Felipe97/llama-cpp-compiled
01.2k
1#include "arg.h"2#include "common.h"3#include "imatrix-loader.h"4#include "log.h"5#include "llama.h"6#include "gguf.h"7 8#include <algorithm>9#include <chrono>10#include <clocale>11#include <cmath>12#include <cstdio>13#include <cstring>14#include <ctime>15#include <thread>16#include <mutex>17#include <vector>18#include <fstream>19#include <unordered_map>20#include <map>21#include <regex>22#include <numeric>23 24#if defined(_MSC_VER)25#pragma warning(disable: 4244 4267) // possible loss of data26#endif27 28static void print_usage(int, char ** argv) {29 LOG("\nexample usage:\n");30 LOG("\n %s \\\n"31 " -m model.gguf -f some-text.txt [-o imatrix.gguf] [--output-format {gguf,dat}] [--no-ppl] \\\n"32 " [--process-output] [--chunk 123] [--save-frequency 0] [--output-frequency 10] \\\n"33 " [--in-file imatrix-prev-0.gguf --in-file imatrix-prev-1.gguf ...] [--parse-special] \\\n"34 " [--show-statistics] [...]\n" , argv[0]);35 LOG("\n");36}37 38struct Stats {39 std::vector<float> values;40 std::vector<int64_t> counts;41};42 43struct tensor_statistics {44 std::string tensor;45 Stats stats;46 float total_sqract = 0.0f;47 float mean_sqract = 0.0f;48 float max_sqract = 0.0f;49 float min_sqract = 0.0f;50 int elements = 0;51 float stddev = 0.0f;52 float active = 0.0f;53 float entropy = 0.0f;54 float zd = 0.0f;55 float cossim = 0.0f;56};57 58class IMatrixCollector {59public:60 IMatrixCollector() = default;61 void set_params(common_params params) { m_params = std::move(params); }62 bool collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data);63 void save_imatrix_legacy(int32_t ncall = -1) const;64 void save_imatrix(int32_t n_chunk = -1) const;65 bool load_imatrix(const char * file_name);66 const std::unordered_map<std::string, Stats> & get_mstats() const { return m_stats; }67private:68 std::unordered_map<std::string, Stats> m_stats;69 common_params m_params;70 std::mutex m_mutex;71 std::vector<std::string> m_datasets;72 int32_t m_last_chunk = 0;73 std::vector<char> m_src1_data;74 std::vector<char> m_ids; // the expert ids from ggml_mul_mat_id75};76 77// remove any prefix and suffixes from the name78// CUDA0#blk.0.attn_k.weight#0 => blk.0.attn_k.weight79static std::string filter_tensor_name(const char * name) {80 std::string wname;81 const char * p = strchr(name, '#');82 if (p != NULL) {83 p = p + 1;84 const char * q = strchr(p, '#');85 if (q != NULL) {86 wname = std::string(p, q - p);87 } else {88 wname = p;89 }90 } else {91 wname = name;92 }93 return wname;94}95 96static void process_tensor_name(const std::string & input, std::string & layer, std::string & tensor) {97 std::vector<std::string> name;98 std::istringstream stream(input);99 std::string item;100 101 while (std::getline(stream, item, '.')) {102 name.push_back(item);103 }104 for (size_t i = 0; i < name.size(); ++i) {105 if (name[i] == "blk" && i + 1 < name.size()) {106 layer = name[i + 1];107 break;108 }109 }110 for (size_t i = 0; i < name.size(); ++i) {111 if (name[i] == "weight" && i > 0) {112 tensor = name[i - 1];113 break;114 }115 }116 117 if (tensor.empty()) {118 tensor = input;119 }120 if (layer.empty()) {121 layer = "-";122 }123}124 125static void compute_statistics(std::vector<tensor_statistics> & tstats, const std::string & name, const Stats & e) {126 if (e.values.size() % e.counts.size() != 0) {127 LOG_ERR("%s: activation size mismatch for tensor %s (%zu vs %zu)\n", __func__, name.c_str(), e.counts.size(), e.values.size());128 return;129 }130 if (e.counts.empty()) {131 LOG_ERR("%s: there are no activations for tensor %s. The imatrix may be suboptimal\n", __func__, name.c_str());132 return;133 }134 135 const int n_mat = e.counts.size();136 const int row_size = e.values.size() / n_mat;137 138 std::vector<float> activations;139 activations.reserve(e.values.size());140 141 for (int i = 0; i < n_mat; ++i) {142 if (e.counts[i] == 0) {143 LOG_DBG("%s: skipping tensor %s due to zero count at index %d\n", __func__, name.c_str(), i);144 continue;145 }146 for (int j = 0; j < row_size; ++j) {147 activations.push_back(e.values[i*row_size + j] / e.counts[i]);148 }149 }150 151 if (activations.empty()) {152 LOG_ERR("%s: all counts are zero for tensor %s, skipping statistics computation\n", __func__, name.c_str());153 return;154 }155 156 const float act_total = std::accumulate(activations.begin(), activations.end(), 0.0f);157 const float act_max = *std::max_element(activations.begin(), activations.end());158 const float act_min = *std::min_element(activations.begin(), activations.end());159 const float act_mean = act_total / activations.size();160 const float act_sqr_total = std::inner_product(activations.begin(), activations.end(), activations.begin(), 0.0f);161 const float act_var = (act_sqr_total / activations.size()) - (act_mean * act_mean);162 const float act_dev = std::sqrt(std::max(0.0f, act_var));163 