tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-30 12:55:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

imatrix : two-way conversion between old format and GGUF

Browse Source
This commit is contained in:
Francis Couture-Harpin
2025-04-15 17:29:57 -04:00
parent 16202d6f96
commit a5165a6ca9
2 changed files with 365 additions and 73 deletions
Download Patch File Download Diff File Expand all files Collapse all files

293
examples/imatrix/imatrix.cpp
View File

@ -12,6 +12,7 @@
#include <thread>
#include <mutex>
#include <vector>
#include <fstream>
#include <unordered_map>
#include <map>
#include <algorithm>
@ -29,15 +30,19 @@ static void print_usage(int, char ** argv) {
LOG("\n");
}
static bool str_has_suffix(const std::string & str, const std::string & suffix) {
return str.size() >= suffix.size() && str.compare(str.size() - suffix.size(), str.size(), suffix) == 0;
}
static bool str_remove_suffix(std::string & str, const std::string & suffix) {
bool has_suffix = str.size() >= suffix.size() && str.compare(str.size() - suffix.size(), str.size(), suffix) == 0;
bool has_suffix = str_has_suffix(str, suffix);
if (has_suffix) {
str = str.substr(0, str.size() - suffix.size());
}
return has_suffix;
}
static const char * const LLM_KV_IMATRIX_DATASET = "imatrix.dataset";
static const char * const LLM_KV_IMATRIX_DATASETS = "imatrix.datasets";
static const char * const LLM_KV_IMATRIX_CHUNK_COUNT = "imatrix.chunk_count";
static const char * const LLM_KV_IMATRIX_CHUNK_SIZE = "imatrix.chunk_size";
@ -51,12 +56,15 @@ public:
IMatrixCollector() = default;
void set_params(common_params params) { m_params = std::move(params); }
bool collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data);
void save_imatrix_legacy(int32_t ncall = -1) const;
void save_imatrix(int32_t n_chunk = -1) const;
bool load_imatrix_legacy(const char * fname);
bool load_imatrix(const char * file_name);
private:
std::unordered_map<std::string, Stats> m_stats;
common_params m_params;
std::mutex m_mutex;
std::vector<std::string> m_datasets;
int32_t m_last_chunk = 0;
std::vector<float> m_src1_data;
std::vector<char> m_ids; // the expert ids from ggml_mul_mat_id
@ -88,6 +96,8 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
const struct ggml_tensor * src1 = t->src[1];
std::string wname = filter_tensor_name(src0->name);
const int32_t chunk_size = m_params.n_ctx / m_params.n_parallel;
// when ask is true, the scheduler wants to know if we are interested in data from this tensor
// if we return true, a follow-up call will be made with ask=false in which we can do the actual collection
if (ask) {
@ -175,7 +185,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
}
}
}
const int32_t n_chunk = e.counts[ex] / (m_params.n_ctx / m_params.n_parallel);
const int32_t n_chunk = e.counts[ex] / chunk_size;
if (n_chunk > m_last_chunk) {
const int32_t chunk_step = n_chunk - m_last_chunk;
m_last_chunk = n_chunk;
@ -214,7 +224,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
}
}
}
const int32_t n_chunk = e.counts[0] / (m_params.n_ctx / m_params.n_parallel);
const int32_t n_chunk = e.counts[0] / chunk_size;
if (n_chunk > m_last_chunk) {
const int32_t chunk_step = n_chunk - m_last_chunk;
m_last_chunk = n_chunk;
