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

ggml/ex: calculate accuracy in graph, adapt MNIST (ggml/980)

Browse source
This commit is contained in:
Johannes Gäßler 2024-10-03 17:29:59 +02:00 • committed by Georgi Gerganov
parent eee39bdc96
commit fabdc3bda3
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
11 changed files with 389 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

79
tests/test-backend-ops.cpp
View file

@ -116,6 +116,11 @@ static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float m
} else if (tensor->type == GGML_TYPE_I8 || tensor->type == GGML_TYPE_I16 || tensor->type == GGML_TYPE_I32) {
// This is going to create some weird integers though.
ggml_backend_tensor_set(tensor, data.data(), 0, ggml_nbytes(tensor));
} else if (tensor->type == GGML_TYPE_I64) {
// Integers with a size of 8 bytes can be set by mirroring the float data, the specific values are again not really meaningful.
const size_t nbytes_half = ggml_nbytes(tensor)/2;
ggml_backend_tensor_set(tensor, data.data(), 0*nbytes_half, nbytes_half);
ggml_backend_tensor_set(tensor, data.data(), 1*nbytes_half, nbytes_half);
} else {
GGML_ABORT("fatal error");
}
@ -145,6 +150,8 @@ static std::vector<float> tensor_to_float(const ggml_tensor * t) {
tv.push_back(ggml_bf16_to_fp32(*(ggml_bf16_t*)&buf[i]));
} else if (t->type == GGML_TYPE_F32) {
tv.push_back(*(float *) &buf[i]);
} else if (t->type == GGML_TYPE_I64) {
tv.push_back((float)*(int64_t *) &buf[i]);
} else if (t->type == GGML_TYPE_I32) {
tv.push_back((float)*(int32_t *) &buf[i]);
} else if (t->type == GGML_TYPE_I16) {
@ -1116,6 +1123,71 @@ struct test_get_rows : public test_case {
}
};
// GGML_OP_ARGMAX
struct test_argmax : public test_case {
const ggml_type type;
const std::array<int64_t, 4> ne;
std::string vars() override {
return VARS_TO_STR2(type, ne);
}
test_argmax(ggml_type type = GGML_TYPE_F32,
std::array<int64_t, 4> ne = {10, 100, 1, 1})
: type(type), ne(ne) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
ggml_set_name(a, "a");
ggml_tensor * out = ggml_argmax(ctx, a);
ggml_set_name(out, "out");
return out;
}
double max_nmse_err() override {
return 0.0;
}
};
// GGML_OP_COUNT_EQUAL
struct test_count_equal : public test_case {
const ggml_type type;
const std::array<int64_t, 4> ne;
std::string vars() override {
return VARS_TO_STR2(type, ne);
}
test_count_equal(ggml_type type = GGML_TYPE_F32,
std::array<int64_t, 4> ne = {4, 500, 1, 1})
: type(type), ne(ne) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
ggml_set_name(a, "a");
ggml_tensor * a_argmax = ggml_argmax(ctx, a);
ggml_set_name(a_argmax, "a_argmax");
ggml_tensor * b = ggml_new_tensor(ctx, type, 4, ne.data());
ggml_set_name(b, "b");
ggml_tensor * b_argmax = ggml_argmax(ctx, a);
ggml_set_name(b_argmax, "b_argmax");
ggml_tensor * out = ggml_count_equal(ctx, a_argmax, b_argmax);
ggml_set_name(out, "out");
return out;
}
double max_nmse_err() override {
return 0.0;
}
};
// GGML_OP_REPEAT
struct test_repeat : public test_case {
const ggml_type type;
@ -3260,6 +3332,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {3,1,2,1}, 1, 0, 1));
test_cases.emplace_back(new test_conv_transpose_1d({2,1,1,1}, {3,1,1,1}, 1, 0, 1));
test_cases.emplace_back(new test_argmax());
test_cases.emplace_back(new test_count_equal());
for (int ne3 : {1, 3}) { // CUDA backward pass only supports ne3 == 1
test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 5, 4, ne3}, {1, 1, 1, 1}));
test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 5, 4, ne3}, {2, 1, 1, 1}));
@ -3278,8 +3353,8 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_dup(GGML_TYPE_F16, {10, 10, 5, 1}, {0, 2, 1, 3})); // dup by rows
test_cases.emplace_back(new test_dup(GGML_TYPE_F32, {10, 10, 5, 1}, {1, 0, 2, 3}));
test_cases.emplace_back(new test_dup(GGML_TYPE_F16, {10, 10, 5, 1}, {1, 0, 2, 3})); // dup dst not-contiguous
test_cases.emplace_back(new test_dup(GGML_TYPE_I16, {10, 8, 3, 1}, {0, 2, 1, 3}));
test_cases.emplace_back(new test_dup(GGML_TYPE_I16, {10, 8, 3, 1}, {1, 2, 0, 3}));
test_cases.emplace_back(new test_dup(GGML_TYPE_I16, {10, 8, 3, 1}, {0, 2, 1, 3}));
test_cases.emplace_back(new test_dup(GGML_TYPE_I16, {10, 8, 3, 1}, {1, 2, 0, 3}));
for (int dim = 1; dim < GGML_MAX_DIMS; ++dim) {
test_cases.emplace_back(new test_set(GGML_TYPE_F32, GGML_TYPE_F32, {6, 5, 4, 3}, dim));