1#include "arg.h"
2#include "common.h"
3#include "log.h"
4#include "llama.h"
5#include "gguf.h"
6
7#include <algorithm>
8#include <chrono>
9#include <cmath>
10#include <cstdio>
11#include <cstring>
12#include <ctime>
13#include <thread>
14#include <mutex>
15#include <vector>
16#include <fstream>
17#include <unordered_map>
18#include <map>
19#include <regex>
20#include <numeric>
21
22#if defined(_MSC_VER)
23#pragma warning(disable: 4244 4267) // possible loss of data
24#endif
25
26static void print_usage(int, char ** argv) {
27 LOG("\nexample usage:\n");
28 LOG("\n %s \\\n"
29 " -m model.gguf -f some-text.txt [-o imatrix.gguf] [--output-format {gguf,dat}] [--no-ppl] \\\n"
30 " [--process-output] [--chunk 123] [--save-frequency 0] [--output-frequency 10] \\\n"
31 " [--in-file imatrix-prev-0.gguf --in-file imatrix-prev-1.gguf ...] [--parse-special] \\\n"
32 " [--show-statistics] [...]\n" , argv[0]);
33 LOG("\n");
34}
35
36static const char * const LLM_KV_IMATRIX_DATASETS = "imatrix.datasets";
37static const char * const LLM_KV_IMATRIX_CHUNK_COUNT = "imatrix.chunk_count";
38static const char * const LLM_KV_IMATRIX_CHUNK_SIZE = "imatrix.chunk_size";
39
40struct Stats {
41 std::vector<float> values;
42 std::vector<int64_t> counts;
43};
44
45struct tensor_statistics {
46 std::string tensor;
47 Stats stats;
48 float total_sqract = 0.0f;
49 float mean_sqract = 0.0f;
50 float max_sqract = 0.0f;
51 float min_sqract = 0.0f;
52 int elements = 0;
53 float stddev = 0.0f;
54 float active = 0.0f;
55 float entropy = 0.0f;
56 float zd = 0.0f;
57 float cossim = 0.0f;
58};
59
60class IMatrixCollector {
61public:
62 IMatrixCollector() = default;
63 void set_params(common_params params) { m_params = std::move(params); }
64 bool collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data);
65 void save_imatrix_legacy(int32_t ncall = -1) const;
66 void save_imatrix(int32_t n_chunk = -1) const;
67 bool load_imatrix_legacy(const char * fname);
68 bool load_imatrix(const char * file_name);
69 const std::unordered_map<std::string, Stats> & get_mstats() const { return m_stats; }
70private:
71 std::unordered_map<std::string, Stats> m_stats;
72 common_params m_params;
73 std::mutex m_mutex;
74 std::vector<std::string> m_datasets;
75 int32_t m_last_chunk = 0;
76 std::vector<char> m_src1_data;
77 std::vector<char> m_ids; // the expert ids from ggml_mul_mat_id
78};
79
80// remove any prefix and suffixes from the name
81// CUDA0#blk.0.attn_k.weight#0 => blk.0.attn_k.weight
82static std::string filter_tensor_name(const char * name) {
83 std::string wname;
84 const char * p = strchr(s: name, c: '#');
85 if (p != NULL) {
86 p = p + 1;
87 const char * q = strchr(s: p, c: '#');
88 if (q != NULL) {
89 wname = std::string(p, q - p);
90 } else {
91 wname = p;
92 }
93 } else {
94 wname = name;
95 }
96 return wname;
97}
98
99static void process_tensor_name(const std::string & input, std::string & layer, std::string & tensor) {
100 std::vector<std::string> name;
101 std::istringstream stream(input);
102 std::string item;
103
104 while (std::getline(in&: stream, str&: item, delim: '.')) {
105 name.push_back(x: item);
106 }
107 for (size_t i = 0; i < name.size(); ++i) {
108 if (name[i] == "blk" && i + 1 < name.size()) {
109 layer = name[i + 1];
110 break;
111 }
112 }
113 for (size_t i = 0; i < name.size(); ++i) {
114 if (name[i] == "weight" && i > 0) {
115 tensor = name[i - 1];
116 break;
117 }
118 }
119
120 if (tensor.empty()) {
121 tensor = input;
122 }
123 if (layer.empty()) {
124 layer = "-";
125 }
126}
127
128static void compute_statistics(std::vector<tensor_statistics> & tstats, const std::string & name, const Stats & e) {
129 if (e.values.size() % e.counts.size() != 0) {
130 LOG_ERR("%s: activation size mismatch for tensor %s (%zu vs %zu)\n", __func__, name.c_str(), e.counts.size(), e.values.size());
131 return;
132 }
133 if (e.counts.empty()) {
134 LOG_ERR("%s: there are no activations for tensor %s. The imatrix may be suboptimal\n", __func__, name.c_str());
135 return;
136 }
137
138 const int n_mat = e.counts.size();
139 const int row_size = e.values.size() / n_mat;
140
141 std::vector<float> activations;
142 activations.reserve(n: e.values.size());
143
144 for (int i = 0; i < n_mat; ++i) {
145 for (int j = 0; j < row_size; ++j) {
146 activations.push_back(x: e.values[i*row_size + j] / e.counts[i]);
147 }
148 }
149
150 const float act_total = std::accumulate(first: activations.begin(), last: activations.end(), init: 0.0f);
151 const float act_max = *std::max_element(first: activations.begin(), last: activations.end());
152 const float act_min = *std::min_element(first: activations.begin(), last: activations.end());
153 const float act_mean = act_total / activations.size();
