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

metal : initial working version

Browse source
This commit is contained in:
Georgi Gerganov 2024-02-07 11:20:04 +02:00
parent 099afc6274
commit 92a0c17474
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
3 changed files with 439 additions and 59 deletions
Download patch file Download diff file Expand all files Collapse all files

38
tests/test-backend-ops.cpp
View file

@ -480,12 +480,13 @@ struct test_case {
double err = nmse(f1.data(), f2.data(), f1.size());
if (err > ud->max_err) {
printf("[%s] NMSE = %.9f > %.9f ", ggml_op_desc(t1), err, ud->max_err);
//for (int i = 0; i < (int) f1.size(); i++) {
// printf("%5d %9.6f %9.6f, diff = %9.6f\n", i, f1[i], f2[i], f1[i] - f2[i]);
//}
//printf("\n");
//exit(1);
printf("[%s] NMSE = %.9f > %.9f", ggml_op_desc(t1), err, ud->max_err);
printf("\n");
for (int i = 0; i < (int) f1.size(); i++) {
printf("%5d %9.6f %9.6f, diff = %9.6f\n", i, f1[i], f2[i], f1[i] - f2[i]);
}
printf("\n");
exit(1);
ud->ok = false;
}
return true;
@ -572,9 +573,19 @@ struct test_case {
// duplicate the op
size_t target_size = ggml_backend_is_cpu(backend) ? 1ULL << 33 : 1ULL << 35; // 8 GB CPU, 32 GB GPU
int n_runs = std::min((size_t)gf->size - gf->n_nodes, target_size / op_size(out)) + 1;
#if 0
for (int i = 1; i < n_runs; i++) {
gf->nodes[gf->n_nodes++] = out;
}
#else
n_runs = 256;
int n_nodes = gf->n_nodes;
for (int i = 0; i < n_runs; i++) {
for (int j = 0; j < n_nodes; j++) {
gf->nodes[gf->n_nodes++] = gf->nodes[j];
}
}
#endif
// calculate memory
size_t mem = n_runs * op_size(out);
@ -2044,6 +2055,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
test_cases.emplace_back(new test_rms_norm(GGML_TYPE_F32, {64, 10, 10, 10}, eps));
}
#if 0
for (ggml_type type_a : all_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}));
@ -2063,6 +2075,20 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {10, 10}, {2, 2}));
}
}
#else
for (ggml_type type_a : {GGML_TYPE_Q4_0, GGML_TYPE_Q8_0, GGML_TYPE_F16}) {
for (ggml_type type_b : {GGML_TYPE_F32}) {
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 4096, 1, 4096, { 1, 1}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 4096, 2, 4096, { 1, 1}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 4096, 3, 4096, { 1, 1}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 4096, 4, 4096, { 1, 1}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 4096, 5, 4096, { 1, 1}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 4096, 6, 4096, { 1, 1}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 4096, 7, 4096, { 1, 1}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 4096, 8, 4096, { 1, 1}, {1, 1}));
}
}
#endif
for (ggml_type type_a : all_types) {
for (ggml_type type_b : {GGML_TYPE_F32 /*, GGML_TYPE_F16 */}) {