@ -230,19 +240,19 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
return true;
}
void IMatrixCollector::save_imatrix(int32_t n_chunk) const {
void IMatrixCollector::save_imatrix_legacy(int32_t ncall) const {
auto fname = m_params.out_file;
if (n_chunk > 0) {
if (ncall > 0) {
fname += ".at_";
fname += std::to_string(n_chunk);
fname += std::to_string(ncall);
}
// avoid writing imatrix entries that do not have full data
// this can happen with MoE models where some of the experts end up not being exercised by the provided training data
int n_entries = 0;
std::vector<std::string> to_store;
size_t data_size = 0;
bool is_first = true; // for printing
for (const auto & kv : m_stats) {
@ -274,9 +284,8 @@ void IMatrixCollector::save_imatrix(int32_t n_chunk) const {
continue;
}
n_entries++;
to_store.push_back(kv.first);
data_size += GGML_PAD(ggml_tensor_overhead() + sizeof(float) * kv.second.values.size(), GGML_MEM_ALIGN);
data_size += GGML_PAD(ggml_tensor_overhead() + sizeof(float) * kv.second.counts.size(), GGML_MEM_ALIGN);
}
if (to_store.size() < m_stats.size()) {
@ -286,6 +295,79 @@ void IMatrixCollector::save_imatrix(int32_t n_chunk) const {
// deterministic tensor name order
std::sort(to_store.begin(), to_store.end());
const int32_t chunk_size = m_params.n_ctx / m_params.n_parallel;
std::ofstream out(fname, std::ios::binary);
out.write((const char *) &n_entries, sizeof(n_entries));
for (const auto & name : to_store) {
const auto & stat = m_stats.at(name);
const int32_t len = name.size();
out.write((const char *) &len, sizeof(len));
out.write(name.c_str(), len);
const int32_t ncall = *std::max_element(stat.counts.begin(), stat.counts.end()) / chunk_size;
out.write((const char *) &ncall, sizeof(ncall));
const int32_t nval = stat.values.size();
const int32_t nmat = stat.counts.size();
out.write((const char *) &nval, sizeof(nval));
if (nval > 0 && nmat > 0) {
std::vector<float> tmp(nval);
for (int32_t i = 0; i < nval; i++) {
const float counts = static_cast<float>(stat.counts[i / (nval / nmat)]);
tmp[i] = (stat.values[i] / counts) * static_cast<float>(ncall);
}
out.write((const char *) tmp.data(), nval * sizeof(float));
}
}
// Write the number of call the matrix was computed with
out.write((const char *) &m_last_chunk, sizeof(m_last_chunk));
// Write the input filename at the end of the file to later on specify it in quantize
{
const char * dataset_file = m_params.prompt_file.c_str();
int32_t len = m_params.prompt_file.size();
// When there is no prompt but there were other imatrix files loaded, use the last dataset
if (m_params.prompt_file.empty() && !m_datasets.empty()) {
const std::string & dataset_str = m_datasets[m_datasets.size() - 1];
dataset_file = dataset_str.c_str();
len = dataset_str.size();
}
out.write((const char *) &len, sizeof(len));
out.write(dataset_file, len);
}
LOGV(1, "\n");
LOG_DBGV(1, "%s: stored collected data after %d chunks in %s\n", __func__, m_last_chunk, fname.c_str());
}
void IMatrixCollector::save_imatrix(int32_t n_chunk) const {
auto fname = m_params.out_file;
// TODO: use the new format by default also for .imatrix
if (!str_has_suffix(fname, ".gguf")) {
return this->save_imatrix_legacy(n_chunk);
}
if (n_chunk > 0) {
fname += ".at_";
fname += std::to_string(n_chunk);
}
