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

now faster and smaller

Browse source
This commit is contained in:
mike dupont 2023-11-28 21:50:31 -05:00
parent d1d1cceda7
commit 1807a6e280
3 changed files with 1131 additions and 3650 deletions
Download patch file Download diff file Expand all files Collapse all files

4
Makefile
View file

@ -17,6 +17,8 @@ ifndef UNAME_S
UNAME_S := $(shell uname -s)
endif
LLAMA_DEBUG := 1
ifndef UNAME_P
UNAME_P := $(shell uname -p)
endif
@ -116,7 +118,7 @@ endif
# keep standard at C11 and C++11
MK_CPPFLAGS = -I. -Icommon
MK_CFLAGS = -std=c11 -fPIC
MK_CXXFLAGS = -std=c++20 -fPIC -fpermissive -DCPP_ONLY
MK_CXXFLAGS = -std=c++20 -fPIC -fpermissive -DCPP_ONLY
# -Ofast tends to produce faster code, but may not be available for some compilers.
ifdef LLAMA_FAST

4764
Untitled.ipynb
View file

File diff suppressed because one or more lines are too long

13
ggml.cpp
View file

@ -9440,8 +9440,8 @@ float* ggml_tensor_to_float(const ggml_tensor* tensor) {
// function to create a hash table of the N most common values of a given tensor
void find_n_most_common_values(const char * pname, const ggml_tensor* tensor, int decimal_place, size_t top_n) {
//float* buffer = ggml_tensor_to_float(tensor);
float* buffer = 0;
//float* buffer = 0;
//if(tensor->type == GGML_TYPE_F32)
// {
@ -9451,10 +9451,11 @@ void find_n_most_common_values(const char * pname, const ggml_tensor* tensor, in
// memcpy(buffer, ggml_get_data_f32(tensor), ggml_nbytes(tensor));
//return buffer;
const size_t num_elements = ggml_nbytes(tensor)/sizeof(float);
//buffer = new float[num_elements];
buffer=(float*)ggml_get_data(tensor);
const size_t num_elements = ggml_nelements(tensor);
if (tensor->type != GGML_TYPE_F32) {
return;
}
float* buffer = ggml_get_data_f32(tensor);
auto values = std::unordered_map<double, int>(); // hash table to store the count of each value