another attempt to tip the scales (volta)

Author: LostRuins

CMakeLists.txt

@@ -139,10 +139,10 @@ if (LLAMA_CUBLAS)
         elseif(CUDAToolkit_VERSION VERSION_GREATER 12)
                 add_compile_definitions(GGML_CUDA_USE_GRAPHS) #try enable cuda graphs on cu12 build
                 add_compile_definitions(KCPP_LIMIT_CUDA_MAX_ARCH=800)
-                set(CMAKE_CUDA_ARCHITECTURES "50-virtual;61-virtual;70-virtual;75-virtual;80-virtual") # lowest CUDA 12 standard + lowest for integer intrinsics
+                set(CMAKE_CUDA_ARCHITECTURES "50-virtual;61-virtual;75-virtual;80-virtual") # lowest CUDA 12 standard + lowest for integer intrinsics
         else()
                 add_compile_definitions(KCPP_LIMIT_CUDA_MAX_ARCH=750) #will cause issues with ggml_cuda_highest_compiled_arch if removed
-                set(CMAKE_CUDA_ARCHITECTURES "35-virtual;50-virtual;61-virtual;70-virtual;75-virtual") # lowest CUDA 12 standard + lowest for integer intrinsics
+                set(CMAKE_CUDA_ARCHITECTURES "35-virtual;50-virtual;61-virtual;75-virtual") # lowest CUDA 12 standard + lowest for integer intrinsics
         endif()
     endif()
     message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")

Makefile

@@ -226,15 +226,13 @@ NVCCFLAGS += -Wno-deprecated-gpu-targets \
              -gencode arch=compute_35,code=compute_35 \
              -gencode arch=compute_50,code=compute_50 \
              -gencode arch=compute_61,code=compute_61 \
-             -gencode arch=compute_70,code=compute_70 \
              -gencode arch=compute_75,code=compute_75 \
              -DKCPP_LIMIT_CUDA_MAX_ARCH=750
 
 else ifdef LLAMA_ARCHES_CU12
 NVCCFLAGS += -Wno-deprecated-gpu-targets \
              -gencode arch=compute_50,code=compute_50 \
              -gencode arch=compute_61,code=compute_61 \
-             -gencode arch=compute_70,code=compute_70 \
              -gencode arch=compute_75,code=compute_75 \
              -gencode arch=compute_80,code=compute_80 \
 			 -DKCPP_LIMIT_CUDA_MAX_ARCH=800

ggml/src/ggml-cuda/common.cuh

@@ -136,6 +136,10 @@ static int ggml_cuda_highest_compiled_arch(const int arch) {
 #ifndef KCPP_LIMIT_CUDA_MAX_ARCH
     return arch;
 #else
+    if(arch==GGML_CUDA_CC_VOLTA) //fix for kcpp, if volta try return 610 instead
+    {
+        return GGML_CUDA_CC_DP4A;
+    }
     return (arch > KCPP_LIMIT_CUDA_MAX_ARCH ? KCPP_LIMIT_CUDA_MAX_ARCH : arch);
 #endif
 }