// write imatrix entries even if they don't have full data. (can be corrected when reading)
// this can happen with MoE models where some of the experts end up not being exercised by the provided training data
std::vector<std::string> to_store;
size_t data_size = 0;
for (const auto & kv : m_stats) {
to_store.push_back(kv.first);
data_size += GGML_PAD(ggml_tensor_overhead() + sizeof(float) * kv.second.values.size(), GGML_MEM_ALIGN);
data_size += GGML_PAD(ggml_tensor_overhead() + sizeof(float) * kv.second.counts.size(), GGML_MEM_ALIGN);
}
// deterministic tensor name order
std::sort(to_store.begin(), to_store.end());
struct ggml_init_params params = {
/* .mem_size = */ data_size,
/* .mem_buffer = */ NULL,
@ -294,31 +376,42 @@ void IMatrixCollector::save_imatrix(int32_t n_chunk) const {
struct ggml_context * ctx = ggml_init(params);
struct gguf_context * ctx_gguf = gguf_init_empty();
gguf_set_val_str(ctx_gguf, "general.type", "imatrix");
// Write the input filename to later on specify it in quantize
gguf_set_val_str(ctx_gguf, LLM_KV_IMATRIX_DATASET, m_params.prompt_file.c_str());
// Write the number of chunks the matrix was computed with
gguf_set_val_u32(ctx_gguf, LLM_KV_IMATRIX_CHUNK_COUNT, m_last_chunk);
gguf_set_val_u32(ctx_gguf, LLM_KV_IMATRIX_CHUNK_SIZE, m_params.n_ctx / m_params.n_parallel);
{
std::vector<const char *> datasets;
datasets.reserve(m_datasets.size() + 1);
for (size_t i = 0; i < m_datasets.size(); ++i) {
datasets.push_back(m_datasets[i].c_str());
}
if (!m_params.prompt_file.empty()) {
datasets.push_back(m_params.prompt_file.c_str());
}
gguf_set_val_str(ctx_gguf, "general.type", "imatrix");
// Write the dataset paths
gguf_set_arr_str(ctx_gguf, LLM_KV_IMATRIX_DATASETS, datasets.data(), datasets.size());
// Write the number of chunks the matrix was computed with
gguf_set_val_u32(ctx_gguf, LLM_KV_IMATRIX_CHUNK_COUNT, m_last_chunk);
gguf_set_val_u32(ctx_gguf, LLM_KV_IMATRIX_CHUNK_SIZE, m_params.n_ctx / m_params.n_parallel);
}
for (const auto & name : to_store) {
const auto & stat = m_stats.at(name);
const int32_t nval = (int32_t) stat.values.size();
const int32_t nmat = (int32_t) stat.counts.size();
if (nval > 0) {
struct ggml_tensor * sums = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, nval / nmat, nmat);
struct ggml_tensor * counts = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 1, nmat);
ggml_format_name(sums, "%s.sums", name.c_str());
if (nval > 0 && nmat > 0) {
struct ggml_tensor * in_sum2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, nval / nmat, nmat);
struct ggml_tensor * counts = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 1, nmat);
ggml_format_name(in_sum2, "%s.in_sum2", name.c_str());
ggml_format_name(counts, "%s.counts", name.c_str());
for (int32_t j = 0; j < nval; ++j) {
((float *) sums->data)[j] = (float) stat.values[j];
((float *) in_sum2->data)[j] = (float) stat.values[j];
}
for (int32_t j = 0; j < nmat; ++j) {
((float *) counts->data)[j] = (float) stat.counts[j];
}
gguf_add_tensor(ctx_gguf, sums);
gguf_add_tensor(ctx_gguf, in_sum2);
gguf_add_tensor(ctx_gguf, counts);
}
}
@ -332,6 +425,105 @@ void IMatrixCollector::save_imatrix(int32_t n_chunk) const {
ggml_free(ctx);
}
bool IMatrixCollector::load_imatrix_legacy(const char * fname) {
std::ifstream in(fname, std::ios::binary);
