VMM)
     endif()

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -71,6 +71,10 @@
 #include <string>
 #include <vector>
 
+#ifdef GGML_CUDA_NO_TURING_MMA
+#define CUBLAS_COMPUTE_16F CUBLAS_COMPUTE_16F_PEDANTIC
+#endif
+
 static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
 
 [[noreturn]]

ggml/src/ggml-cuda/mmq.cuh

@@ -9,6 +9,9 @@
 
 using namespace ggml_cuda_mma;
 
+#ifdef GGML_CUDA_NO_TURING_MMA
+#undef TURING_MMA_AVAILABLE
+#endif
 #define MMQ_DP4A_MAX_BATCH_SIZE 64 // Max. batch size to use for dp4a MMQ kernels when FP16 tensor cores are available.
 #define MMQ_ITER_K 256
 #define MMQ_NWARPS 8