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

now crashing

Browse source
This commit is contained in:
mike dupont 2023-11-27 07:30:23 -05:00
parent b484674707
commit 7ac56bdc62
1 changed files with 33 additions and 16 deletions
Download patch file Download diff file Expand all files Collapse all files

49
ggml.cpp
View file

@ -9426,23 +9426,37 @@ void ggml_tensor_checksum(const char * name,const struct ggml_tensor * tensor);
void ggml_tensor_hash(const char * name,const struct ggml_tensor * tensor, int decimalPlace); void ggml_tensor_hash(const char * name,const struct ggml_tensor * tensor, int decimalPlace);
#include "ggml-backend-impl.h" #include "ggml-backend-impl.h"
// helper function to convert the tensor buffer to a float array // helper function to convert the tensor buffer to a float array
float* ggml_tensor_to_float(const ggml_tensor& tensor, size_t* out_size) { float* ggml_tensor_to_float(const ggml_tensor* tensor) {
//if (tensor->type != GGML_TYPE_FLOAT) { //if (tensor->type != GGML_TYPE_FLOAT) {
//throw std::runtime_error("Only support for floating-point tensors"); //throw std::runtime_error("Only support for floating-point tensors");
//} //}
const size_t num_elements = tensor->n_dims > 0 ? std::accumulate(tensor->nb, tensor->nb + tensor->n_dims, 1) : 0; //if (out_size) {
float* buffer = new float[num_elements]; // *out_size = num_elements;
if (out_size) { // }
*out_size = num_elements;
} if(tensor->type == GGML_TYPE_F32)
memcpy(buffer, ggml_get_data_f32(tensor), ggml_nbytes(tensor)); {
const size_t num_elements = tensor->n_dims > 0 ? std::accumulate(tensor->nb, tensor->nb + tensor->n_dims, 1) : 0;
float* buffer = new float[num_elements];
memcpy(buffer, ggml_get_data_f32(tensor), ggml_nbytes(tensor));
return buffer;
}
else
{
const size_t num_elements = ggml_nbytes(tensor)/sizeof(float);
float* buffer = new float[num_elements];
memcpy(buffer, (float*)ggml_get_data(tensor), ggml_nbytes(tensor));
return buffer;
}
//memcpy(vec, ggml_get_data_f32(embeddings), ggml_nbytes(embeddings)); //memcpy(vec, ggml_get_data_f32(embeddings), ggml_nbytes(embeddings));
return buffer;
} }
// function to create a hash table of the N most common values of a given tensor // function to create a hash table of the N most common values of a given tensor
std::vector<double> find_n_most_common_values(const ggml_tensor& tensor, int decimal_place, size_t top_n) { std::vector<double> find_n_most_common_values(const ggml_tensor* tensor, int decimal_place, size_t top_n) {
float* buffer = ggml_tensor_to_float(tensor, nullptr); float* buffer = ggml_tensor_to_float(tensor);
auto values = std::unordered_map<double, int>(); // hash table to store the count of each value auto values = std::unordered_map<double, int>(); // hash table to store the count of each value
if (decimal_place <= 0 || top_n <= 0) { if (decimal_place <= 0 || top_n <= 0) {
@ -9450,16 +9464,19 @@ std::vector<double> find_n_most_common_values(const ggml_tensor& tensor, int dec
} }
// find N most common values by counting the frequency of each value with truncated decimal places // find N most common values by counting the frequency of each value with truncated decimal places
for (size_t i = 0; i < buffer->size(); ++i) { auto size = ggml_nbytes(tensor)/sizeof(float);
for (size_t i = 0; i < size; ++i) {
const double value = std::pow(10, static_cast<double>(decimal_place)); const double value = std::pow(10, static_cast<double>(decimal_place));
buffer[i] *= value; // multiply by value to truncate decimal places buffer[i] *= value; // multiply by value to truncate decimal places
int count = values.find(buffer[i])->second + 1; if (values.find(buffer[i]) != values.end()){
if (count > top_n) { int count = values.find(buffer[i])->second + 1;
if (count > top_n) {
continue; continue;
} }
if (decimal_place <= 0 || count >= top_n) { if (decimal_place <= 0 || count >= top_n) {
break; break;
} }
}
} }
// sort the values in descending order of frequency // sort the values in descending order of frequency