if (!in) {
LOG_ERR("%s: failed to open %s\n", __func__, fname);
return false;
}
int n_entries;
in.read((char *) &n_entries, sizeof(n_entries));
if (in.fail() || n_entries < 1) {
LOG_ERR("%s: no data in file %s\n", __func__, fname);
return false;
}
// Guess the chunk size because it's not stored in the file
const int32_t chunk_size = m_params.n_ctx / m_params.n_parallel;
for (int i = 0; i < n_entries; ++i) {
int32_t len = 0;
in.read((char *) &len, sizeof(len));
std::vector<char> name_as_vec(len + 1);
in.read((char *) name_as_vec.data(), len);
if (in.fail()) {
LOG_ERR("%s: failed reading name for entry %d from %s\n", __func__, i + 1, fname);
return false;
}
name_as_vec[len] = 0;
std::string name{ name_as_vec.data() };
auto & e = m_stats[std::move(name)];
int32_t ncall = 0;
in.read((char *) &ncall, sizeof(ncall));
int32_t nval = 0;
in.read((char *) &nval, sizeof(nval));
if (in.fail() || nval < 1) {
LOG_ERR("%s: failed reading number of values for entry %d\n", __func__, i);
m_stats = {};
return false;
}
if (e.values.empty()) {
e.values.resize(nval, 0.0f);
e.counts.resize(1, 0);
}
std::vector<float> tmp(nval);
in.read((char *) tmp.data(), nval * sizeof(float));
if (in.fail()) {
LOG_ERR("%s: failed reading data for entry %d\n", __func__, i);
m_stats = {};
return false;
}
// Recreate the state as expected by save_imatrix(), and correct for weighted sum.
for (int i = 0; i < nval; i++) {
e.values[i] += tmp[i] * chunk_size;
}
// The legacy format doesn't distinguish the counts for different experts
for (size_t j = 0; j < e.counts.size(); ++j) {
e.counts[j] += ncall * chunk_size;
}
}
{
// TODO: extract into its own method; this is also used by the GGUF-based format
// Calculate the last chunk count
int64_t max_count = 0;
for (const auto & stats : m_stats) {
for (int64_t count : stats.second.counts) {
if (count > max_count) {
max_count = count;
}
}
}
m_last_chunk = max_count / (chunk_size);
}
{
// Read the number of calls the matrix was computed with
int32_t n_calls;
in.read((char *) &n_calls, sizeof(n_calls));
// ignore it because it's not important
}
// Read the dataset path to include it when writing to GGUF
if (!in.fail()){
int32_t len = 0;
in.read((char *) &len, sizeof(len));
if (!in.fail()) {
std::vector<char> dataset;
dataset.resize(len + 1, 0);
in.read(dataset.data(), len);
if (!in.fail()) {
m_datasets.push_back(dataset.data());
}
}
}
return true;
}
// Using GGUF as the file format, for greater extensibility
bool IMatrixCollector::load_imatrix(const char * file_name) {
struct ggml_context * ctx = nullptr;
struct gguf_init_params meta_gguf_params = {
@ -340,7 +532,7 @@ bool IMatrixCollector::load_imatrix(const char * file_name) {
};
struct gguf_context * ctx_gguf = gguf_init_from_file(file_name, meta_gguf_params);
if (!ctx_gguf) {
return false;
return this->load_imatrix_legacy(file_name);
}
const int32_t n_entries = gguf_get_n_tensors(ctx_gguf);
if (n_entries < 1) {
@ -350,8 +542,17 @@ bool IMatrixCollector::load_imatrix(const char * file_name) {
return false;
}
const std::string sums_suffix{".sums"};
const std::string counts_suffix{".counts"};
const int64_t datasets_key = gguf_find_key(ctx_gguf, LLM_KV_IMATRIX_DATASETS);
if (datasets_key != -1 && gguf_get_arr_type(ctx_gguf, datasets_key) == GGUF_TYPE_STRING) {