154 const float act_sqr_total = std::inner_product(first1: activations.begin(), last1: activations.end(), first2: activations.begin(), init: 0.0f);
155 const float act_var = (act_sqr_total / activations.size()) - (act_mean * act_mean);
156 const float act_dev = std::sqrt(x: std::max(a: 0.0f, b: act_var));
157 float threshold = 1e-5f;
158 const int inactive_count = std::count_if(first: activations.begin(), last: activations.end(),
159 pred: [threshold](const float v) { return fabsf(x: v) <= threshold; });
160 const float active_ratio = 1 - static_cast<float>(inactive_count) / activations.size();
161
162 float entropy = 0;
163 if (act_total > 0) {
164 for (const auto act : activations) {
165 if (const float p = act / act_total; p > 0) {
166 entropy -= p * std::log2(x: p);
167 }
168 }
169 }
170
171 int z_score = 0;
172 if (act_dev > 0.0f) {
173 for (const auto act : activations) {
174 if (const float p = (act - act_mean) / act_dev; p > 1) {
175 z_score++;
176 }
177 }
178 }
179
180 auto & ts = tstats.emplace_back();
181 ts.tensor = name;
182 ts.stats = e;
183 ts.total_sqract = act_total;
184 ts.mean_sqract = act_mean;
185 ts.max_sqract = act_max;
186 ts.min_sqract = act_min;
187 ts.elements = static_cast<int>(activations.size());
188 ts.stddev = act_dev;
189 ts.active = active_ratio;
190 ts.entropy = entropy;
191 ts.zd = static_cast<float>(z_score) / ts.elements;
192}
193
194static void compute_cossim(std::vector<tensor_statistics> & tstats) {
195 static const std::regex pattern(R"(blk\.(\d+)\.)");
196 for (auto & ts : tstats) {
197 if (std::smatch match; std::regex_search(s: ts.tensor, m&: match, e: pattern)) {
198 const int blk = std::stoi(str: match[1]);
199 std::string tname(ts.tensor);
200 tname.replace(pos: match.position(sub: 1), n: match.length(sub: 1), str: std::to_string(val: blk-1));
201 auto prev = std::find_if(first: tstats.begin(), last: tstats.end(),
202 pred: [tname](const tensor_statistics & t) { return t.tensor == tname; });
203 if (prev != tstats.end()) {
204 const float dp = std::inner_product(first1: ts.stats.values.begin(), last1: ts.stats.values.end(),
205 first2: prev->stats.values.begin(), init: 0.0f);
206 const float curr_mag = std::sqrt(x: std::inner_product(first1: ts.stats.values.begin(), last1: ts.stats.values.end(),
207 first2: ts.stats.values.begin(), init: 0.0f));
208 const float prev_mag = std::sqrt(x: std::inner_product(first1: prev->stats.values.begin(), last1: prev->stats.values.end(),
209 first2: prev->stats.values.begin(), init: 0.0f));
210 const float cs = dp / (curr_mag * prev_mag);
211 ts.cossim = cs;
212 }
213 } else {
214 ts.cossim = 0;
215 }
216 }
217}
218
219bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data) {
220 GGML_UNUSED(user_data);
221
222 const struct ggml_tensor * src0 = t->src[0];
223 const struct ggml_tensor * src1 = t->src[1];
224 std::string wname = filter_tensor_name(name: src0->name);
225
226 const int32_t chunk_size = m_params.n_ctx / m_params.n_parallel;
227
228 // when ask is true, the scheduler wants to know if we are interested in data from this tensor
229 // if we return true, a follow-up call will be made with ask=false in which we can do the actual collection
230 if (ask) {
231 if (t->op == GGML_OP_MUL_MAT_ID) return true; // collect all indirect matrix multiplications
232 if (t->op != GGML_OP_MUL_MAT) return false;
233 // why are small batches ignored (<16 tokens)?
234 if (src1->ne[1] < 16 || src1->type != GGML_TYPE_F32) return false;
235 if (!(wname.substr(pos: 0, n: 4) == "blk." || (m_params.process_output && wname == "output.weight"))) return false;
236 return true;
237 }
238
239 std::lock_guard<std::mutex> lock(m_mutex);
240
241 // copy the data from the GPU memory if needed
242 const bool is_host = ggml_backend_buffer_is_host(buffer: src1->buffer);
243
244 if (!is_host) {
245 const size_t src1_nbytes = ggml_nbytes(tensor: src1);
246 m_src1_data.resize(new_size: src1_nbytes);
247 ggml_backend_tensor_get(tensor: src1, data: m_src1_data.data(), offset: 0, size: src1_nbytes);
248 }
249
250 const char * data = is_host ? (const char *) src1->data : m_src1_data.data();
251 GGML_ASSERT(src1->nb[0] == ggml_element_size(src1));
252
253 // this has been adapted to the new format of storing merged experts in a single 3d tensor
254 // ref: https://github.com/ggml-org/llama.cpp/pull/6387
255 if (t->op == GGML_OP_MUL_MAT_ID) {
256 // ids -> [n_experts_used, n_tokens]
257 // src1 -> [cols, n_expert_used, n_tokens]
258 const ggml_tensor * ids = t->src[2];
259 const int64_t n_as = src0->ne[2];
260 const int64_t n_ids = ids->ne[0];
261
262 // the top-k selected expert ids are stored in the ids tensor
263 // for simplicity, always copy ids to host, because it is small
264 // take into account that ids is not contiguous!
