Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-20 01:31:42 +02:00
Code Issues Packages Projects Releases Wiki Activity

lets try without volta specific kernels, fattn should fall back to tile

Browse Source
This commit is contained in:
Concedo
2025-12-06 15:56:07 +08:00
parent eaf61c3fd7
commit 1a14ae1183
2 changed files with 8 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/common.cuh
+4 -3
View File
@@ -237,7 +237,7 @@ static const char * cu_get_error_str(CUresult err) {
// The Volta instructions are in principle available on Turing or newer but they are effectively unusable:
#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
#define VOLTA_MMA_AVAILABLE
// #define VOLTA_MMA_AVAILABLE
#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
@@ -275,7 +275,7 @@ static bool fast_fp16_hardware_available(const int cc) {
// To be used for feature selection of external libraries, e.g. cuBLAS.
static bool fp16_mma_hardware_available(const int cc) {
return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_VOLTA) ||
return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc > GGML_CUDA_CC_VOLTA) ||
GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc) ||
(GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_QY2);
}
@@ -303,7 +303,8 @@ static bool amd_wmma_available(const int cc) {
}
static bool volta_mma_available(const int cc) {
return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) == GGML_CUDA_CC_VOLTA;
return false;
// return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) == GGML_CUDA_CC_VOLTA;
}
static bool turing_mma_available(const int cc) {
ggml/src/ggml-cuda/fattn-mma-f16.cuh
+4 -4
View File
@@ -89,10 +89,10 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
}
static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_config_volta(const int DKQ, const int DV, const int ncols) {
GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 8, 64, 4, 32, 288, 256, 64, 1, false);
GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16, 64, 4, 32, 288, 256, 64, 1, false);
GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2, 32, 160, 128, 64, 1, false);
GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1, 32, 160, 128, 64, 1, false);
// GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 8, 64, 4, 32, 288, 256, 64, 1, false);
// GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16, 64, 4, 32, 288, 256, 64, 1, false);
// GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2, 32, 160, 128, 64, 1, false);
// GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1, 32, 160, 128, 64, 1, false);
// TODO tune specifically for Volta
return ggml_cuda_fattn_mma_get_config_ampere(DKQ, DV, ncols);