CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-08 10:07:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

CUDA: fix crash on uneven context without FA (#16988)

Browse Source
This commit is contained in:
Johannes Gäßler
2025-11-06 14:05:47 +01:00
committed by GitHub
parent 5b180c3d60
commit aa374175c3
7 changed files with 44 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

44
tests/test-backend-ops.cpp
View File

@@ -3385,11 +3385,11 @@ struct test_mul_mat : public test_case {
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
const bool v; // whether a and b are non-contiguous views
const int64_t k_v; // size of k in memory, resulting in a non-contiguous view for k_v > k, no view for k_v == 0
const uint32_t o; // number of outputs
std::string vars() override {
return VARS_TO_STR10(type_a, type_b, m, n, k, bs, nr, per, v, o);
return VARS_TO_STR10(type_a, type_b, m, n, k, bs, nr, per, k_v, o);
}
double max_nmse_err() override {
@@ -3410,8 +3410,8 @@ struct test_mul_mat : public test_case {
std::array<int64_t, 2> bs = {10, 10},
std::array<int64_t, 2> nr = {2, 2},
std::array<int64_t, 4> per = {0, 1, 2, 3},
bool v = false, uint32_t o = 1)
: type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr), per(per), v(v), o(o) {}
int64_t k_v = 0, uint32_t o = 1)
: type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr), per(per), k_v(k_v), o(o) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
// C^T = A * B^T: (k, m) * (k, n) => (m, n)
@@ -3421,7 +3421,7 @@ struct test_mul_mat : public test_case {
const int npermuted = (per[0] != 0) + (per[1] != 1) + (per[2] != 2) + (per[3] != 3);
if (npermuted > 0) {
GGML_ASSERT(npermuted == 2);
GGML_ASSERT(!v); // not handled
GGML_ASSERT(k_v == 0); // not handled
GGML_ASSERT(!ggml_is_quantized(type_a) || per[0] == 0);
GGML_ASSERT(!ggml_is_quantized(type_b) || per[0] == 0);
@@ -3445,29 +3445,21 @@ struct test_mul_mat : public test_case {
ggml_set_name(a, "a_permuted");
ggml_set_name(b, "b_permuted");
} else {
if (v) {
a = ggml_new_tensor_4d(ctx, type_a, k*2, m, bs[0], bs[1]);
b = ggml_new_tensor_4d(ctx, type_b, k*2, n, bs[0]*nr[0], bs[1]*nr[1]);
const int64_t k_physical = k_v == 0 ? k : k_v;
a = ggml_new_tensor_4d(ctx, type_a, k_physical, m, bs[0], bs[1]);
b = ggml_new_tensor_4d(ctx, type_b, k_physical, n, bs[0]*nr[0], bs[1]*nr[1]);
if (!ggml_is_quantized(type_a)) {
if (bs[1] == 1 && nr[1] == 1) {
ggml_set_param(a);
}
ggml_set_param(b);
if (!ggml_is_quantized(type_a)) {
if (bs[1] == 1 && nr[1] == 1) {
ggml_set_param(a);
}
ggml_set_param(b);
}
if (k_v != 0) {
GGML_ASSERT(k_v > k);
a = ggml_view_4d(ctx, a, k, m, bs[0], bs[1], a->nb[1], a->nb[2], a->nb[3], 0);
b = ggml_view_4d(ctx, b, k, n, bs[0]*nr[0], bs[1]*nr[1], b->nb[1], b->nb[2], b->nb[3], 0);
} 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]);
if (!ggml_is_quantized(type_a)) {
if (bs[1] == 1 && nr[1] == 1) {
ggml_set_param(a);
}
ggml_set_param(b);
}
}
ggml_set_name(a, "a");
ggml_set_name(b, "b");
@@ -6901,7 +6893,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128, 45, 64, { 8, 1}, {4, 1}));
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056, 1, 193, {1, 1}, {4, 1}, {0, 2, 1, 3}));
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056, 1, 67, {1, 1}, {4, 1}, {0, 2, 1, 3}));
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 16, 32, 32, { 1, 1}, {1, 1}, {0, 1, 2, 3}, true, 3));
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 16, 32, 32, { 1, 1}, {1, 1}, {0, 1, 2, 3}, 64, 3));
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 64, 77, 77, {12,1}, {1,1}));
#if 0
@@ -6927,7 +6919,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
for (uint32_t k = 0; k < 2; ++k) {
for (ggml_type type: {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_F32}) {
test_cases.emplace_back(new test_mul_mat(type, GGML_TYPE_F32, 1056 + m, 1, 128 + k, {bs, bs2}, {nr, 1}, {0, 2, 1, 3}));
test_cases.emplace_back(new test_mul_mat(type, GGML_TYPE_F32, 128 + m, 1, 1056 + k, {bs, bs2}, {nr, 1}, {0, 1, 2, 3}, true));
test_cases.emplace_back(new test_mul_mat(type, GGML_TYPE_F32, 128 + m, 1, 1056 + k, {bs, bs2}, {nr, 1}, {0, 1, 2, 3}, 2*1056 + k));
}
}
}
@@ -7432,7 +7424,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
test_cases.emplace_back(new test_pad_reflect_1d(GGML_TYPE_F32, {3000, 384, 4, 1}));
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 16416, 1, 128, {8, 1}, {4, 1}, {0, 2, 1, 3}));
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128, 1, 16416, {8, 1}, {4, 1}, {0, 1, 2, 3}, true));
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128, 1, 16416, {8, 1}, {4, 1}, {0, 1, 2, 3}, 2*16416));
for (int bs : {1, 2, 3, 4, 5, 8, 512}) {
for (ggml_type type_a : all_types) {