float threshold = 1e-5f;164 const int inactive_count = std::count_if(activations.begin(), activations.end(),165 [threshold](const float v) { return fabsf(v) <= threshold; });166 const float active_ratio = 1 - static_cast<float>(inactive_count) / activations.size();167 168 float entropy = 0;169 if (act_total > 0) {170 for (const auto act : activations) {171 if (const float p = act / act_total; p > 0) {172 entropy -= p * std::log2(p);173 }174 }175 }176 177 int z_score = 0;178 if (act_dev > 0.0f) {179 for (const auto act : activations) {180 if (const float p = (act - act_mean) / act_dev; p > 1) {181 z_score++;182 }183 }184 }185 186 auto & ts = tstats.emplace_back();187 ts.tensor = name;188 ts.stats = e;189 ts.total_sqract = act_total;190 ts.mean_sqract = act_mean;191 ts.max_sqract = act_max;192 ts.min_sqract = act_min;193 ts.elements = static_cast<int>(activations.size());194 ts.stddev = act_dev;195 ts.active = active_ratio;196 ts.entropy = entropy;197 ts.zd = static_cast<float>(z_score) / ts.elements;198}199 200static void compute_cossim(std::vector<tensor_statistics> & tstats) {201 static const std::regex pattern(R"(blk\.(\d+)\.)");202 for (auto & ts : tstats) {203 if (std::smatch match; std::regex_search(ts.tensor, match, pattern)) {204 const int blk = std::stoi(match[1]);205 std::string tname(ts.tensor);206 tname.replace(match.position(1), match.length(1), std::to_string(blk-1));207 auto prev = std::find_if(tstats.begin(), tstats.end(),208 [tname](const tensor_statistics & t) { return t.tensor == tname; });209 if (prev != tstats.end()) {210 const float dp = std::inner_product(ts.stats.values.begin(), ts.stats.values.end(),211 prev->stats.values.begin(), 0.0f);212 const float curr_mag = std::sqrt(std::inner_product(ts.stats.values.begin(), ts.stats.values.end(),213 ts.stats.values.begin(), 0.0f));214 const float prev_mag = std::sqrt(std::inner_product(prev->stats.values.begin(), prev->stats.values.end(),215 prev->stats.values.begin(), 0.0f));216 const float cs = dp / (curr_mag * prev_mag);217 ts.cossim = cs;218 }219 } else {220 ts.cossim = 0;221 }222 }223}224 225static bool all_finite(const float * v, size_t n) {226 for (size_t i = 0; i < n; ++i) {227 if (!std::isfinite(v[i])) {228 return false;229 }230 }231 return true;232}233 234bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data) {235 GGML_UNUSED(user_data);236 237 const struct ggml_tensor * src0 = t->src[0];238 const struct ggml_tensor * src1 = t->src[1];239 std::string wname = filter_tensor_name(src0->name);240 241 const int32_t chunk_size = m_params.n_ctx / m_params.n_parallel;242 243 // when ask is true, the scheduler wants to know if we are interested in data from this tensor244 // if we return true, a follow-up call will be made with ask=false in which we can do the actual collection245 if (ask) {246 if (t->op == GGML_OP_MUL_MAT_ID) return true; // collect all indirect matrix multiplications247 if (t->op != GGML_OP_MUL_MAT) return false;248 // why are small batches ignored (<16 tokens)?249 if (src1->ne[1] < 16 || src1->type != GGML_TYPE_F32) return false;250 if (!(wname.substr(0, 4) == "blk." || (m_params.process_output && wname == "output.weight"))) return false;251 return true;252 }253 254 std::lock_guard<std::mutex> lock(m_mutex);255 256 // copy the data from the GPU memory if needed257 const bool is_host = ggml_backend_buffer_is_host(src1->buffer);258 259 if (!is_host) {260 const size_t src1_nbytes = ggml_nbytes(src1);261 m_src1_data.resize(src1_nbytes);262 ggml_backend_tensor_get(src1, m_src1_data.data(), 0, src1_nbytes);263 }264 265 const char * data = is_host ? (const char *) src1->data : m_src1_data.data();266 GGML_ASSERT(src1->nb[0] == ggml_element_size(src1));267 268 // this has been adapted to the new format of storing merged experts in a single 3d tensor269 // ref: https://github.com/ggml-org/llama.cpp/pull/6387270 if (t->op == GGML_OP_MUL_MAT_ID) {271 // ids -> [n_experts_used, n_tokens]272 // src1 -> [cols, n_expert_used, n_tokens]273 const ggml_tensor * ids = t->src[2];274 const int64_t n_as = src0->ne[2];275 const int64_t n_ids = ids->ne[0];276 277 // the top-k selected expert ids are stored in the ids tensor278 // for simplicity, always copy ids to host, because it is small279 // take into account that ids is not contiguous!280 281 GGML_ASSERT(ids->ne[1] == src1->ne[2]);282 283 // the extra dimension would need to be stored somewhere to be reflected in the imatrix file284 if (ggml_nrows(src1) != src1->ne[1] * src1->ne[2]) {285 LOG_ERR("%s: tensor has more than 3 dimensions: %s", __func__, wname.c_str());286 GGML_ASSERT(false);287 }288 289 m_ids.resize(ggml_nbytes(ids));290 ggml_backend_tensor_get(ids, m_ids.data(), 0, ggml_nbytes(ids));291 292 auto & e = m_stats[wname];293 294 if (e.counts.size() == 1 && n_as > 1) {295 // broadcast, when loading an old imatrix296 e.counts.resize(n_as, e.counts[0]);297 }298 if (e.values.empty()) {299 e.values.resize(src1->ne[0]*n_as, 0);300 e.counts.resize(n_as, 0);301 }302 else if (e.values.size() != (size_t)src1->ne[0]*n_as) {303 LOG_ERR("%s: inconsistent size for %s (%d vs %d)\n", __func__, wname.c_str(), (int)e.values.size(), (int)(src1->ne[0]*n_as));304 exit(1); //GGML_ABORT("fatal error");305 }306 else if (e.counts.size() != (size_t)n_as) {307 LOG_ERR("%s: inconsistent expert count for %s (%d vs %d)\n", __func__, wname.c_str(), (int)e.counts.size(), (int)n_as);308 exit(1); //GGML_ABORT("fatal error");309 }310 LOG_DBGV(2, "%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_chunk, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[2], (int)src1->type);311 312 const int64_t ne0 = src1->ne[0];313 const int64_t n_tokens = src1->ne[2];314 315 // single pass over the routing ids316 std::vector<uint8_t> touched(n_as, 0);317 for (int64_t idx = 0; idx < n_ids; ++idx) {318 for (int64_t row = 0; row < n_tokens; ++row) {319 const int32_t ex = *(const int32_t *) (m_ids.data() + row * ids->nb[1] + idx * ids->nb[0]);320 321 GGML_ASSERT(ex >= 0 && ex < n_as); // sanity check322 323 const int64_t i11 = idx % src1->ne[1];324 const float * x = (const float *) (data + i11 * src1->nb[1] + row * src1->nb[2]);325 float * acc = e.values.data() + ex * ne0;326 327 e.counts[ex]++;328 touched[ex] = 1;329 for (int64_t j = 0; j < ne0; ++j) {330 acc[j] += x[j] * x[j];331 }332 }333 }334 335 // check for non-finite values, only checking experts that were routed to and touched336 for (int64_t ex = 0; ex < n_as; ++ex) {337 if (touched[ex] && !all_finite(e.values.data() + ex * ne0, ne0)) {338 LOG_ERR("%s: non-finite values detected in %s\n", __func__, wname.c_str());339 exit(1);340 }341 }342 343 for (int64_t ex = 0; ex < n_as; ++ex) {344 const int32_t n_chunk = e.counts[ex] / chunk_size;345 if (n_chunk > m_last_chunk) {346 const int32_t chunk_step = n_chunk - m_last_chunk;347 m_last_chunk = n_chunk;348 if ((m_last_chunk % m_params.n_out_freq) / chunk_step == 0) {349 save_imatrix();350 }351 if (m_params.n_save_freq > 0 && (m_last_chunk % m_params.n_save_freq) / chunk_step == 0) {352 save_imatrix(m_last_chunk);353 }354 }355 }356 } else {357 auto & e = m_stats[wname];358 const int64_t n_mat = src0->ne[2] * src0->ne[3];359 360 // use a single count per dense tensor361 // (necessary when merging older GGUF-imatrix files with 3d tensors)362 if (e.counts.size() > 1) {363 bool all_equal = true;364 for (size_t i = 1; i < e.counts.size(); ++i) {365 if (e.counts[0] != e.counts[i]) {366 all_equal = false;367 break;368 }369 }370 if (all_equal) {371 e.counts.resize(1);372 }373 }374 if (e.values.empty()) {375 e.values.resize(src1->ne[0] * n_mat, 0);376 e.counts.resize(1, 0);377 }378 else if (e.values.size() != (size_t)(src1->ne[0] * n_mat)) {379 LOG_ERR("%s: inconsistent size for %s (%d vs %d)\n", __func__, wname.c_str(), (int)e.values.size(), (int)(src1->ne[0] * n_mat));380 exit(1); //GGML_ABORT("fatal error");381 }382 LOG_DBGV(2, "%s[%d]: %32s, %s, %5d x %5d x %5d, %d\n", __func__, m_last_chunk, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->ne[2], (int)src1->type);383 384 const int64_t ne0 = src1->ne[0];385 386 for (int64_t i3 = 0; i3 < src1->ne[3]; ++i3) {387 for (int64_t i2 = 0; i2 < src1->ne[2]; ++i2) {388 // handle 3D+ tensors, but flatten 3D+ activations when model tensor is 2D389 const int64_t mat_id = (i3 % src0->ne[3]) * src0->ne[2] + (i2 % src0->ne[2]);390 float * acc = e.values.data() + mat_id * ne0;391 392 for (int64_t row = 0; row < src1->ne[1]; ++row) {393 const float * x = (const float *) (data + row * src1->nb[1] + i2 * src1->nb[2] + i3 * src1->nb[3]);394 for (int64_t j = 0; j < ne0; ++j) {395 acc[j] += x[j] * x[j];396 }397 }398 }399 }400 401 // check for non-finite values402 if (!all_finite(e.values.data(), e.values.size())) {403 LOG_ERR("%s: non-finite values detected in %s\n", __func__, wname.c_str());404 exit(1);405 }406 // only 1 count in practice, except when a tensor is used for both MUL_MAT_ID and MUL_MAT407 for (size_t i = 0; i < e.counts.size(); ++i) {408 e.counts[i] += ggml_nrows(src1) / n_mat;409 const int32_t n_chunk = e.counts[i] / chunk_size;410 if (n_chunk > m_last_chunk) {411 const int32_t chunk_step = n_chunk - m_last_chunk;412 m_last_chunk = n_chunk;413 if ((m_last_chunk % m_params.n_out_freq) / chunk_step == 0) {414 save_imatrix();415 }416 if (m_params.n_save_freq > 0 && (m_last_chunk % m_params.n_save_freq) / chunk_step == 0) {417 save_imatrix(m_last_chunk);418 }419 }420 }421 }422 423 return true;424}425 426void IMatrixCollector::save_imatrix_legacy(int32_t ncall) const {427 auto fname = m_params.out_file;428 429 if (ncall > 0) {430 fname += ".at_";431 fname += std::to_string(ncall);432 }433 434 // warn when writing imatrix entries that do not have full data435 // this can happen with MoE models where some of the experts end up not being exercised by the