265
266 GGML_ASSERT(ids->ne[1] == src1->ne[2]);
267
268 // the extra dimension would need to be stored somewhere to be reflected in the imatrix file
269 if (ggml_nrows(tensor: src1) != src1->ne[1] * src1->ne[2]) {
270 LOG_ERR("%s: tensor has more than 3 dimensions: %s", __func__, wname.c_str());
271 GGML_ASSERT(false);
272 }
273
274 m_ids.resize(new_size: ggml_nbytes(tensor: ids));
275 ggml_backend_tensor_get(tensor: ids, data: m_ids.data(), offset: 0, size: ggml_nbytes(tensor: ids));
276
277 auto & e = m_stats[wname];
278
279 if (e.counts.size() == 1 && n_as > 1) {
280 // broadcast, when loading an old imatrix
281 e.counts.resize(new_size: n_as, x: e.counts[0]);
282 }
283 if (e.values.empty()) {
284 e.values.resize(new_size: src1->ne[0]*n_as, x: 0);
285 e.counts.resize(new_size: n_as, x: 0);
286 }
287 else if (e.values.size() != (size_t)src1->ne[0]*n_as) {
288 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));
289 exit(status: 1); //GGML_ABORT("fatal error");
290 }
291 else if (e.counts.size() != (size_t)n_as) {
292 LOG_ERR("%s: inconsistent expert count for %s (%d vs %d)\n", __func__, wname.c_str(), (int)e.counts.size(), (int)n_as);
293 exit(status: 1); //GGML_ABORT("fatal error");
294 }
295 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);
296 // loop over all possible experts, regardless if they are used or not in the batch
297 for (int64_t ex = 0; ex < n_as; ++ex) {
298 size_t e_start = ex*src1->ne[0];
299
300 for (int64_t idx = 0; idx < n_ids; ++idx) {
301 for (int64_t row = 0; row < src1->ne[2]; ++row) {
302 const int excur = *(const int32_t *) (m_ids.data() + row*ids->nb[1] + idx*ids->nb[0]);
303
304 GGML_ASSERT(excur >= 0 && excur < n_as); // sanity check
305
306 if (excur != ex) continue;
307
308 const int64_t i11 = idx % src1->ne[1];
309 const int64_t i12 = row;
310 const float * x = (const float *)(data + i11*src1->nb[1] + i12*src1->nb[2]);
311
312 e.counts[ex]++;
313
314 for (int64_t j = 0; j < src1->ne[0]; ++j) {
315 e.values[e_start + j] += x[j] * x[j];
316 if (!std::isfinite(x: (float)e.values[e_start + j])) {
317 LOG_ERR("%f detected in %s\n", (float)e.values[e_start + j], wname.c_str());
318 exit(status: 1);
319 }
320 }
321 }
322 }
323 const int32_t n_chunk = e.counts[ex] / chunk_size;
324 if (n_chunk > m_last_chunk) {
325 const int32_t chunk_step = n_chunk - m_last_chunk;
326 m_last_chunk = n_chunk;
327 if ((m_last_chunk % m_params.n_out_freq) / chunk_step == 0) {
328 save_imatrix();
329 }
330 if (m_params.n_save_freq > 0 && (m_last_chunk % m_params.n_save_freq) / chunk_step == 0) {
331 save_imatrix(n_chunk: m_last_chunk);
332 }
333 }
334 }
335 } else {
336 auto & e = m_stats[wname];
337 const int64_t n_mat = src0->ne[2] * src0->ne[3];
338
339 // use a single count per dense tensor
340 // (necessary when merging older GGUF-imatrix files with 3d tensors)
341 if (e.counts.size() > 1) {
342 bool all_equal = true;
343 for (size_t i = 1; i < e.counts.size(); ++i) {
344 if (e.counts[0] != e.counts[i]) {
345 all_equal = false;
346 break;
347 }
348 }
349 if (all_equal) {
350 e.counts.resize(new_size: 1);
351 }
352 }
353 if (e.values.empty()) {
354 e.values.resize(new_size: src1->ne[0] * n_mat, x: 0);
355 e.counts.resize(new_size: 1, x: 0);
356 }
357 else if (e.values.size() != (size_t)(src1->ne[0] * n_mat)) {
358 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));
359 exit(status: 1); //GGML_ABORT("fatal error");
360 }
361 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);
362
363 for (int64_t i3 = 0; i3 < src1->ne[3]; ++i3) {
364 for (int64_t i2 = 0; i2 < src1->ne[2]; ++i2) {
365 // handle 3D+ tensors, but flatten 3D+ activations when model tensor is 2D
366 const int64_t mat_id = (i3 % src0->ne[3]) * src0->ne[2] + (i2 % src0->ne[2]);
367 const int64_t mat_start = mat_id * src1->ne[0];
368
369 for (int64_t row = 0; row < src1->ne[1]; ++row) {
370 const float * x = (const float *) (data + row * src1->nb[1] + i2 * src1->nb[2] + i3 * src1->nb[3]);
371 for (int64_t j = 0; j < src1->ne[0]; ++j) {
372 e.values[mat_start + j] += x[j] * x[j];
373 if (!std::isfinite(x: (float)e.values[j])) {
374 LOG_ERR("%f detected in %s\n", (float)e.values[j], wname.c_str());
375 exit(status: 1);
376 }
377 }
378 }
379 }
380 }
381 // only 1 count in practice, except when a tensor is used for both MUL_MAT_ID and MUL_MAT