const int64_t n = gguf_get_arr_n(ctx_gguf, datasets_key);
m_datasets.reserve(m_datasets.size() + n);
for (int64_t i = 0; i < n; ++i) {
m_datasets.push_back(gguf_get_arr_str(ctx_gguf, datasets_key, i));
}
}
const std::string in_sum2_suffix{ ".in_sum2" };
const std::string counts_suffix{ ".counts" };
// Could re-use m_stats instead, but this allows
// checking for completeness of *each* loaded imatrix file
@ -364,26 +565,23 @@ bool IMatrixCollector::load_imatrix(const char * file_name) {
if (name.empty()) { continue; }
if (str_remove_suffix(name, sums_suffix)) {
// sums
sums_counts_for[name].first = cur;
if (str_remove_suffix(name, in_sum2_suffix)) {
// in_sum2
sums_counts_for[std::move(name)].first = cur;
} else if (str_remove_suffix(name, counts_suffix)) {
// counts
sums_counts_for[name].second = cur;
sums_counts_for[std::move(name)].second = cur;
} else {
LOG_ERR("%s: invalid imatrix tensor name: %s\n", __func__, name.c_str());
gguf_free(ctx_gguf);
ggml_free(ctx);
return false;
// ignore other tensors
}
}
for (const auto & sc : sums_counts_for) {
const std::string & name = sc.first;
const struct ggml_tensor * sums = sc.second.first;
const struct ggml_tensor * counts = sc.second.second;
const std::string & name = sc.first;
const struct ggml_tensor * in_sum2 = sc.second.first;
const struct ggml_tensor * counts = sc.second.second;
if (!sums || !counts) {
if (!in_sum2 || !counts) {
LOG_ERR("%s: mismatched sums and counts for %s\n", __func__, name.c_str());
gguf_free(ctx_gguf);
ggml_free(ctx);
@ -392,9 +590,9 @@ bool IMatrixCollector::load_imatrix(const char * file_name) {
auto & e = m_stats[name];
int64_t nval = ggml_nelements(sums);
int64_t nval = ggml_nelements(in_sum2);
if (e.values.empty()) {
e.values.resize(nval, 0);
e.values.resize(nval, 0.0f);
} else if ((size_t) nval != e.values.size()) {
LOG_ERR("%s: mismatched sums size for %s: %zu != %zu\n", __func__, name.c_str(), (size_t) nval, e.values.size());
gguf_free(ctx_gguf);
@ -417,12 +615,25 @@ bool IMatrixCollector::load_imatrix(const char * file_name) {
// Recreate the state as expected by save_imatrix()
for (int64_t j = 0; j < nval; j++) {
e.values[j] += ((const float *) sums->data)[j];
e.values[j] += ((const float *) in_sum2->data)[j];
}
for (int64_t j = 0; j < ncounts; j++) {
e.counts[j] += std::lround(((const float *) counts->data)[j]);
}
}
// TODO: extract into its own method; this is also used by the legacy format
// Calculate the last chunk count
int64_t max_count = 0;
for (const auto & stats : m_stats) {
for (int64_t count : stats.second.counts) {
if (count > max_count) {
max_count = count;
}
}
}
m_last_chunk = max_count / (m_params.n_ctx / m_params.n_parallel);
gguf_free(ctx_gguf);
ggml_free(ctx);
return true;
@ -685,7 +896,7 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params, c
int main(int argc, char ** argv) {
common_params params;
params.out_file = "imatrix.dat" ;
params.out_file = "imatrix.gguf" ;
params.n_ctx = 512;
params.logits_all = true;

145
examples/quantize/quantize.cpp
View File

@ -64,7 +64,7 @@ static const char * const LLM_KV_QUANTIZE_IMATRIX_N_ENTRIES = "quantize.imatrix
static const char * const LLM_KV_QUANTIZE_IMATRIX_N_CHUNKS = "quantize.imatrix.chunks_count";