provided training data436 437 int n_entries = 0;438 std::vector<std::string> to_store;439 440 bool is_first = true; // for printing441 for (const auto & kv : m_stats) {442 const int n_all = kv.second.counts.size();443 444 if (n_all == 0) {445 continue;446 }447 448 int n_zeros = 0;449 for (const int c : kv.second.counts) {450 if (c == 0) {451 n_zeros++;452 }453 }454 455 if (n_zeros != 0 && is_first) {456 LOG_INF("\n");457 is_first = false;458 }459 460 if (n_zeros == n_all) {461 LOG_WRN("%s: entry '%40s' has no data - skipping\n", __func__, kv.first.c_str());462 continue;463 }464 465 if (n_zeros > 0) {466 LOG_WRN("%s: entry '%40s' has partial data (%.2f%%)\n", __func__, kv.first.c_str(), 100.0f * (n_all - n_zeros) / n_all);467 }468 469 n_entries++;470 to_store.push_back(kv.first);471 }472 473 if (to_store.size() < m_stats.size()) {474 LOG_WRN("%s: storing only %zu out of %zu entries\n", __func__, to_store.size(), m_stats.size());475 }476 477 // deterministic tensor name order478 std::sort(to_store.begin(), to_store.end());479 480 const int32_t chunk_size = m_params.n_ctx / m_params.n_parallel;481 482 std::ofstream out(fname, std::ios::binary);483 out.write((const char *) &n_entries, sizeof(n_entries));484 for (const auto & name : to_store) {485 const auto & stat = m_stats.at(name);486 const int32_t len = name.size();487 out.write((const char *) &len, sizeof(len));488 out.write(name.c_str(), len);489 // ceiling division to avoid accidental zeros490 const int32_t ncall = (*std::max_element(stat.counts.begin(), stat.counts.end()) + (chunk_size - 1)) / chunk_size;491 out.write((const char *) &ncall, sizeof(ncall));492 const int32_t nval = stat.values.size();493 const int32_t nmat = stat.counts.size();494 out.write((const char *) &nval, sizeof(nval));495 if (nval > 0 && nmat > 0) {496 std::vector<float> tmp(nval);497 for (int32_t i = 0; i < nval; i++) {498 float count = static_cast<float>(stat.counts[i / (nval / nmat)]);499 float value = stat.values[i];500 if (count == 0.0f) {501 // store 1 for partial data502 value = 1.0f;503 count = 1.0f;504 }505 tmp[i] = (value / count) * static_cast<float>(ncall);506 }507 out.write((const char *) tmp.data(), nval * sizeof(float));508 }509 }510 511 // Write the number of call the matrix was computed with512 out.write((const char *) &m_last_chunk, sizeof(m_last_chunk));513 514 // Write the input filename at the end of the file to later on specify it in quantize515 {516 const char * dataset_file = m_params.prompt_file.c_str();517 int32_t len = m_params.prompt_file.size();518 // When there is no prompt but there were other imatrix files loaded, use the last dataset519 if (m_params.prompt_file.empty() && !m_datasets.empty()) {520 const std::string & dataset_str = m_datasets[m_datasets.size() - 1];521 dataset_file = dataset_str.c_str();522 len = dataset_str.size();523 }524 out.write((const char *) &len, sizeof(len));525 out.write(dataset_file, len);526 }527 528 LOGV(1, "\n");529 LOG_DBGV(1, "%s: stored collected data after %d chunks in %s\n", __func__, m_last_chunk, fname.c_str());530}531 532void IMatrixCollector::save_imatrix(int32_t n_chunk) const {533 auto fname = m_params.out_file;534 int8_t use_legacy_format = m_params.imat_dat;535 536 if (use_legacy_format > 0) {537 this->save_imatrix_legacy(n_chunk);538 return;539 }540 // only warn when `--output-format gguf` is not specified541 if (use_legacy_format == 0 && !string_ends_with(fname, ".gguf")) {542 LOG_WRN("\n%s: saving imatrix using GGUF format with a different suffix than .gguf\n", __func__);543 LOG_WRN("%s: if you want the previous imatrix format, use --output-format dat\n", __func__);544 }545 546 if (n_chunk > 0) {547 fname += ".at_";548 fname += std::to_string(n_chunk);549 }550 551 // write imatrix entries even if they don't have full data. (can be corrected when reading)552 // this can happen with MoE models where some of the experts end up not being exercised by the provided training data553 554 std::vector<std::string> to_store;555 size_t data_size = 0;556 557 bool is_first = true; // for printing558 for (const auto & kv : m_stats) {559 const int n_all = kv.second.counts.size();560 561 int n_zeros = 0;562 for (const auto c : kv.second.counts) {563 if (c == 0) {564 n_zeros++;565 }566 }567 568 if (n_zeros != 0 && is_first) {569 LOG_INF("\n");570 is_first = false;571 }572 573 if (n_zeros > 0) {574 LOG_WRN("%s: entry '%40s' has partial data (%.2f%%)\n", __func__, kv.first.c_str(), 100.0f * (n_all - n_zeros) / n_all);575 }576 577 to_store.push_back(kv.first);578 data_size += GGML_PAD(ggml_tensor_overhead() + sizeof(float) * kv.second.values.size(), GGML_MEM_ALIGN);579 data_size += GGML_PAD(ggml_tensor_overhead() + sizeof(float) * kv.second.counts.size(), GGML_MEM_ALIGN);580 }581 582 // deterministic tensor name order583 std::sort(to_store.begin(), to_store.end());584 585 struct ggml_init_params params = {586 /* .mem_size = */ data_size,587 /* .mem_buffer = */ NULL,588 /* .no_alloc = */ false,589 };590 struct ggml_context * ctx = ggml_init(params);591 struct gguf_context * ctx_gguf = gguf_init_empty();592 593 {594 std::vector<const char *> datasets;595 datasets.reserve(m_datasets.size() + 1);596 for (size_t i = 0; i < m_datasets.size(); ++i) {597 datasets.push_back(m_datasets[i].c_str());598 }599 if (!m_params.prompt_file.empty()) {600 datasets.push_back(m_params.prompt_file.c_str());601 }602 603 gguf_set_val_str(ctx_gguf, "general.type", "imatrix");604 // Write the dataset paths605 gguf_set_arr_str(ctx_gguf, LLM_KV_IMATRIX_DATASETS, datasets.data(), datasets.size());606 // Write the number of chunks the matrix was computed with607 gguf_set_val_u32(ctx_gguf, LLM_KV_IMATRIX_CHUNK_COUNT, m_last_chunk);608 gguf_set_val_u32(ctx_gguf, LLM_KV_IMATRIX_CHUNK_SIZE, m_params.n_ctx / m_params.n_parallel);609 }610 611 for (const auto & name : to_store) {612 const auto & stat = m_stats.at(name);613 const int32_t nval = (int32_t) stat.values.size();614 const int32_t nmat = (int32_t) stat.counts.size();615 if (nval > 0 && nmat > 0) {616 struct ggml_tensor * in_sum2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, nval / nmat, nmat);617 struct ggml_tensor * counts = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 1, nmat);618 ggml_format_name(in_sum2, "%s.in_sum2", name.c_str());619 ggml_format_name(counts, "%s.counts", name.c_str());620 621 for (int32_t j = 0; j < nval; ++j) {622 ((float *) in_sum2->data)[j] = (float) stat.values[j];623 }624 for (int32_t j = 0; j < nmat; ++j) {625 ((float *) counts->data)[j] = (float) stat.counts[j];626 }627 628 gguf_add_tensor(ctx_gguf, in_sum2);629 gguf_add_tensor(ctx_gguf, counts);630 }631 }632 633 gguf_write_to_file(ctx_gguf, fname.c_str(), false);634 635 LOGV(1, "\n");636 LOG_DBGV(1, "%s: stored collected data after %d chunks in %s\n", __func__, m_last_chunk, fname.c_str());637 638 gguf_free(ctx_gguf);639 ggml_free(ctx);640}641 642bool IMatrixCollector::load_imatrix(const char * file_name) {643 common_imatrix loaded;644 if (!common_imatrix_load(file_name, loaded)) {645 return false;646 }647 648 const int32_t chunk_size = m_params.n_ctx / m_params.n_parallel;649 const bool is_legacy = loaded.is_legacy;650 651 for (auto & [name, entry] : loaded.entries) {652 auto & e = m_stats[name];653 654 if (is_legacy) {655 // Legacy format: sums contain (raw_sum/raw_count)*ncall, counts contain {ncall}656 // Reconstruct raw form by multiplying by chunk_size657 if (e.values.empty()) {658 e.values.resize(entry.sums.size(), 0.0f);659 e.counts.resize(1, 0);660 }661 for (size_t j = 0; j < entry.sums.size(); ++j) {662 e.values[j] += entry.sums[j] * chunk_size;663 }664 for (size_t j = 0; j < e.counts.size(); ++j) {665 e.counts[j] += entry.counts[0] * chunk_size;666 }667 } else {668 // GGUF format: raw sums and counts, accumulate directly669 const int64_t nval = entry.sums.size();670 const int64_t ncounts = entry.counts.size();671 672 if (e.values.empty()) {673 e.values.resize(nval, 0.0f);674 } else if ((size_t) nval != e.values.size()) {675 LOG_ERR("%s: mismatched sums size for %s: %zu != %zu\n", __func__, name.c_str(), (size_t) nval, e.values.size());676 return false;677 }678 679 if (e.counts.empty()) {680 e.counts.resize(ncounts, 0);681 } else if (e.counts.size() == 1 && ncounts > 1) {682 e.counts.resize(ncounts, e.counts[0]);683 } else if ((size_t) ncounts != e.counts.size()) {684 LOG_ERR("%s: mismatched counts size for %s: %zu != %zu\n", __func__, name.c_str(), (size_t) ncounts, e.counts.size());685 return false;686 }687 688 for (int64_t j = 0; j < nval; ++j) {689 e.values[j] += entry.sums[j];690 }691 for (int64_t j = 0; j < ncounts; ++j) {692 e.counts[j] += entry.counts[j];693 }694 }695 }696 697 m_datasets.insert(m_datasets.end(), loaded.datasets.begin(), loaded.datasets.end());698 699 // Calculate the last chunk count700 int64_t max_count = 0;701 for (const auto & stats : m_stats) {702 for (int64_t count : stats.second.counts) {703 if (count > max_count) {704 max_count = count;705 }706 }707 }708 m_last_chunk = max_count / chunk_size;709 710 return true;711}712 713static IMatrixCollector g_collector;714 715static bool ik_collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data) {716 return g_collector.collect_imatrix(t, ask, user_data);717}718 719struct results_log_softmax {720 double log_softmax;721 float logit;722 float prob;723};724 725static std::vector<float> softmax(const std::vector<float> & logits) {726 std::vector<float> probs(logits.size());727 float max_logit = logits[0];728 for (float v : logits) {729 max_logit = std::max(max_logit, v);730 }731 double sum_exp = 0.0;732 for (size_t i = 0; i < logits.size(); i++) {733 // Subtract the maximum logit value from the current logit value for numerical stability734 const float logit = logits[i] - max_logit;735 const