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

vulkan: handle mat_mul with A matrix > 4GB (#16176)

Browse Source 
* vulkan: handle mat_mul with A matrix > 4GB

This change splits mat_mul operations with huge A matrix into chunks in the M
dimension. This works well for stable-diffusion use cases where the im2col
matrix has very large M.

Fix the order of setting the stride in mul_mm_cm2 - setting the dimension
clobbers the stride, so stride should be set after.

* build fixes
This commit is contained in:
Jeff Bolz
2025-09-27 20:36:34 -05:00
committed by GitHub
parent e6d65fb02d
commit 1384abf8b8
3 changed files with 59 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
tests/test-backend-ops.cpp
View File

@@ -6200,6 +6200,14 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_gla(GGML_TYPE_F32, 32, 64, 32, 4));
test_cases.emplace_back(new test_gla(GGML_TYPE_F32, 32, 64, 128, 4));
#if 0
// > 4GB A matrix. Too slow to be enabled by default.
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F16, 900000, 3, 2592, {1, 1}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F16, 1700000, 96, 2592, {1, 1}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F16, 1700000, 3, 2592, {1, 1}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F16, 1700000, 1, 2592, {1, 1}, {1, 1}));
#endif
for (ggml_type type_a : all_types) {
for (int i = 1; i < 10; ++i) {
test_cases.emplace_back(new test_mul_mat(type_a, GGML_TYPE_F32, 16, i, 256, { 1, 1}, {1, 1}));