// TODO: share with imatrix.cpp
static const char * const LLM_KV_IMATRIX_DATASET = "imatrix.dataset";
static const char * const LLM_KV_IMATRIX_DATASETS = "imatrix.datasets";
static const char * const LLM_KV_IMATRIX_CHUNK_COUNT = "imatrix.chunk_count";
static const char * const LLM_KV_IMATRIX_CHUNK_SIZE = "imatrix.chunk_size";
@ -84,7 +84,7 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
for (auto ch : ftype_str_in) {
ftype_str.push_back(std::toupper(ch));
}
for (auto & it : QUANT_OPTIONS) {
for (const auto & it : QUANT_OPTIONS) {
if (striequals(it.name.c_str(), ftype_str.c_str())) {
ftype = it.ftype;
ftype_str_out = it.name;
@ -93,7 +93,7 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
}
try {
int ftype_int = std::stoi(ftype_str);
for (auto & it : QUANT_OPTIONS) {
for (const auto & it : QUANT_OPTIONS) {
if (it.ftype == ftype_int) {
ftype = it.ftype;
ftype_str_out = it.name;
@ -126,7 +126,7 @@ static void usage(const char * executable) {
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("\nAllowed quantization types:\n");
for (auto & it : QUANT_OPTIONS) {
for (const auto & it : QUANT_OPTIONS) {
if (it.name != "COPY") {
printf(" %2d or ", it.ftype);
} else {
@ -146,7 +146,71 @@ static bool str_remove_suffix(std::string & str, const std::string & suffix) {
return has_suffix;
}
static int load_imatrix(const std::string & imatrix_file, std::string & imatrix_dataset, std::unordered_map<std::string, std::vector<float>> & imatrix_data) {
static int load_legacy_imatrix(const std::string & imatrix_file, std::vector<std::string> & imatrix_datasets, std::unordered_map<std::string, std::vector<float>> & imatrix_data) {
std::ifstream in(imatrix_file.c_str(), std::ios::binary);
if (!in) {
printf("%s: failed to open %s\n",__func__, imatrix_file.c_str());
exit(1);
}
int n_entries;
in.read((char *)&n_entries, sizeof(n_entries));
if (in.fail() || n_entries < 1) {
printf("%s: no data in file %s\n", __func__, imatrix_file.c_str());
exit(1);
}
for (int i = 0; i < n_entries; ++i) {
int len; in.read((char *)&len, sizeof(len));
std::vector<char> name_as_vec(len+1);
in.read((char *)name_as_vec.data(), len);
if (in.fail()) {
printf("%s: failed reading name for entry %d from %s\n", __func__, i+1, imatrix_file.c_str());
exit(1);
}
name_as_vec[len] = 0;
std::string name{name_as_vec.data()};
auto & e = imatrix_data[name];
int ncall;
in.read((char *)&ncall, sizeof(ncall));
int nval;
in.read((char *)&nval, sizeof(nval));
if (in.fail() || nval < 1) {
printf("%s: failed reading number of values for entry %d\n", __func__, i);
imatrix_data = {};
exit(1);
}
e.resize(nval);
in.read((char *)e.data(), nval*sizeof(float));
if (in.fail()) {
printf("%s: failed reading data for entry %d\n", __func__, i);
imatrix_data = {};
exit(1);
}
if (ncall > 0) {
for (auto& v : e) v /= ncall;
}
if (getenv("LLAMA_TRACE")) {
printf("%s: loaded data (size = %6d, ncall = %6d) for '%s'\n", __func__, int(e.size()), ncall, name.c_str());
}
}
// latest imatrix version contains the dataset filename at the end of the file
int m_last_call = 0;
if (in.peek() != EOF) {
in.read((char *)&m_last_call, sizeof(m_last_call));
int dataset_len;
in.read((char *)&dataset_len, sizeof(dataset_len));
std::vector<char> dataset_as_vec(dataset_len);
