nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 14 Packages Projects Releases Wiki Activity

ggml-webgpu: improve MTP inference by using mat-vec path for small batches (#24811)

Browse Source 
* ggml-webgpu: improve small batches decoding

* Add barrier to the NUM_COLS loop in mul-mat-vec
This commit is contained in:
Masashi Yoshimura
2026-06-23 17:13:55 +09:00
committed by GitHub
parent 035cd8f9a6
commit 7c908502ea
8 changed files with 682 additions and 591 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/test-backend-ops.cpp
+2
View File
@@ -8433,6 +8433,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, k, {3, 2}, {2, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, k, {3, 2}, {1, 2}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, k, {3, 2}, {2, 2}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 4, k, {3, 2}, {2, 2}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {1, 1}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {1, 1}, {2, 1}));
@@ -8449,6 +8450,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, k, {2, 3}, {1, 1}, {0, 1, 3, 2}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, k, {2, 3}, {1, 1}, {0, 3, 2, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 4, k, {2, 3}, {1, 1}, {0, 3, 2, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 8, k, {2, 3}, {1, 1}, {0, 2, 1, 3}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 8, k, {2, 3}, {1, 1}, {0, 1, 3, 2}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 8, k, {2, 3}, {1, 1}, {0, 3, 2, 1}));