382 for (size_t i = 0; i < e.counts.size(); ++i) {
383 e.counts[i] += ggml_nrows(tensor: src1) / n_mat;
384 const int32_t n_chunk = e.counts[i] / chunk_size;
385 if (n_chunk > m_last_chunk) {
386 const int32_t chunk_step = n_chunk - m_last_chunk;
387 m_last_chunk = n_chunk;
388 if ((m_last_chunk % m_params.n_out_freq) / chunk_step == 0) {
389 save_imatrix();
390 }
391 if (m_params.n_save_freq > 0 && (m_last_chunk % m_params.n_save_freq) / chunk_step == 0) {
392 save_imatrix(n_chunk: m_last_chunk);
393 }
394 }
395 }
396 }
397
398 return true;
399}
400
401void IMatrixCollector::save_imatrix_legacy(int32_t ncall) const {
402 auto fname = m_params.out_file;
403
404 if (ncall > 0) {
405 fname += ".at_";
406 fname += std::to_string(val: ncall);
407 }
408
409 // warn when writing imatrix entries that do not have full data
410 // this can happen with MoE models where some of the experts end up not being exercised by the provided training data
411
412 int n_entries = 0;
413 std::vector<std::string> to_store;
414
415 bool is_first = true; // for printing
416 for (const auto & kv : m_stats) {
417 const int n_all = kv.second.counts.size();
418
419 if (n_all == 0) {
420 continue;
421 }
422
423 int n_zeros = 0;
424 for (const int c : kv.second.counts) {
425 if (c == 0) {
426 n_zeros++;
427 }
428 }
429
430 if (n_zeros != 0 && is_first) {
431 LOG_INF("\n");
432 is_first = false;
433 }
434
435 if (n_zeros == n_all) {
436 LOG_WRN("%s: entry '%40s' has no data - skipping\n", __func__, kv.first.c_str());
437 continue;
438 }
439
440 if (n_zeros > 0) {
441 LOG_WRN("%s: entry '%40s' has partial data (%.2f%%)\n", __func__, kv.first.c_str(), 100.0f * (n_all - n_zeros) / n_all);
442 }
443
444 n_entries++;
445 to_store.push_back(x: kv.first);
446 }
447
448 if (to_store.size() < m_stats.size()) {
449 LOG_WRN("%s: storing only %zu out of %zu entries\n", __func__, to_store.size(), m_stats.size());
450 }
451
452 // deterministic tensor name order
453 std::sort(first: to_store.begin(), last: to_store.end());
454
455 const int32_t chunk_size = m_params.n_ctx / m_params.n_parallel;
456
457 std::ofstream out(fname, std::ios::binary);
458 out.write(s: (const char *) &n_entries, n: sizeof(n_entries));
459 for (const auto & name : to_store) {
460 const auto & stat = m_stats.at(k: name);
461 const int32_t len = name.size();
462 out.write(s: (const char *) &len, n: sizeof(len));
463 out.write(s: name.c_str(), n: len);
464 // ceiling division to avoid accidental zeros
465 const int32_t ncall = (*std::max_element(first: stat.counts.begin(), last: stat.counts.end()) + (chunk_size - 1)) / chunk_size;
466 out.write(s: (const char *) &ncall, n: sizeof(ncall));
467 const int32_t nval = stat.values.size();
468 const int32_t nmat = stat.counts.size();
469 out.write(s: (const char *) &nval, n: sizeof(nval));
470 if (nval > 0 && nmat > 0) {
471 std::vector<float> tmp(nval);
472 for (int32_t i = 0; i < nval; i++) {
473 float count = static_cast<float>(stat.counts[i / (nval / nmat)]);
474 float value = stat.values[i];
475 if (count == 0.0f) {
476 // store 1 for partial data
477 value = 1.0f;
478 count = 1.0f;
479 }
480 tmp[i] = (value / count) * static_cast<float>(ncall);
481 }
482 out.write(s: (const char *) tmp.data(), n: nval * sizeof(float));
483 }
484 }
485
486 // Write the number of call the matrix was computed with
487 out.write(s: (const char *) &m_last_chunk, n: sizeof(m_last_chunk));
488
489 // Write the input filename at the end of the file to later on specify it in quantize
490 {
491 const char * dataset_file = m_params.prompt_file.c_str();
492 int32_t len = m_params.prompt_file.size();
493 // When there is no prompt but there were other imatrix files loaded, use the last dataset
494 if (m_params.prompt_file.empty() && !m_datasets.empty()) {
495 const std::string & dataset_str = m_datasets[m_datasets.size() - 1];
496 dataset_file = dataset_str.c_str();
497 len = dataset_str.size();
498 }
499 out.write(s: (const char *) &len, n: sizeof(len));
500 out.write(s: dataset_file, n: len);
501 }
502
503 LOGV(1, "\n");
504 LOG_DBGV(1, "%s: stored collected data after %d chunks in %s\n", __func__, m_last_chunk, fname.c_str());
505}
506
507void IMatrixCollector::save_imatrix(int32_t n_chunk) const {
