DarthAffe/stable-diffusion.cpp
1
0 Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-11-12 11:56:36 +00:00
Code Issues Packages Projects Releases Wiki Activity

refactor: reuse ggml internals in backend extensions

Browse Source
This commit is contained in:
leejet committed 2026-11-01 04:01:26 +08:00
1 parent 95773aa216
commit 6cfc3572f7
26 files changed
+358 -590
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

src/core/cuda/sage_attention.cu
+3 -4
View File
@@ -1,5 +1,5 @@
#include "backend.h"
#include "core/ggml_extensions.h"
#include "runtime.h"
#if __CUDACC_VER_MAJOR__ >= 12
@@ -353,9 +353,8 @@ namespace sd_cuda {
if (info.device < 0)
return;
int previous;
cudaDeviceProp properties;
cudaFuncAttributes attributes;
if (cudaGetDeviceProperties(&properties, info.device) != cudaSuccess || cudaGetDevice(&previous) != cudaSuccess ||
if (cudaGetDevice(&previous) != cudaSuccess ||
cudaSetDevice(info.device) != cudaSuccess)
return;
const auto error = cudaFuncGetAttributes(&attributes, sage_output);
@@ -364,7 +363,7 @@ namespace sd_cuda {
// PTX compiled for an older target retains that target's instruction guards after JIT.
static std::map<int, SageDevice> devices;
auto& data = devices[info.device];
data = {properties.major * 10 + properties.minor, std::min(attributes.binaryVersion, attributes.ptxVersion)};
data = {info.cc / 10, std::min(attributes.binaryVersion, attributes.ptxVersion)};
ggml_ext_kernel_desc desc = {};
desc.abi_version = GGML_EXT_ABI_VERSION;
desc.name = "sd.sage_attn.cuda.v1";