mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 28 Packages Projects Releases Wiki Activity

cuda : extend GGML_OP_PAD to work with non-cont src0 (#19429)

Browse Source 
* cuda : extend GGML_OP_PAD to work with non-cont src0

* tests : add permuted pad
This commit is contained in:
Georgi Gerganov
2026-02-10 08:07:16 +02:00
committed by GitHub
parent 98e57ca422
commit a0d585537c
4 changed files with 28 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/test-backend-ops.cpp
+12 -9
View File
@@ -5894,33 +5894,36 @@ struct test_pad_ext : public test_case {
const int rp2;
const int lp3;
const int rp3;
const bool v;
const int tfrm; // 0 - none, 1 - non-cont, 2 - perm
const bool circular;
std::string vars() override {
return VARS_TO_STR12(type, ne_a, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3, v, circular);
return VARS_TO_STR12(type, ne_a, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3, tfrm, circular);
}
test_pad_ext(ggml_type type = GGML_TYPE_F32,
std::array<int64_t, 4> ne_a = {512, 512, 3, 1},
int lp0 = 1, int rp0 = 1, int lp1 = 1, int rp1 = 1,
int lp2 = 1, int rp2 = 1, int lp3 = 1, int rp3 = 1,
bool v = false, bool circular = false)
int tfrm = 0, bool circular = false)
: type(type), ne_a(ne_a), lp0(lp0), rp0(rp0), lp1(lp1), rp1(rp1), lp2(lp2), rp2(rp2), lp3(lp3), rp3(rp3),
v(v), circular(circular) {}
tfrm(tfrm), circular(circular) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne_a.data());
ggml_set_name(a, "a");
if (v) {
if (tfrm == 1) {
a = ggml_view_4d(ctx, a, (a->ne[0] + 1) / 2, (a->ne[1] + 1) / 2, (a->ne[2] + 1) / 2, (a->ne[3] + 1) / 2, a->nb[1], a->nb[2], a->nb[3], 0);
ggml_set_name(a, "view of a");
} else if (tfrm == 2) {
a = ggml_permute(ctx, a, 2, 1, 0, 3);
ggml_set_name(a, "permuted a");
}
ggml_tensor * out = circular
? ggml_pad_ext_circular(ctx, a, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3)
: ggml_pad_ext(ctx, a, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3);
: ggml_pad_ext (ctx, a, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3);
ggml_set_name(out, "out");
return out;
@@ -8198,10 +8201,10 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_solve_tri(GGML_TYPE_F32, { 64, 64, 4, 4 }, { 200, 64, 4, 4 }));
test_cases.emplace_back(new test_solve_tri(GGML_TYPE_F32, { 64, 64, 4, 4 }, { 384, 64, 4, 4 }));
for (bool v : {false, true}) {
for (int tfrm : {0, 1, 2}) {
for (bool circular : {false, true}) {
test_cases.emplace_back(new test_pad_ext(GGML_TYPE_F32, {512, 512, 1, 1}, 0, 1, 0, 1, 0, 0, 0, 0, v, circular));
test_cases.emplace_back(new test_pad_ext(GGML_TYPE_F32, {11, 22, 33, 44}, 1, 2, 3, 4, 5, 6, 7, 8, v, circular));
test_cases.emplace_back(new test_pad_ext(GGML_TYPE_F32, {512, 512, 1, 1}, 0, 1, 0, 1, 0, 0, 0, 0, tfrm, circular));
test_cases.emplace_back(new test_pad_ext(GGML_TYPE_F32, {11, 22, 33, 44}, 1, 2, 3, 4, 5, 6, 7, 8, tfrm, circular));
}
}