508 auto fname = m_params.out_file;
509 int8_t use_legacy_format = m_params.imat_dat;
510
511 if (use_legacy_format > 0) {
512 this->save_imatrix_legacy(ncall: n_chunk);
513 return;
514 }
515 // only warn when `--output-format gguf` is not specified
516 if (use_legacy_format == 0 && !string_ends_with(str: fname, suffix: ".gguf")) {
517 LOG_WRN("\n%s: saving imatrix using GGUF format with a different suffix than .gguf\n", __func__);
518 LOG_WRN("%s: if you want the previous imatrix format, use --output-format dat\n", __func__);
519 }
520
521 if (n_chunk > 0) {
522 fname += ".at_";
523 fname += std::to_string(val: n_chunk);
524 }
525
526 // write imatrix entries even if they don't have full data. (can be corrected when reading)
527 // this can happen with MoE models where some of the experts end up not being exercised by the provided training data
528
529 std::vector<std::string> to_store;
530 size_t data_size = 0;
531
532 bool is_first = true; // for printing
533 for (const auto & kv : m_stats) {
534 const int n_all = kv.second.counts.size();
535
536 int n_zeros = 0;
537 for (const auto c : kv.second.counts) {
538 if (c == 0) {
539 n_zeros++;
540 }
541 }
542
543 if (n_zeros != 0 && is_first) {
544 LOG_INF("\n");
545 is_first = false;
546 }
547
548 if (n_zeros > 0) {
549 LOG_WRN("%s: entry '%40s' has partial data (%.2f%%)\n", __func__, kv.first.c_str(), 100.0f * (n_all - n_zeros) / n_all);
550 }
551
552 to_store.push_back(x: kv.first);
553 data_size += GGML_PAD(ggml_tensor_overhead() + sizeof(float) * kv.second.values.size(), GGML_MEM_ALIGN);
554 data_size += GGML_PAD(ggml_tensor_overhead() + sizeof(float) * kv.second.counts.size(), GGML_MEM_ALIGN);
555 }
556
557 // deterministic tensor name order
558 std::sort(first: to_store.begin(), last: to_store.end());
559
560 struct ggml_init_params params = {
561 /* .mem_size = */ data_size,
562 /* .mem_buffer = */ NULL,
563 /* .no_alloc = */ false,
564 };
565 struct ggml_context * ctx = ggml_init(params);
566 struct gguf_context * ctx_gguf = gguf_init_empty();
567
568 {
569 std::vector<const char *> datasets;
570 datasets.reserve(n: m_datasets.size() + 1);
571 for (size_t i = 0; i < m_datasets.size(); ++i) {
572 datasets.push_back(x: m_datasets[i].c_str());
573 }
574 if (!m_params.prompt_file.empty()) {
575 datasets.push_back(x: m_params.prompt_file.c_str());
576 }
577
578 gguf_set_val_str(ctx: ctx_gguf, key: "general.type", val: "imatrix");
579 // Write the dataset paths
580 gguf_set_arr_str(ctx: ctx_gguf, key: LLM_KV_IMATRIX_DATASETS, data: datasets.data(), n: datasets.size());
581 // Write the number of chunks the matrix was computed with
582 gguf_set_val_u32(ctx: ctx_gguf, key: LLM_KV_IMATRIX_CHUNK_COUNT, val: m_last_chunk);
583 gguf_set_val_u32(ctx: ctx_gguf, key: LLM_KV_IMATRIX_CHUNK_SIZE, val: m_params.n_ctx / m_params.n_parallel);
584 }
585
586 for (const auto & name : to_store) {
587 const auto & stat = m_stats.at(k: name);
588 const int32_t nval = (int32_t) stat.values.size();
589 const int32_t nmat = (int32_t) stat.counts.size();
590 if (nval > 0 && nmat > 0) {
591 struct ggml_tensor * in_sum2 = ggml_new_tensor_2d(ctx, type: GGML_TYPE_F32, ne0: nval / nmat, ne1: nmat);
592 struct ggml_tensor * counts = ggml_new_tensor_2d(ctx, type: GGML_TYPE_F32, ne0: 1, ne1: nmat);
593 ggml_format_name(tensor: in_sum2, fmt: "%s.in_sum2", name.c_str());
594 ggml_format_name(tensor: counts, fmt: "%s.counts", name.c_str());
595
596 for (int32_t j = 0; j < nval; ++j) {
597 ((float *) in_sum2->data)[j] = (float) stat.values[j];
598 }
599 for (int32_t j = 0; j < nmat; ++j) {
600 ((float *) counts->data)[j] = (float) stat.counts[j];
601 }
602
603 gguf_add_tensor(ctx: ctx_gguf, tensor: in_sum2);
604 gguf_add_tensor(ctx: ctx_gguf, tensor: counts);
605 }
606 }
607
608 gguf_write_to_file(ctx: ctx_gguf, fname: fname.c_str(), only_meta: false);
609
610 LOGV(1, "\n");
611 LOG_DBGV(1, "%s: stored collected data after %d chunks in %s\n", __func__, m_last_chunk, fname.c_str());
612
613 gguf_free(ctx: ctx_gguf);
614 ggml_free(ctx);
615}
616
617bool IMatrixCollector::load_imatrix_legacy(const char * fname) {
618 std::ifstream in(fname, std::ios::binary);
619 if (!in) {
620 LOG_ERR("%s: failed to open %s\n", __func__, fname);
621 return false;
622 }
623 int n_entries;
624 in.read(s: (char *) &n_entries, n: sizeof(n_entries));