in.read(dataset_as_vec.data(), dataset_len);
imatrix_datasets.resize(1);
imatrix_datasets[0].assign(dataset_as_vec.begin(), dataset_as_vec.end());
printf("%s: imatrix dataset='%s'\n", __func__, imatrix_datasets[0].c_str());
}
printf("%s: loaded %d importance matrix entries from %s computed on %d chunks\n", __func__, int(imatrix_data.size()), imatrix_file.c_str(), 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) {
struct ggml_context * ctx = nullptr;
struct gguf_init_params meta_gguf_params = {
@ -155,8 +219,8 @@ static int load_imatrix(const std::string & imatrix_file, std::string & imatrix_
};
struct gguf_context * ctx_gguf = gguf_init_from_file(imatrix_file.c_str(), meta_gguf_params);
if (!ctx_gguf) {
fprintf(stderr, "%s: if this is an older imatrix file, make sure to convert it to the GGUF-based imatrix format\n", __func__);
exit(1);
fprintf(stderr, "%s: imatrix file '%s' is using old format\n", __func__, imatrix_file.c_str());
return load_legacy_imatrix(imatrix_file, imatrix_datasets, imatrix_data);
}
const int32_t n_entries = gguf_get_n_tensors(ctx_gguf);
if (n_entries < 1) {
@ -166,7 +230,7 @@ static int load_imatrix(const std::string & imatrix_file, std::string & imatrix_
exit(1);
}
const int dataset_idx = gguf_find_key(ctx_gguf, LLM_KV_IMATRIX_DATASET);
const int dataset_idx = gguf_find_key(ctx_gguf, LLM_KV_IMATRIX_DATASETS);
const int chunk_count_idx = gguf_find_key(ctx_gguf, LLM_KV_IMATRIX_CHUNK_COUNT);
const int chunk_size_idx = gguf_find_key(ctx_gguf, LLM_KV_IMATRIX_CHUNK_SIZE);
if (dataset_idx < 0 || chunk_count_idx < 0 || chunk_size_idx < 0) {
@ -178,8 +242,8 @@ static int load_imatrix(const std::string & imatrix_file, std::string & imatrix_
const uint32_t chunk_size = gguf_get_val_u32(ctx_gguf, chunk_size_idx);
const std::string sums_suffix{".sums"};
const std::string counts_suffix{".counts"};
const std::string sums_suffix{ ".in_sum2" };
const std::string counts_suffix{ ".counts" };
// Using an ordered map to get a deterministic iteration order.
std::map<std::string, std::pair<struct ggml_tensor *, struct ggml_tensor *>> sums_counts_for;
@ -190,16 +254,13 @@ static int load_imatrix(const std::string & imatrix_file, std::string & imatrix_
if (name.empty()) { continue; }
if (str_remove_suffix(name, sums_suffix)) {
// sums
sums_counts_for[name].first = cur;
// in_sum2
sums_counts_for[std::move(name)].first = cur;
} else if (str_remove_suffix(name, counts_suffix)) {
// counts
sums_counts_for[name].second = cur;
sums_counts_for[std::move(name)].second = cur;
} else {
fprintf(stderr, "%s: invalid imatrix tensor name: %s\n", __func__, name.c_str());
gguf_free(ctx_gguf);
ggml_free(ctx);
exit(1);
// ignore other tensors
}
}
@ -223,8 +284,15 @@ static int load_imatrix(const std::string & imatrix_file, std::string & imatrix_
float max_count = 0.0f;
for (int64_t j = 0; j < ne1; ++j) {
const float count = ((const float *) counts->data)[j];
for (int64_t i = 0; i < ne0; ++i) {
e[j*ne0 + i] = ((const float *) sums->data)[j*ne0 + i] / count;
if (count > 0.0f) {
for (int64_t i = 0; i < ne0; ++i) {
e[j*ne0 + i] = ((const float *) sums->data)[j*ne0 + i] / count;
}
} else {
// Partial imatrix data, this tensor never got any input during calibration
for (int64_t i = 0; i < ne0; ++i) {
e[j*ne0 + i] = 1;
}
}
