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

ggml : extend ggml_get_rows, ggml_repeat, ggml_concat (ggml/639)

Browse source 
* add more int ops

* ggml_compute_forward_dup_bytes

* add tests

* PR comments

* tests : minor indentations

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Guillaume Wenzek 2023-12-29 18:07:03 +01:00 • committed by Georgi Gerganov
parent f2eb19bd8b
commit 5f66ebca9c
2 changed files with 198 additions and 10 deletions
Download patch file Download diff file Expand all files Collapse all files

42
tests/test-backend-ops.cpp
View file

@ -58,6 +58,9 @@ static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float m
int64_t hist[16];
ggml_quantize_chunk(tensor->type, data.data(), dataq.data(), 0, size, hist);
ggml_backend_tensor_set(tensor, dataq.data(), 0, dataq.size());
} 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 {
GGML_ASSERT(false);
}
@ -87,8 +90,13 @@ static std::vector<float> tensor_to_float(const ggml_tensor * t) {
tv.push_back(*(float *) &buf[i]);
} else if (t->type == GGML_TYPE_I32) {
tv.push_back((float)*(int32_t *) &buf[i]);
} else if (t->type == GGML_TYPE_I16) {
tv.push_back((float)*(int16_t *) &buf[i]);
} else if (t->type == GGML_TYPE_I8) {
tv.push_back((float)*(int8_t *) &buf[i]);
} else if (quantized) {
tt.to_float(&buf[i], vq.data(), bs);
std::vector<float> vq(ggml_blck_size(t->type));
tt.to_float(&buf[i], vq.data(), ggml_blck_size(t->type));
tv.insert(tv.end(), vq.begin(), vq.end());
} else {
GGML_ASSERT(false);
@ -661,17 +669,26 @@ struct test_repeat : public test_case {
struct test_dup : public test_case {
const ggml_type type;
const std::array<int64_t, 4> ne;
const std::array<int64_t, 4> permute;
bool _use_permute;
std::string vars() override {
return VARS_TO_STR2(type, ne);
std::string v = VARS_TO_STR2(type, ne);
if (_use_permute) v += "," + VAR_TO_STR(permute);
return v;
}
test_dup(ggml_type type = GGML_TYPE_F32,
std::array<int64_t, 4> ne = {10, 10, 10, 1})
: type(type), ne(ne) {}
std::array<int64_t, 4> ne = {10, 10, 10, 1},
std::array<int64_t, 4> permute = {0, 0, 0, 0})
: type(type), ne(ne), permute(permute),
_use_permute(permute[0] + permute[1] + permute[2] + permute[3] > 0) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * src = ggml_new_tensor(ctx, type, 4, ne.data());
if (_use_permute) {
src = ggml_permute(ctx, src, permute[0], permute[1], permute[2], permute[3]);
}
ggml_tensor * out = ggml_dup(ctx, src);
return out;
}
@ -1450,14 +1467,26 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
}
}
}
for (int b : {1, 7}) {
for (bool v : {false, true}) {
test_cases.emplace_back(new test_get_rows(GGML_TYPE_I32, 256, 5, 4, b, v));
}
}
test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {1, 1, 1, 1}));
test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {2, 1, 1, 1}));
test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {1, 2, 1, 1}));
test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {1, 1, 2, 1}));
test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {1, 1, 1, 2}));
test_cases.emplace_back(new test_repeat(GGML_TYPE_I32, {10, 10, 10, 10}, {2, 1, 1, 1}));
test_cases.emplace_back(new test_repeat(GGML_TYPE_I16, {10, 10, 10, 10}, {1, 1, 1, 2}));
test_cases.emplace_back(new test_dup());
test_cases.emplace_back(new test_dup(GGML_TYPE_F32));
test_cases.emplace_back(new test_dup(GGML_TYPE_F16));
test_cases.emplace_back(new test_dup(GGML_TYPE_I32));
test_cases.emplace_back(new test_dup(GGML_TYPE_I16));
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 (ggml_type type : all_types) {
test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, type, {256, 10, 10, 1}));
@ -1565,7 +1594,8 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
test_cases.emplace_back(new test_alibi());
test_cases.emplace_back(new test_im2col());
test_cases.emplace_back(new test_concat());
test_cases.emplace_back(new test_concat(GGML_TYPE_F32));
test_cases.emplace_back(new test_concat(GGML_TYPE_I32));
for (ggml_sort_order order : {GGML_SORT_ASC, GGML_SORT_DESC}) {
test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {8, 1, 1, 1}, order));