625 if (in.fail() || n_entries < 1) {
626 LOG_ERR("%s: no data in file %s\n", __func__, fname);
627 return false;
628 }
629 // Guess the chunk size because it's not stored in the file
630 const int32_t chunk_size = m_params.n_ctx / m_params.n_parallel;
631
632 for (int i = 0; i < n_entries; ++i) {
633 int32_t len = 0;
634 in.read(s: (char *) &len, n: sizeof(len));
635 std::vector<char> name_as_vec(len + 1);
636 in.read(s: (char *) name_as_vec.data(), n: len);
637 if (in.fail()) {
638 LOG_ERR("%s: failed reading name for entry %d from %s\n", __func__, i + 1, fname);
639 return false;
640 }
641 name_as_vec[len] = 0;
642 std::string name{ name_as_vec.data() };
643 auto & e = m_stats[std::move(name)];
644 int32_t ncall = 0;
645 in.read(s: (char *) &ncall, n: sizeof(ncall));
646 int32_t nval = 0;
647 in.read(s: (char *) &nval, n: sizeof(nval));
648 if (in.fail() || nval < 1) {
649 LOG_ERR("%s: failed reading number of values for entry %d\n", __func__, i);
650 m_stats = {};
651 return false;
652 }
653
654 if (e.values.empty()) {
655 e.values.resize(new_size: nval, x: 0.0f);
656 e.counts.resize(new_size: 1, x: 0);
657 }
658
659 std::vector<float> tmp(nval);
660 in.read(s: (char *) tmp.data(), n: nval * sizeof(float));
661 if (in.fail()) {
662 LOG_ERR("%s: failed reading data for entry %d\n", __func__, i);
663 m_stats = {};
664 return false;
665 }
666
667 // Recreate the state as expected by save_imatrix(), and correct for weighted sum.
668 for (int i = 0; i < nval; i++) {
669 e.values[i] += tmp[i] * chunk_size;
670 }
671 // The legacy format doesn't distinguish the counts for different experts
672 for (size_t j = 0; j < e.counts.size(); ++j) {
673 e.counts[j] += ncall * chunk_size;
674 }
675 }
676
677 {
678 // TODO: extract into its own method; this is also used by the GGUF-based format
679 // Calculate the last chunk count
680 int64_t max_count = 0;
681 for (const auto & stats : m_stats) {
682 for (int64_t count : stats.second.counts) {
683 if (count > max_count) {
684 max_count = count;
685 }
686 }
687 }
688 m_last_chunk = max_count / (chunk_size);
689 }
690
691 {
692 // Read the number of calls the matrix was computed with
693 int32_t n_calls;
694 in.read(s: (char *) &n_calls, n: sizeof(n_calls));
695 // ignore it because it's not important
696 }
697
698 // Read the dataset path to include it when writing to GGUF
699 if (!in.fail()){
700 int32_t len = 0;
701 in.read(s: (char *) &len, n: sizeof(len));
702 if (!in.fail()) {
703 std::vector<char> dataset;
704 dataset.resize(new_size: len + 1, x: 0);
705 in.read(s: dataset.data(), n: len);
706 if (!in.fail()) {
707 m_datasets.push_back(x: dataset.data());
708 }
709 }
710 }
711
712 return true;
713}
714
715// Using GGUF as the file format, for greater extensibility
716bool IMatrixCollector::load_imatrix(const char * file_name) {
717 struct ggml_context * ctx = nullptr;
718 struct gguf_init_params meta_gguf_params = {
719 /* .no_alloc = */ false, // the data is needed
720 /* .ctx = */ &ctx,
721 };
722 struct gguf_context * ctx_gguf = gguf_init_from_file(fname: file_name, params: meta_gguf_params);
723 if (!ctx_gguf) {
724 return this->load_imatrix_legacy(fname: file_name);
725 }
726 const int32_t n_entries = gguf_get_n_tensors(ctx: ctx_gguf);
727 if (n_entries < 1) {
728 LOG_ERR("%s: no data in file %s\n", __func__, file_name);
729 gguf_free(ctx: ctx_gguf);
730 ggml_free(ctx);
731 return false;
732 }
733
734 const int64_t datasets_key = gguf_find_key(ctx: ctx_gguf, key: LLM_KV_IMATRIX_DATASETS);
735 if (datasets_key != -1 && gguf_get_arr_type(ctx: ctx_gguf, key_id: datasets_key) == GGUF_TYPE_STRING) {
736 const int64_t n = gguf_get_arr_n(ctx: ctx_gguf, key_id: datasets_key);
737 m_datasets.reserve(n: m_datasets.size() + n);
738 for (int64_t i = 0; i < n; ++i) {
739 m_datasets.push_back(x: gguf_get_arr_str(ctx: ctx_gguf, key_id: datasets_key, i));
740 }
741 }
742
743 const std::string in_sum2_suffix{ ".in_sum2" };
744 const std::string counts_suffix{ ".counts" };
745
746 // Could re-use m_stats instead, but this allows
747 // checking for completeness of *each* loaded imatrix file
748 // and also makes it easier to re-use a similar implementation in quantize.cpp
749 // Using an ordered map to get a deterministic iteration order.