float exp_logit = expf(logit);736 sum_exp += exp_logit;737 probs[i] = exp_logit;738 }739 for (size_t i = 0; i < probs.size(); i++) {740 probs[i] /= sum_exp;741 }742 return probs;743}744 745static results_log_softmax log_softmax(int n_vocab, const float * logits, int tok) {746 float max_logit = logits[0];747 for (int i = 1; i < n_vocab; ++i) {748 max_logit = std::max(max_logit, logits[i]);749 }750 double sum_exp = 0.0;751 for (int i = 0; i < n_vocab; ++i) {752 sum_exp += expf(logits[i] - max_logit);753 }754 return {logits[tok] - max_logit - log(sum_exp), logits[tok], expf(logits[tok] - max_logit) / (float) sum_exp};755}756 757static void process_logits(758 int n_vocab, const float * logits, const int * tokens, int n_token, std::vector<std::thread> & workers,759 double & nll, double & nll2, float * logit_history, float * prob_history) {760 std::mutex mutex;761 int counter = 0;762 auto compute = [&mutex, &counter, &nll, &nll2, logit_history, prob_history, n_vocab, logits, tokens, n_token] () {763 double local_nll = 0;764 double local_nll2 = 0;765 while (true) {766 std::unique_lock<std::mutex> lock(mutex);767 int i = counter++;768 if (i >= n_token) {769 nll += local_nll; nll2 += local_nll2;770 break;771 }772 lock.unlock();773 const results_log_softmax results = log_softmax(n_vocab, logits + i*n_vocab, tokens[i+1]);774 const double v = -results.log_softmax;775 local_nll += v;776 local_nll2 += v*v;777 778 logit_history[i] = results.logit;779 prob_history[i] = results.prob;780 }781 };782 for (auto & w : workers) {783 w = std::thread(compute);784 }785 compute();786 for (auto & w : workers) {787 w.join();788 }789}790 791static bool compute_imatrix(llama_context * ctx, const common_params & params, const int32_t n_ctx) {792 const llama_model * model = llama_get_model(ctx);793 const llama_vocab * vocab = llama_model_get_vocab(model);794 795 const bool add_bos = llama_vocab_get_add_bos(vocab);796 797 if (llama_pooling_type(ctx) != LLAMA_POOLING_TYPE_LAST) {798 GGML_ASSERT(!llama_vocab_get_add_eos(vocab));799 }800 801 auto tim1 = std::chrono::high_resolution_clock::now();802 LOG_INF("%s: tokenizing the input ..\n", __func__);803 804 std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, true, params.parse_special);805 806 auto tim2 = std::chrono::high_resolution_clock::now();807 LOG_INF("%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());808 809 if (params.i_chunk > 0) {810 if (size_t((params.i_chunk + 2)*n_ctx) >= tokens.size()) {811 LOG_ERR("%s: there will be not enough tokens left after removing %d chunks\n", __func__, params.i_chunk);812 return false;813 }814 LOG_INF("%s: removing initial %d chunks (%d tokens)\n", __func__, params.i_chunk, params.i_chunk*n_ctx);815 tokens.erase(tokens.begin(), tokens.begin() + params.i_chunk*n_ctx);816 }817 818 if (int(tokens.size()) < 2*n_ctx) {819 LOG_ERR("%s: you need at least %d tokens for a context of %d tokens\n", __func__, 2*n_ctx, n_ctx);820 LOG_ERR("%s: the data file you provided tokenizes to only %zu tokens\n", __func__, tokens.size());821 return false;822 }823 824 std::vector<float> logit_history;825 std::vector<float> prob_history;826 827 if (params.compute_ppl) {828 logit_history.resize(tokens.size());829 prob_history.resize(tokens.size());830 }831 832 const int n_chunk_max = tokens.size() / n_ctx;833 834 const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);835 const int n_vocab = llama_vocab_n_tokens(vocab);836 const int n_batch = params.n_batch;837 838 int count = 0;839 double nll = 0.0;840 double nll2 = 0.0;841 842 const int num_batches = (n_ctx + n_batch - 1) / n_batch;843 const int n_seq = std::max(1, n_batch / n_ctx);844 845 GGML_ASSERT(n_batch < n_ctx || n_batch % n_ctx == 0);846 GGML_ASSERT(params.n_ctx == n_seq * n_ctx);847 848 llama_batch batch = llama_batch_init(std::min(n_batch, n_ctx*n_seq), 0, 1);849 850 std::vector<float> logits;851 if (params.compute_ppl && num_batches > 1) {852 logits.reserve((size_t)n_ctx * n_vocab);853 }854 855 LOG_INF("%s: computing over %d chunks, n_ctx=%d, batch_size=%d, n_seq=%d\n", __func__, n_chunk, n_ctx, n_batch, n_seq);856 857 std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);858 859 for (int i = 0; i < n_chunk; i += n_seq) {860 const int start = i * n_ctx;861 const int end = start + n_ctx;862 863 const int n_seq_batch = std::min(n_seq, n_chunk - i);864 865 const auto t_start = std::chrono::high_resolution_clock::now();866 867 // clear the KV cache868 llama_memory_clear(llama_get_memory(ctx), true);869 870 for (int j = 0; j < num_batches; ++j) {871 const int batch_start = start + j * n_batch;872 const int batch_size = std::min(end - batch_start, n_batch);873 874 // clear the batch875 common_batch_clear(batch);876 877 for (int seq = 0; seq < n_seq_batch; seq++) {878 int seq_start = batch_start + seq*n_ctx;879 880 // save original token and restore it after eval881 const auto token_org = tokens[seq_start];882 883 // add BOS token for the first batch of each chunk884 if (add_bos && j == 0) {885 tokens[seq_start] = llama_vocab_bos(vocab);886 }887 for (int k = 0; k < batch_size; ++k) {888 // NOTE: specifying all logits to get activations for the output.weight tensor889 // and also for the perplexity calculation.890 // TODO: only get outputs when (params.process_output || params.compute_ppl)891 // (not possible when this skips FFN computation of the last layer)892 common_batch_add(batch, tokens[seq_start + k], j*n_batch + k, { seq }, true);893 }894 895 // restore the original token in case it was set to BOS896 tokens[seq_start] = token_org;897 }898 899 if (llama_decode(ctx, batch)) {900 LOG_ERR("%s : failed to eval\n", __func__);901 llama_batch_free(batch);902 return false;903 }904 905 if (params.compute_ppl && num_batches > 1) {906 const auto * batch_logits = llama_get_logits(ctx);907 logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);908 }909 }910 911 912 if (i == 0) {913 llama_synchronize(ctx);914 const auto t_end = std::chrono::high_resolution_clock::now();915 const float t_total = std::chrono::duration<float>(t_end - t_start).count();916 LOG_INF("%s: %.2f seconds per pass - ETA ", __func__, t_total);917 int total_seconds = (int)(t_total * n_chunk / n_seq);918 if (total_seconds >= 60*60) {919 LOG("%d hours ", total_seconds / (60*60));920 total_seconds = total_seconds % (60*60);921 }922 LOG("%.2f minutes\n", total_seconds / 60.0);923 }924 925 if (params.compute_ppl) {926 const int first = n_ctx/2;927 for (int seq = 0; seq < n_seq_batch; seq++) {928 const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits_ith(ctx, seq*n_ctx);929 930 llama_token * tokens_data = tokens.data() + start + seq*n_ctx + first;931 932 process_logits(n_vocab, all_logits + first*n_vocab,933 tokens_data, n_ctx - 1 - first,934 workers, nll, nll2,935 logit_history.data() + start + seq*n_ctx + first,936 prob_history.data() + start + seq*n_ctx + first);937 938 count += n_ctx - first - 1;939 940 LOG("[%d]%.4lf,", i + seq + 1, std::exp(nll / count));941 }942 fflush(stdout);943 944 logits.clear();945 }946 }947 948 LOG("\n");949 950 if (params.compute_ppl) {951 nll2 /= count;952 nll /= count;953 const double ppl = exp(nll);954 nll2 -= nll * nll;955 if (nll2 > 0) {956 nll2 = sqrt(nll2/(count-1));957 LOG("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl);958 } else {959 LOG("Unexpected negative standard deviation of log(prob)\n");960 }961 }962 963 llama_batch_free(batch);964 965 return true;966}967 968static bool show_statistics(const common_params & params) {969 std::vector<tensor_statistics> ts;970 if (params.in_files.empty() || params.in_files.size() > 1) {971 LOG_ERR("\nError: a single imatrix file is required to compute tensor statistics\n\n");972 return false;973 }974 if (g_collector.load_imatrix(params.in_files[0].c_str())) {975 for (const auto & [name, stats] :g_collector.get_mstats()) {976 compute_statistics(ts, name, stats);977 }978 } else {979 LOG_ERR("\nError: %s is not a valid imatrix file\n\n", params.in_files[0].c_str());980 return false;981 }982 if (!ts.empty()) {983 compute_cossim(ts);984 } else {985 LOG_ERR("Error: cannot compute statistics for %s\n\n", params.in_files[0].c_str());986 return false;987 }988 989 struct tensor_comparer {990 bool operator()(const tensor_statistics & a, const tensor_statistics & b) const {991 std::string layer, name_a, name_b;992 ;993 process_tensor_name(a.tensor, layer, name_a);994 process_tensor_name(b.tensor, layer, name_b);995 return name_a < name_b || (name_a == name_b && a.total_sqract > b.total_sqract);996 }997 };998 std::sort(ts.begin(), ts.end(), tensor_comparer());999 1000 struct weighted_stats {1001 float weighted_bias = 0.0f;1002 float weighted_zd = 0.0f;1003 float weighted_cossim = 0.0f;1004 int total_elements = 0;1005 };1006 std::map<int, weighted_stats> ws;1007 1008 LOG_INF("\nComputing statistics for %s (%d tensors)\n", params.in_files[0].c_str(), static_cast<int>(ts.size()));1009 LOG_INF("\n%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\n", " Layer", " Tensor", " Σ(Act²)",1010 " Min", " Max", " μ", " σ", " % Active", "N", " Entropy", "E (norm)", "ZD",1011 " CosSim");1012 LOG_INF(1013 "=============================================================================================================="1014 "===========================================================\n");1015 for (const auto & tstat : ts) {1016 std::string layer, name;1017 process_tensor_name(tstat.tensor, layer, name);1018 1019 int blk;1020 try {1021 blk = std::stoi(layer);1022 } catch (const std::exception & e) {1023 blk = -1; // not a block layer1024 }1025 1026 const float entropy_norm = (tstat.elements > 0) ? 