vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

ggml : sync latest ggml_mul_mat_id

Browse source
This commit is contained in:
Georgi Gerganov 2023-12-09 11:19:46 +02:00
parent a3eefe95a8
commit 861cd67899
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
4 changed files with 114 additions and 75 deletions
Download patch file Download diff file Expand all files Collapse all files

34
tests/test-backend-ops.cpp
View file

@ -770,11 +770,9 @@ struct test_mul_mat_id : public test_case {
const int64_t m;
const int64_t n;
const int64_t k;
const std::array<int64_t, 2> bs; // dims 3 and 4
const std::array<int64_t, 2> nr; // repeat in dims 3 and 4
std::string vars() override {
return VARS_TO_STR9(type_a, type_b, n_mats, id, m, n, k, bs, nr);
return VARS_TO_STR7(type_a, type_b, n_mats, id, m, n, k);
}
double max_nmse_err() override {
@ -782,7 +780,7 @@ struct test_mul_mat_id : public test_case {
}
size_t op_size(ggml_tensor * t) override {
size_t a = ggml_nbytes(t->src[2]) * n * nr[0] * nr[1];
size_t a = ggml_nbytes(t->src[2]) * n;
size_t b = ggml_nbytes(t->src[1]) * m;
size_t c = ggml_nbytes(t);
return a + b + c;
@ -792,35 +790,37 @@ struct test_mul_mat_id : public test_case {
test_mul_mat_id(ggml_type type_a = GGML_TYPE_F32, ggml_type type_b = GGML_TYPE_F32,
int n_mats = 2, int id = 0,
int64_t m = 32, int64_t n = 32, int64_t k = 32,
std::array<int64_t, 2> bs = {10, 10},
std::array<int64_t, 2> nr = {2, 2})
int64_t m = 32, int64_t n = 32, int64_t k = 32)
: type_a(type_a), type_b(type_b), n_mats(n_mats), id(id),
m(m), n(n), k(k), bs(bs), nr(nr) {}
m(m), n(n), k(k) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
// C^T = A * B^T: (k, m) * (k, n) => (m, n)
std::vector<ggml_tensor *> mats;
for (int i = 0; i < n_mats; i++) {
ggml_tensor * a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0], bs[1]);
ggml_tensor * a = ggml_new_tensor_2d(ctx, type_a, k, m);
mats.push_back(a);
}
ggml_tensor * ids = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, n_mats);
ggml_tensor * b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]);
ggml_tensor * ids = ggml_new_tensor_2d(ctx, GGML_TYPE_I32, n_mats, n);
ggml_tensor * b = ggml_new_tensor_2d(ctx, type_b, k, n);
ggml_tensor * out = ggml_mul_mat_id(ctx, mats.data(), ids, id, b);
return out;
}
void initialize_tensors(ggml_context * ctx) override {
std::random_device rd;
std::default_random_engine rng(rd());
for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
if (t->type == GGML_TYPE_I32) {
// ids
std::vector<int> data(n_mats);
for (int i = 0; i < n_mats; i++) {
data[i] = i;
for (int64_t r = 0; r < ggml_nrows(t); r++) {
std::vector<int32_t> data(t->ne[0]);
for (int i = 0; i < t->ne[0]; i++) {
data[i] = i;
}
std::shuffle(data.begin(), data.end(), rng);
ggml_backend_tensor_set(t, data.data(), r * t->nb[1], t->ne[0] * sizeof(int32_t));
}
std::shuffle(data.begin(), data.end(), std::default_random_engine(std::random_device()()));
ggml_backend_tensor_set(t, data.data(), 0, n_mats * sizeof(int));
} else {
init_tensor_uniform(t);
}
@ -1215,7 +1215,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
for (ggml_type type_b : {GGML_TYPE_F32 /*, GGML_TYPE_F16 */}) {
for (int n_mats : {1, 2, 4}) {
for (int id = 0; id < n_mats; id++) {
test_cases.emplace_back(new test_mul_mat_id(type_a, type_b, n_mats, id, 16, 16, 256, {1, 1}, {1, 1}));
test_cases.emplace_back(new test_mul_mat_id(type_a, type_b, n_mats, id, 16, 16, 256));
}
}
}