sleepy/llama.cpp
1
0
Fork 0
Code Issues 16 Pull Requests Actions 139 Packages Projects Releases Wiki Activity

CUDA: use shared mem for ssm_conv (#20128)

Browse Source 
* CUDA: use shared mem for ssm_conv

* fuse silu + ssm_conv

* fuse unary + mul

* enable for fp16

* formatting

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>

---------

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
This commit is contained in:
Aman Gupta
2026-03-06 23:09:59 +08:00
committed by GitHub
parent 388baabc06
commit 1e38a7a6fa
6 changed files with 168 additions and 33 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/test-backend-ops.cpp
+3
View File
@@ -7663,6 +7663,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {d_conv, d_inner, 1, 1}, {d_conv, d_inner, 1, 1}));
test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {2 * d_conv, d_inner, 1, 1}, {d_conv, d_inner, 1, 1}));
test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {d_conv, d_inner, 4, 1}, {d_conv, d_inner, 1, 1}));
// long token (n_t > 32, exercises the long_token kernel path)
test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {d_conv - 1 + 64, d_inner, 1, 1}, {d_conv, d_inner, 1, 1}));
test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {d_conv - 1 + 64, d_inner, 4, 1}, {d_conv, d_inner, 1, 1}));
}
}