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

compress: format

Browse source
This commit is contained in:
Stéphane du Hamel 2024-09-25 01:26:39 +02:00
parent b9a32f464f
commit bec83989be
1 changed files with 33 additions and 27 deletions
Download patch file Download diff file Expand all files Collapse all files

32
examples/compress/compress.cpp
View file

@ -78,7 +78,8 @@ std::vector<uint8_t> encode(llama_context *ctx, std::vector<llama_token> inp, gp
break;
}
}
if(match<0){
if (match < 0)
{
LOG_ERR("\n couldn't match %s", llama_token_to_piece(ctx, inp[index]).c_str());
exit(1);
}
@ -140,7 +141,6 @@ std::vector<uint8_t> encode(llama_context *ctx, std::vector<llama_token> inp, gp
}
sample_ids_bitpacked[block_start + 1] = block_size & 0xff;
// put last bytes
if (PAD)
{
@ -544,7 +544,8 @@ int main(int argc, char **argv)
auto t_enc_end = ggml_time_us();
LOG("\n");
if(!params.no_perf){
if (!params.no_perf)
{
LOG("\nInput: %d characters (%d tokens)", params.prompt.length(), inp.size());
float compressed_bits_per_token = 8 * (float)sample_ids_bitpacked.size() / (float)inp.size();
@ -555,19 +556,22 @@ int main(int argc, char **argv)
LOG("\nPPL (over)estimation: %04f (%04f with padding)", exp2(compressed_bits_per_token - total_pad / (float)inp.size()), exp2(compressed_bits_per_token));
}
// maybe this needs to be changed
if(params.out_file != "imatrix.dat"){
if (params.out_file != "imatrix.dat")
{
// dump uint8array to bin file
std::ofstream ofs(params.out_file.c_str(), std::ios::binary);
ofs.write((char *)&sample_ids_bitpacked[0], sample_ids_bitpacked.size());
ofs.close();
}else{
}
else
{
LOG("\n------------\n");
// print as hex to stdout
for (int i = 0; i < sample_ids_bitpacked.size(); i++){
for (int i = 0; i < sample_ids_bitpacked.size(); i++)
{
LOG("%02X ", sample_ids_bitpacked[i]);
}
}
}
else if (params.compress_mode == 2)
{
@ -575,7 +579,8 @@ int main(int argc, char **argv)
// load sample_ids_bitpacked from params.prompt_file
std::ifstream ifs(params.prompt_file.c_str(), std::ios::binary);
if (!ifs) {
if (!ifs)
{
LOG_ERR("%s: failed to open file\n", __func__);
return -1;
}
@ -588,14 +593,16 @@ int main(int argc, char **argv)
std::vector<uint8_t> sample_ids_bitpacked(fileSize);
// Read the ifs into the vector
if (!ifs.read(reinterpret_cast<char*>(sample_ids_bitpacked.data()), fileSize)) {
if (!ifs.read(reinterpret_cast<char *>(sample_ids_bitpacked.data()), fileSize))
{
LOG_ERR("%s: failed to read file\n", __func__);
return -1;
}
ifs.close();
// Debug: print as hex
for (int i = 0; i < sample_ids_bitpacked.size(); i++){
for (int i = 0; i < sample_ids_bitpacked.size(); i++)
{
LOG("%02X ", sample_ids_bitpacked[i]);
}
LOG("\n");
@ -612,12 +619,12 @@ int main(int argc, char **argv)
std::vector<llama_token> out = decode(ctx, smpl, sample_ids_bitpacked);
gpt_sampler_free(smpl);
auto t_dec_end = ggml_time_us();
// maybe this needs to be changed
if(params.out_file != "imatrix.dat"){
if (params.out_file != "imatrix.dat")
{
// dump as string to file
std::string out_str = ::llama_detokenize(ctx, out);
@ -628,7 +635,6 @@ int main(int argc, char **argv)
llama_free(ctx);
llama_free_model(model);
}
llama_backend_free();