750 std::map<std::string, std::pair<struct ggml_tensor *, struct ggml_tensor *>> sums_counts_for;
751
752 for (struct ggml_tensor * cur = ggml_get_first_tensor(ctx); cur; cur = ggml_get_next_tensor(ctx, tensor: cur)) {
753 std::string name = cur->name;
754
755 if (name.empty()) { continue; }
756
757 if (string_remove_suffix(str&: name, suffix: in_sum2_suffix)) {
758 // in_sum2
759 sums_counts_for[std::move(name)].first = cur;
760 } else if (string_remove_suffix(str&: name, suffix: counts_suffix)) {
761 // counts
762 sums_counts_for[std::move(name)].second = cur;
763 } else {
764 // ignore other tensors
765 }
766 }
767
768 for (const auto & sc : sums_counts_for) {
769 const std::string & name = sc.first;
770 const struct ggml_tensor * in_sum2 = sc.second.first;
771 const struct ggml_tensor * counts = sc.second.second;
772
773 if (!in_sum2 || !counts) {
774 LOG_ERR("%s: mismatched sums and counts for %s\n", __func__, name.c_str());
775 gguf_free(ctx: ctx_gguf);
776 ggml_free(ctx);
777 return false;
778 }
779
780 auto & e = m_stats[name];
781
782 int64_t nval = ggml_nelements(tensor: in_sum2);
783 if (e.values.empty()) {
784 e.values.resize(new_size: nval, x: 0.0f);
785 } else if ((size_t) nval != e.values.size()) {
786 LOG_ERR("%s: mismatched sums size for %s: %zu != %zu\n", __func__, name.c_str(), (size_t) nval, e.values.size());
787 gguf_free(ctx: ctx_gguf);
788 ggml_free(ctx);
789 return false;
790 }
791
792 int64_t ncounts = ggml_nelements(tensor: counts);
793 if (e.counts.empty()) {
794 e.counts.resize(new_size: ncounts, x: 0);
795 } else if (e.counts.size() == 1 && ncounts > 1) {
796 // broadcast, when loading an old imatrix
797 e.counts.resize(new_size: ncounts, x: e.counts[0]);
798 } else if ((size_t) ncounts != e.counts.size()) {
799 LOG_ERR("%s: mismatched counts size for %s: %zu != %zu\n", __func__, name.c_str(), (size_t) ncounts, e.counts.size());
800 gguf_free(ctx: ctx_gguf);
801 ggml_free(ctx);
802 return false;
803 }
804
805 // Recreate the state as expected by save_imatrix()
806 for (int64_t j = 0; j < nval; j++) {
807 e.values[j] += ((const float *) in_sum2->data)[j];
808 }
809 for (int64_t j = 0; j < ncounts; j++) {
810 e.counts[j] += std::lround(x: ((const float *) counts->data)[j]);
811 }
812 }
813
814 // TODO: extract into its own method; this is also used by the legacy format
815 // Calculate the last chunk count
816 int64_t max_count = 0;
817 for (const auto & stats : m_stats) {
818 for (int64_t count : stats.second.counts) {
819 if (count > max_count) {
820 max_count = count;
821 }
822 }
823 }
824 m_last_chunk = max_count / (m_params.n_ctx / m_params.n_parallel);
825
826 gguf_free(ctx: ctx_gguf);
827 ggml_free(ctx);
828 return true;
829}
830
831static IMatrixCollector g_collector;
832
833static bool ik_collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data) {
834 return g_collector.collect_imatrix(t, ask, user_data);
835}
836
837struct results_log_softmax {
838 double log_softmax;
839 float logit;
840 float prob;
841};
842
843static std::vector<float> softmax(const std::vector<float> & logits) {
844 std::vector<float> probs(logits.size());
845 float max_logit = logits[0];
846 for (float v : logits) {
847 max_logit = std::max(a: max_logit, b: v);
848 }
849 double sum_exp = 0.0;
850 for (size_t i = 0; i < logits.size(); i++) {
851 // Subtract the maximum logit value from the current logit value for numerical stability
852 const float logit = logits[i] - max_logit;
853 const float exp_logit = expf(x: logit);
854 sum_exp += exp_logit;
855 probs[i] = exp_logit;
856 }
857 for (size_t i = 0; i < probs.size(); i++) {
858 probs[i] /= sum_exp;
859 }
860 return probs;
861}
862
863static results_log_softmax log_softmax(int n_vocab, const float * logits, int tok) {
864 float max_logit = logits[0];
865 for (int i = 1; i < n_vocab; ++i) {
866 max_logit = std::max(a: max_logit, b: logits[i]);
867 }
868 double sum_exp = 0.0;
869 for (int i = 0; i < n_vocab; ++i) {
870 sum_exp += expf(x: logits[i] - max_logit);
871 }
872 return {.log_softmax: logits[tok] - max_logit - log(x: sum_exp), .logit: logits[tok], .prob: expf(x: logits[tok] - max_logit) / (float) sum_exp};
873}
874
875static void process_logits(
876 int n_vocab, const float * logits, const int * tokens, int n_token, std::vector<std::thread> & workers,
877 double & nll, double & nll2, float * logit_history, float * prob_history) {
878 std::mutex mutex;
879 int counter = 0;
880 auto compute = [&mutex, &counter, &nll, &nll2, logit_history, prob_history, n_vocab, logits, tokens, n_token] () {
881 double local_nll = 0;
882 double local_nll2 = 0;
