CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-27 08:21:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

opencl: add q8_0 mm support (#16469)

Browse Source 
* opencl: add mm_q8_0_f32

* opencl: fix data loading for incomplete tile

* opencl: use q8_0 mm for larger matrix

* opencl: add some tests to cover the path
This commit is contained in:
lhez
2025-10-15 10:51:04 -07:00
committed by GitHub
parent d93f8439b0
commit 0cb7a0683b
6 changed files with 271 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
tests/test-backend-ops.cpp
View File

@@ -6365,6 +6365,19 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
}
}
#if 0
{
// Test paths in OpenCL
std::vector<int> ns = {32, 64, 128, 256, 512, 1024, 4096};
std::vector<int> ks = {896, 1536, 4096};
for (auto n : ns) {
for (auto k : ks) {
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q8_0, GGML_TYPE_F32, 1024, n, k, {1, 1}, {1, 1}));
}
}
}
#endif
#if 1
for (ggml_type type_a : base_types) {
for (ggml_type type_b : {GGML_TYPE_F32, GGML_TYPE_F16}) {