if (count > max_count) {
max_count = count;
@ -236,9 +304,18 @@ static int load_imatrix(const std::string & imatrix_file, std::string & imatrix_
}
int m_last_chunk = gguf_get_val_u32(ctx_gguf, chunk_count_idx);
imatrix_dataset = gguf_get_val_str(ctx_gguf, dataset_idx);
printf("%s: imatrix dataset='%s'\n", __func__, imatrix_dataset.c_str());
int64_t n_datasets = gguf_get_arr_n(ctx_gguf, dataset_idx);
imatrix_datasets.resize(n_datasets);
for (int64_t i = 0; i < n_datasets; ++i) {
imatrix_datasets.push_back(gguf_get_val_str(ctx_gguf, dataset_idx));
}
printf("%s: imatrix datasets=['%s'", __func__, imatrix_datasets[0].c_str());
for (size_t i = 1; i < imatrix_datasets.size(); ++i) {
printf(", '%s'", imatrix_datasets[i].c_str());
}
printf("]\n");
printf("%s: loaded %d importance matrix entries from %s computed on %d chunks\n", __func__, int(imatrix_data.size()), imatrix_file.c_str(), m_last_chunk);
gguf_free(ctx_gguf);
@ -248,7 +325,7 @@ static int load_imatrix(const std::string & imatrix_file, std::string & imatrix_
}
static int prepare_imatrix(const std::string & imatrix_file,
std::string & imatrix_dataset,
std::vector<std::string> & imatrix_dataset,
const std::vector<std::string> & included_weights,
const std::vector<std::string> & excluded_weights,
std::unordered_map<std::string, std::vector<float>> & imatrix_data) {
@ -260,18 +337,21 @@ static int prepare_imatrix(const std::string & imatrix_file,
return m_last_call;
}
if (!excluded_weights.empty()) {
for (auto& name : excluded_weights) {
for (auto it = imatrix_data.begin(); it != imatrix_data.end(); ) {
for (const auto & name : excluded_weights) {
for (auto it = imatrix_data.begin(); it != imatrix_data.end();) {
auto pos = it->first.find(name);
if (pos != std::string::npos) it = imatrix_data.erase(it);
else ++it;
if (pos != std::string::npos) {
it = imatrix_data.erase(it);
} else {
++it;
}
}
}
}
if (!included_weights.empty()) {
std::unordered_map<std::string, std::vector<float>> tmp;
for (auto& name : included_weights) {
for (auto& e : imatrix_data) {
for (const auto & name : included_weights) {
for (auto & e : imatrix_data) {
auto pos = e.first.find(name);
if (pos != std::string::npos) {
tmp.emplace(std::move(e));
@ -372,9 +452,9 @@ int main(int argc, char ** argv) {
usage(argv[0]);
}
std::string imatrix_dataset;
std::vector<std::string> imatrix_datasets;
std::unordered_map<std::string, std::vector<float>> imatrix_data;
int m_last_call = prepare_imatrix(imatrix_file, imatrix_dataset, included_weights, excluded_weights, imatrix_data);
int m_last_call = prepare_imatrix(imatrix_file, imatrix_datasets, included_weights, excluded_weights, imatrix_data);
if (!imatrix_data.empty()) {
params.imatrix = &imatrix_data;
{
@ -385,11 +465,12 @@ int main(int argc, char ** argv) {
kvo.val_str[127] = '\0';
kv_overrides.emplace_back(std::move(kvo));
}
if (!imatrix_dataset.empty()) {
if (!imatrix_datasets.empty()) {
llama_model_kv_override kvo;
// TODO: list multiple datasets when there are more than one
std::strcpy(kvo.key, LLM_KV_QUANTIZE_IMATRIX_DATASET);
kvo.tag = LLAMA_KV_OVERRIDE_TYPE_STR;
strncpy(kvo.val_str, imatrix_dataset.c_str(), 127);
strncpy(kvo.val_str, imatrix_datasets[0].c_str(), 127);
kvo.val_str[127] = '\0';
kv_overrides.emplace_back(std::move(kvo));
}