CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-31 08:51:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'master' into compilade/mamba2

Browse Source
This commit is contained in:
Francis Couture-Harpin
2024-11-01 11:12:18 -04:00
parent 805512a73b d865d1478c
commit 7d16e1bc8c
101 changed files with 12679 additions and 5471 deletions
Download Patch File Download Diff File Expand all files Collapse all files

122
tests/test-backend-ops.cpp
View File

@@ -1681,11 +1681,12 @@ struct test_mul_mat : 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
const std::array<int64_t, 2> bs; // dims 3 and 4
const std::array<int64_t, 2> nr; // repeat in dims 3 and 4
const std::array<int64_t, 4> per; // permutation of dimensions
std::string vars() override {
return VARS_TO_STR7(type_a, type_b, m, n, k, bs, nr);
return VARS_TO_STR8(type_a, type_b, m, n, k, bs, nr, per);
}
double max_nmse_err() override {
@@ -1700,17 +1701,44 @@ struct test_mul_mat : public test_case {
test_mul_mat(ggml_type type_a = GGML_TYPE_F32, ggml_type type_b = GGML_TYPE_F32,
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})
: type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr) {}
std::array<int64_t, 2> nr = {2, 2},
std::array<int64_t, 4> per = {0, 1, 2, 3})
: type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr), per(per) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
// C^T = A * B^T: (k, m) * (k, n) => (m, n)
ggml_tensor * a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0] , bs[1]);
ggml_tensor * b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]);
ggml_set_param(ctx, a);
ggml_set_param(ctx, b);
ggml_set_name(a, "a");
ggml_set_name(b, "b");
ggml_tensor * a;
ggml_tensor * b;
const int npermuted = (per[0] != 0) + (per[1] != 1) + (per[2] != 2) + (per[3] != 3);
if (npermuted > 0) {
GGML_ASSERT(npermuted == 2);
GGML_ASSERT(!ggml_is_quantized(type_a) || per[0] == 0);
GGML_ASSERT(!ggml_is_quantized(type_b) || per[0] == 0);
// Create tensors with the permuted dimensions, then permute them back to the dimensions given by m,n,k.
const int64_t ne_a[4] = {k, m, bs[0], bs[1]};
const int64_t ne_b[4] = {k, n, bs[0]*nr[0], bs[1]*nr[1]};
a = ggml_new_tensor_4d(ctx, type_a, ne_a[per[0]], ne_a[per[1]], ne_a[per[2]], ne_a[per[3]]);
b = ggml_new_tensor_4d(ctx, type_b, ne_b[per[0]], ne_b[per[1]], ne_b[per[2]], ne_b[per[3]]);
ggml_set_param(ctx, a);
ggml_set_param(ctx, b);
ggml_set_name(a, "a");
ggml_set_name(b, "b");
a = ggml_permute(ctx, a, per[0], per[1], per[2], per[3]);
b = ggml_permute(ctx, b, per[0], per[1], per[2], per[3]);
ggml_set_name(a, "a_permuted");
ggml_set_name(b, "b_permuted");
} else {
a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0], bs[1]);
b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]);
ggml_set_param(ctx, a);
ggml_set_param(ctx, b);
ggml_set_name(a, "a");
ggml_set_name(b, "b");
}
ggml_tensor * out = ggml_mul_mat(ctx, a, b);
ggml_set_name(out, "out");
@@ -3339,13 +3367,49 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
}
}
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16));
// test cases for 1D im2col
// im2col 1D
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32, {3000, 128, 1, 1}, {3, 128, 1280, 1}, 1, 0, 1, 0, 1, 0, false));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32, {3000, 128, 1, 1}, {3, 128, 1280, 1}, 1, 0, 1, 0, 1, 0, false));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {3000, 128, 1, 1}, {3, 128, 1280, 1}, 1, 0, 1, 0, 1, 0, false));
for (int s0 : {1, 3}) {
for (int p0 : {0, 3}) {
for (int d0 : {1, 3}) {
test_cases.emplace_back(new test_im2col(
GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32, {20, 2, 2, 1}, {3, 2, 2, 1},
s0, 0, p0, 0, d0, 0, false));
}
}
}
// im2col 2D
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16));
for (int s0 : {1, 3}) {
for (int s1 : {1, 3}) {
for (int p0 : {0, 3}) {
for (int p1 : {0, 3}) {
for (int d0 : {1, 3}) {
for (int d1 : {1, 3}) {
test_cases.emplace_back(new test_im2col(
GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32, {20, 20, 2, 2}, {3, 3, 2, 2},
s0, s1, p0, p1, d0, d1, true));
}
}
}
}
}
}
// extra tests for im2col 2D
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 1, 32}, {3, 3, 1, 32}, 1, 1, 1, 1, 1, 1, true));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 2, 32}, {3, 3, 2, 32}, 1, 1, 1, 1, 1, 1, true));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 1, 1024}, {3, 3, 1, 1024}, 1, 1, 1, 1, 1, 1, true));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 2, 1024}, {3, 3, 2, 1024}, 1, 1, 1, 1, 1, 1, true));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 1, 2048}, {3, 3, 1, 2048}, 1, 1, 1, 1, 1, 1, true));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 2, 2048}, {3, 3, 2, 2048}, 1, 1, 1, 1, 1, 1, true));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 1, 2560}, {3, 3, 1, 2560}, 1, 1, 1, 1, 1, 1, true));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 2, 2560}, {3, 3, 2, 2560}, 1, 1, 1, 1, 1, 1, true));
// sycl backend will limit task global_range < MAX_INT
// test cases for 2D im2col with large input W and H (occurs in stable-diffusion)
@@ -3474,13 +3538,14 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
#if 1
for (ggml_type type_a : base_types) {
for (ggml_type type_b : {GGML_TYPE_F32, GGML_TYPE_F16}) {
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, { 1, 1}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 1}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 1}, {2, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 10}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 10}, {2, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 10}, {1, 2}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 10}, {2, 2}));
// test cases without permutation
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, { 1, 1}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 1}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 1}, {2, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 10}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 10}, {2, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 10}, {1, 2}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 10}, {2, 2}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, { 1, 1}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {10, 1}, {1, 1}));
@@ -3489,6 +3554,19 @@ 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, 16, 256, {10, 10}, {2, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {10, 10}, {1, 2}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {10, 10}, {2, 2}));
// test cases with permutation
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {2, 3}, {1, 1}, {0, 2, 1, 3}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {2, 3}, {1, 1}, {0, 1, 3, 2}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {2, 3}, {1, 1}, {0, 3, 2, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 8, 256, {2, 3}, {1, 1}, {0, 2, 1, 3}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 8, 256, {2, 3}, {1, 1}, {0, 1, 3, 2}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 8, 256, {2, 3}, {1, 1}, {0, 3, 2, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 2, 1, 3}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 1, 3, 2}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 3, 2, 1}));
}
}
for (ggml_type type_a : other_types) {