CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-10 10:27:03 +00:00
Code Issues Packages Projects Releases Wiki Activity

quantize : configurable neutral imatrix prior

Browse Source
This commit is contained in:
Francis Couture-Harpin
2025-08-03 16:28:37 -04:00
parent 0a2f5496be
commit 0416ed2bb8
1 changed files with 18 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
tools/quantize/quantize.cpp
View File

@@ -132,6 +132,7 @@ static void usage(const char * executable) {
printf(" --prune-layers L0,L1,L2...comma-separated list of layer numbers to prune from the model\n"); printf(" --prune-layers L0,L1,L2...comma-separated list of layer numbers to prune from the model\n");
printf(" Advanced option to remove all tensors from the given layers\n"); printf(" Advanced option to remove all tensors from the given layers\n");
printf(" --keep-split: will generate quantized model in the same shards as input\n"); printf(" --keep-split: will generate quantized model in the same shards as input\n");
printf(" --prior-weight N: how many tokens the neutral prior is worth (when using imatrix)\n");
printf(" --override-kv KEY=TYPE:VALUE\n"); printf(" --override-kv KEY=TYPE:VALUE\n");
printf(" Advanced option to override model metadata by key in the quantized model. May be specified multiple times.\n"); printf(" Advanced option to override model metadata by key in the quantized model. May be specified multiple times.\n");
printf("Note: --include-weights and --exclude-weights cannot be used together\n"); printf("Note: --include-weights and --exclude-weights cannot be used together\n");
@@ -213,7 +214,7 @@ static int load_legacy_imatrix(const std::string & imatrix_file, std::vector<std
return m_last_call; return m_last_call;
} }
static int load_imatrix(const std::string & imatrix_file, std::vector<std::string> & imatrix_datasets, std::unordered_map<std::string, std::vector<float>> & imatrix_data) { static int load_imatrix(const std::string & imatrix_file, std::vector<std::string> & imatrix_datasets, std::unordered_map<std::string, std::vector<float>> & imatrix_data, float prior_weight) {
struct ggml_context * ctx = nullptr; struct ggml_context * ctx = nullptr;
struct gguf_init_params meta_gguf_params = { struct gguf_init_params meta_gguf_params = {
@@ -289,7 +290,7 @@ static int load_imatrix(const std::string & imatrix_file, std::vector<std::strin
const float count = ((const float *) counts->data)[j]; const float count = ((const float *) counts->data)[j];
if (count > 0.0f) { if (count > 0.0f) {
for (int64_t i = 0; i < ne0; ++i) { for (int64_t i = 0; i < ne0; ++i) {
e[j*ne0 + i] = ((const float *) sums->data)[j*ne0 + i] / count; e[j*ne0 + i] = (((const float *) sums->data)[j*ne0 + i] + prior_weight) / (count + prior_weight);
} }
} else { } else {
// Partial imatrix data, this tensor never got any input during calibration // Partial imatrix data, this tensor never got any input during calibration
@@ -331,10 +332,11 @@ static int prepare_imatrix(const std::string & imatrix_file,
std::vector<std::string> & imatrix_dataset, std::vector<std::string> & imatrix_dataset,
const std::vector<std::string> & included_weights, const std::vector<std::string> & included_weights,
const std::vector<std::string> & excluded_weights, const std::vector<std::string> & excluded_weights,
std::unordered_map<std::string, std::vector<float>> & imatrix_data) { std::unordered_map<std::string, std::vector<float>> & imatrix_data,
float prior_weight) {
int m_last_call = -1; int m_last_call = -1;
if (!imatrix_file.empty()) { if (!imatrix_file.empty()) {
m_last_call = load_imatrix(imatrix_file, imatrix_dataset, imatrix_data); m_last_call = load_imatrix(imatrix_file, imatrix_dataset, imatrix_data, prior_weight);
} }
if (imatrix_data.empty()) { if (imatrix_data.empty()) {
return m_last_call; return m_last_call;
@@ -452,6 +454,7 @@ int main(int argc, char ** argv) {
std::vector<llama_model_kv_override> kv_overrides; std::vector<llama_model_kv_override> kv_overrides;
std::vector<tensor_quantization> tensor_types; std::vector<tensor_quantization> tensor_types;
std::vector<int> prune_layers; std::vector<int> prune_layers;
float prior_weight = 1.0f;
for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) { for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) {
if (strcmp(argv[arg_idx], "--leave-output-tensor") == 0) { if (strcmp(argv[arg_idx], "--leave-output-tensor") == 0) {
@@ -510,6 +513,16 @@ int main(int argc, char ** argv) {
} }
} else if (strcmp(argv[arg_idx], "--keep-split") == 0) { } else if (strcmp(argv[arg_idx], "--keep-split") == 0) {
params.keep_split = true; params.keep_split = true;
} else if (strcmp(argv[arg_idx], "--prior-weight") == 0) {
if (arg_idx < argc-1) {
try {
prior_weight = std::stof(argv[++arg_idx]);
} catch (...) {
usage(argv[0]);
}
} else {
usage(argv[0]);
}
} else { } else {
usage(argv[0]); usage(argv[0]);
} }
@@ -525,7 +538,7 @@ int main(int argc, char ** argv) {
std::vector<std::string> imatrix_datasets; std::vector<std::string> imatrix_datasets;
std::unordered_map<std::string, std::vector<float>> imatrix_data; std::unordered_map<std::string, std::vector<float>> imatrix_data;
int m_last_call = prepare_imatrix(imatrix_file, imatrix_datasets, included_weights, excluded_weights, imatrix_data); int m_last_call = prepare_imatrix(imatrix_file, imatrix_datasets, included_weights, excluded_weights, imatrix_data, prior_weight);
if (!imatrix_data.empty()) { if (!imatrix_data.empty()) {
params.imatrix = &imatrix_data; params.imatrix = &imatrix_data;
{ {