root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-24 10:24:35 +00:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

Expose type name from ggml (#970)

Browse Source 
Avoid duplication of type names in utils

Co-authored-by: Håkon H. Hitland <haakon@likedan.net>
This commit is contained in:
Pavol Rusnak 2023-04-14 20:05:37 +02:00 committed by GitHub
parent f4d277ae17
commit c56b715269
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 27 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
examples/quantize-stats/quantize-stats.cpp
View File

@ -16,9 +16,6 @@
#include <unordered_map>
#include <vector>
static const char * type_strs[] = { "q4_0", "q4_1", "i8", "i16", "i32", "f16", "f32" };
static_assert(sizeof(type_strs) == GGML_TYPE_COUNT * sizeof(char *), "Incomplete type list");
struct quantize_stats_params {
std::string model = "models/7B/ggml-model-f16.bin";
bool verbose = false;
@ -224,7 +221,7 @@ int main(int argc, char ** argv) {
break;
}
int j;
for (j = 0; j < GGML_TYPE_COUNT && strcmp(argv[i], type_strs[j]) != 0; j++) {
for (j = 0; j < GGML_TYPE_COUNT && strcmp(argv[i], ggml_type_name((ggml_type) i)) != 0; j++) {
// find match
}
if (j < GGML_TYPE_COUNT) {
@ -279,7 +276,7 @@ int main(int argc, char ** argv) {
continue;
}
if (params.verbose) {
printf("%s: type %s, size %" PRId64 "\n", kv_tensor.first.c_str(), type_strs[kv_tensor.second->type], ggml_nelements(kv_tensor.second));
printf("%s: type %s, size %" PRId64 "\n", kv_tensor.first.c_str(), ggml_type_name(kv_tensor.second->type), ggml_nelements(kv_tensor.second));
}
if (kv_tensor.second->type == GGML_TYPE_F16) {
is_f16 = true;
@ -304,13 +301,14 @@ int main(int argc, char ** argv) {
// loop throught quantization types
for (int i = 0; i < GGML_TYPE_COUNT; i++) {
const ggml_type type = (ggml_type) i;
if (!params.include_types.empty() && std::find(params.include_types.begin(), params.include_types.end(), i) == params.include_types.end()) {
continue;
}
quantize_fns_t qfns = ggml_internal_get_quantize_fn(i);
if (qfns.quantize_row_q && qfns.dequantize_row_q) {
if (params.verbose) {
printf("testing %s ...\n", type_strs[i]);
printf("testing %s ...\n", ggml_type_name(type));
}
error_stats global_stats {};
@ -322,7 +320,7 @@ int main(int argc, char ** argv) {
if (params.verbose) {
printf(" %s ...\n", kv_tensor.first.c_str());
}
std::string layer_name { type_strs[i] };
std::string layer_name { ggml_type_name(type) };
layer_name += "::" + kv_tensor.first;
test_roundtrip_on_layer(
layer_name,
@ -337,7 +335,7 @@ int main(int argc, char ** argv) {
);
}
print_error_stats(type_strs[i], global_stats, params.print_histogram);
print_error_stats(ggml_type_name(type), global_stats, params.print_histogram);
}
}

17
ggml.c
View File

@ -2671,6 +2671,18 @@ static const size_t GGML_TYPE_SIZE[GGML_TYPE_COUNT] = {
};
static_assert(GGML_TYPE_COUNT == 7, "GGML_TYPE_SIZE is outdated");
static const char * GGML_TYPE_NAME[GGML_TYPE_COUNT] = {
[GGML_TYPE_F32] = "f32",
[GGML_TYPE_F16] = "f16",
[GGML_TYPE_Q4_0] = "q4_0",
[GGML_TYPE_Q4_1] = "q4_1",
[GGML_TYPE_I8] = "i8",
[GGML_TYPE_I16] = "i16",
[GGML_TYPE_I32] = "i32",
};
static_assert(GGML_TYPE_COUNT == 7, "GGML_TYPE_NAME is outdated");
static const char * GGML_OP_LABEL[GGML_OP_COUNT] = {
"NONE",
@ -2895,6 +2907,11 @@ float ggml_type_sizef(enum ggml_type type) {
return ((float)(GGML_TYPE_SIZE[type]))/GGML_BLCK_SIZE[type];
}
const char * ggml_type_name(enum ggml_type type) {
return GGML_TYPE_NAME[type];
}
size_t ggml_element_size(const struct ggml_tensor * tensor) {
return GGML_TYPE_SIZE[tensor->type];
}

2
ggml.h
View File

@ -354,6 +354,8 @@ int ggml_blck_size (enum ggml_type type);
size_t ggml_type_size (enum ggml_type type); // size in bytes for all elements in a block
float ggml_type_sizef(enum ggml_type type); // ggml_type_size()/ggml_blck_size() as float
const char * ggml_type_name(enum ggml_type type);
size_t ggml_element_size(const struct ggml_tensor * tensor);
struct ggml_context * ggml_init(struct ggml_init_params params);

14
llama.cpp
View File

@ -269,16 +269,6 @@ static std::string llama_format_tensor_shape(const std::vector<uint32_t> & ne) {
return ret;
}
static const char * llama_format_type(enum ggml_type type) {
switch (type) {
case GGML_TYPE_F32: return "f32";
case GGML_TYPE_F16: return "f16";
case GGML_TYPE_Q4_0: return "q4_0";
case GGML_TYPE_Q4_1: return "q4_1";
default: LLAMA_ASSERT(false);
}
}
static size_t llama_calc_tensor_size(const std::vector<uint32_t> & ne, enum ggml_type type) {
size_t size = ggml_type_size(type);
for (uint32_t dim : ne) {
@ -1582,7 +1572,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
printf("[%zu/%zu] %36s - %s, type = %6s, ",
++idx, model_loader->tensors_map.tensors.size(),
tensor.name.c_str(), llama_format_tensor_shape(tensor.ne).c_str(),
llama_format_type(tensor.type));
ggml_type_name(tensor.type));
// This used to be a regex, but <regex> has an extreme cost to compile times.
bool quantize = tensor.name.rfind("weight") == tensor.name.size() - 6; // ends with 'weight'?
@ -1615,7 +1605,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
f32_data[i] = ggml_fp16_to_fp32(f16_data[i]);
}
} else {
throw format("type %s unsupported for integer quantization", llama_format_type(tensor.type));
throw format("type %s unsupported for integer quantization", ggml_type_name(tensor.type));
}
printf("quantizing .. ");