100.0f * (tstat.entropy / std::log2(tstat.elements)) : 0.0f;1027 1028 LOG_INF("%5s\t%-20s\t%10.2f\t%8.4f\t%11.4f\t%6.2f\t%6.2f\t%8.2f%%\t%6d\t%10.4f\t%6.2f%%\t%10.2f%%\t%8.4f\n",1029 layer.c_str(), name.c_str(), tstat.total_sqract, tstat.min_sqract, tstat.max_sqract, tstat.mean_sqract,1030 tstat.stddev, tstat.active * 100.0f, tstat.elements, tstat.entropy,1031 entropy_norm, 100.0f * tstat.zd, tstat.cossim);1032 1033 const float weighted_bias = tstat.elements * tstat.total_sqract;1034 const float weighted_zd = tstat.elements * tstat.zd;1035 const float weighted_cossim = tstat.elements * tstat.cossim;1036 1037 if (ws.find(blk) != ws.end()) {1038 ws[blk].weighted_bias += weighted_bias;1039 ws[blk].weighted_zd += weighted_zd;1040 ws[blk].weighted_cossim += weighted_cossim;1041 ws[blk].total_elements += tstat.elements;1042 } else {1043 weighted_stats temp_ws;1044 temp_ws.weighted_bias = weighted_bias;1045 temp_ws.weighted_zd = weighted_zd;1046 temp_ws.weighted_cossim = weighted_cossim;1047 temp_ws.total_elements = tstat.elements;1048 ws[blk] = temp_ws;1049 }1050 }1051 1052 const int layers = std::count_if(ws.begin(), ws.end(), [](const auto & kv) { return kv.first >= 0; });1053 LOG_INF("\nComputing weighted average statistics per layer (%d layers)\n", layers);1054 LOG_INF("\n%s\t%s\t%s\t%s\n", " Layer", " μΣ(Act²)", " μZD", "μCosSim");1055 LOG_INF("================================================\n");1056 for (const auto & [first, second] : ws) {1057 const auto & layer = first;1058 const auto & stats = second;1059 1060 if (stats.total_elements == 0) {1061 continue;1062 }1063 1064 if (layer >= 0) {1065 const float bias = stats.weighted_bias / stats.total_elements;1066 const float zd = stats.weighted_zd / stats.total_elements;1067 const float cossim = stats.weighted_cossim / stats.total_elements;1068 1069 LOG_INF("%5d\t%14.2f\t%10.4f%%\t%6.4f\n", layer, bias, 100.0f * zd, cossim);1070 }1071 }1072 LOG_INF("\n");1073 1074 return true;1075}1076 1077int main(int argc, char ** argv) {1078 std::setlocale(LC_NUMERIC, "C");1079 1080 common_params params;1081 1082 params.out_file = "imatrix.gguf";1083 1084 params.n_ctx = 512;1085 params.escape = false;1086 1087 common_init();1088 1089 if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_IMATRIX, print_usage)) {1090 return 1;1091 }1092 1093 // set_params before show_statistics so load_imatrix has valid n_ctx/n_parallel1094 g_collector.set_params(params);1095 1096 if (params.show_statistics) {1097 if (!show_statistics(params)) {1098 return 1;1099 }1100 return 0;1101 }1102 1103 const int32_t n_ctx = params.n_ctx;1104 1105 if (n_ctx <= 0) {1106 LOG_ERR("%s: imatrix tool requires '--ctx-size' > 0\n", __func__);1107 return 1;1108 }1109 1110 {1111 const int32_t n_seq = std::max(1, params.n_batch / n_ctx);1112 const int32_t n_kv = n_seq * n_ctx;1113 1114 params.n_parallel = n_seq;1115 params.n_ctx = n_kv;1116 1117 params.n_batch = std::min(params.n_batch, n_kv);1118 }1119 1120 g_collector.set_params(params);1121 1122 for (const auto & in_file : params.in_files) {1123 LOG_INF("%s : loading imatrix from '%s'\n", __func__, in_file.c_str());1124 if (!g_collector.load_imatrix(in_file.c_str())) {1125 LOG_ERR("%s : failed to load %s\n", __func__, in_file.c_str());1126 return 1;1127 }1128 }1129 1130 if (params.prompt.empty()) {1131 LOG_INF("No prompt provided; combining precomputed matrices only.\n");1132 1133 if (params.in_files.empty()) {1134 LOG_ERR("Error: No prompt provided and no precomputed matrices (--in-file) to combine.\n");1135 return 1;1136 }1137 1138 if (params.in_files.size() == 1) {1139 LOG_INF("%s : saving imatrix to '%s'\n", __func__, params.out_file.c_str());1140 } else if (params.in_files.size() > 1) {1141 LOG_INF("%s : saving combined imatrix to '%s'\n", __func__, params.out_file.c_str());1142 }1143 1144 g_collector.save_imatrix();1145 1146 return 0;1147 }1148 1149 llama_backend_init();1150 llama_numa_init(params.numa);1151 1152 // pass the callback to the backend scheduler1153 // it will be executed for each node during the graph computation1154 params.cb_eval = ik_collect_imatrix;1155 params.cb_eval_user_data = NULL;1156 params.warmup = false;1157 1158 // init1159 auto llama_init = common_init_from_params(params);1160 1161 auto * model = llama_init->model();1162 auto * ctx = llama_init->context();1163 1164 if (model == nullptr || ctx == nullptr) {1165 LOG_ERR("%s : failed to init\n", __func__);1166 return 1;1167 }1168 1169 const int n_ctx_train = llama_model_n_ctx_train(model);1170 if (params.n_ctx > n_ctx_train) {1171 LOG_WRN("%s: model was trained on only %d context tokens (%d specified)\n",1172 __func__, n_ctx_train, params.n_ctx);1173 }1174 1175 // print system information1176 {1177 LOG_INF("\n");1178 LOG_INF("%s\n", common_params_get_system_info(params).c_str());1179 }1180 1181 if (!compute_imatrix(ctx, params, n_ctx)) {1182 return 1;1183 }1184 1185 g_collector.save_imatrix();1186 1187 LOG("\n");1188 llama_perf_context_print(ctx);1189 1190 llama_backend_free();1191 1192 return 0;1193}1194