883 while (true) {
884 std::unique_lock<std::mutex> lock(mutex);
885 int i = counter++;
886 if (i >= n_token) {
887 nll += local_nll; nll2 += local_nll2;
888 break;
889 }
890 lock.unlock();
891 const results_log_softmax results = log_softmax(n_vocab, logits: logits + i*n_vocab, tok: tokens[i+1]);
892 const double v = -results.log_softmax;
893 local_nll += v;
894 local_nll2 += v*v;
895
896 logit_history[i] = results.logit;
897 prob_history[i] = results.prob;
898 }
899 };
900 for (auto & w : workers) {
901 w = std::thread(compute);
902 }
903 compute();
904 for (auto & w : workers) {
905 w.join();
906 }
907}
908
909static bool compute_imatrix(llama_context * ctx, const common_params & params, const int32_t n_ctx) {
910 const llama_model * model = llama_get_model(ctx);
911 const llama_vocab * vocab = llama_model_get_vocab(model);
912
913 const bool add_bos = llama_vocab_get_add_bos(vocab);
914
915 GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
916
917 auto tim1 = std::chrono::high_resolution_clock::now();
918 LOG_INF("%s: tokenizing the input ..\n", __func__);
919
920 std::vector<llama_token> tokens = common_tokenize(ctx, text: params.prompt, add_special: true, parse_special: params.parse_special);
921
922 auto tim2 = std::chrono::high_resolution_clock::now();
923 LOG_INF("%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
924
925 if (params.i_chunk > 0) {
926 if (size_t((params.i_chunk + 2)*n_ctx) >= tokens.size()) {
927 LOG_ERR("%s: there will be not enough tokens left after removing %d chunks\n", __func__, params.i_chunk);
928 return false;
929 }
930 LOG_INF("%s: removing initial %d chunks (%d tokens)\n", __func__, params.i_chunk, params.i_chunk*n_ctx);
931 tokens.erase(first: tokens.begin(), last: tokens.begin() + params.i_chunk*n_ctx);
932 }
933
934 if (int(tokens.size()) < 2*n_ctx) {
935 LOG_ERR("%s: you need at least %d tokens for a context of %d tokens\n", __func__, 2*n_ctx, n_ctx);
936 LOG_ERR("%s: the data file you provided tokenizes to only %zu tokens\n", __func__, tokens.size());
937 return false;
938 }
939
940 std::vector<float> logit_history;
941 std::vector<float> prob_history;
942
943 if (params.compute_ppl) {
944 logit_history.resize(new_size: tokens.size());
945 prob_history.resize(new_size: tokens.size());
946 }
947
948 const int n_chunk_max = tokens.size() / n_ctx;
949
950 const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(a: params.n_chunks, b: n_chunk_max);
951 const int n_vocab = llama_vocab_n_tokens(vocab);
952 const int n_batch = params.n_batch;
953
954 int count = 0;
955 double nll = 0.0;
956 double nll2 = 0.0;
957
958 const int num_batches = (n_ctx + n_batch - 1) / n_batch;
959 const int n_seq = std::max(a: 1, b: n_batch / n_ctx);
960
961 GGML_ASSERT(n_batch < n_ctx || n_batch % n_ctx == 0);
962 GGML_ASSERT(params.n_ctx == n_seq * n_ctx);
963
964 llama_batch batch = llama_batch_init(n_tokens: std::min(a: n_batch, b: n_ctx*n_seq), embd: 0, n_seq_max: 1);
965
966 std::vector<float> logits;
967 if (params.compute_ppl && num_batches > 1) {
968 logits.reserve(n: (size_t)n_ctx * n_vocab);
969 }
970
971 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);
972
973 std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
974
975 for (int i = 0; i < n_chunk; i += n_seq) {
976 const int start = i * n_ctx;
977 const int end = start + n_ctx;
978
979 const int n_seq_batch = std::min(a: n_seq, b: n_chunk - i);
980
981 const auto t_start = std::chrono::high_resolution_clock::now();
982
983 // clear the KV cache
984 llama_memory_clear(mem: llama_get_memory(ctx), data: true);
985
986 for (int j = 0; j < num_batches; ++j) {
987 const int batch_start = start + j * n_batch;
988 const int batch_size = std::min(a: end - batch_start, b: n_batch);
989
990 // clear the batch
991 common_batch_clear(batch);
992
993 for (int seq = 0; seq < n_seq_batch; seq++) {
994 int seq_start = batch_start + seq*n_ctx;
995
996 // save original token and restore it after eval
997 const auto token_org = tokens[seq_start];
998
999 // add BOS token for the first batch of each chunk
1000 if (add_bos && j == 0) {
1001 tokens[seq_start] = llama_vocab_bos(vocab);
1002 }
1003 for (int k = 0; k < batch_size; ++k) {
1004 // NOTE: specifying all logits to get activations for the output.weight tensor
1005 // and also for the perplexity calculation.
1006 // TODO: only get outputs when (params.process_output || params.compute_ppl)
1007 // (not possible when this skips FFN computation of the last layer)
1008 common_batch_add(batch, id: tokens[seq_start + k], pos: j*n_batch + k, seq_ids: { seq }, logits: true);
1009 }
1010
1011 // restore the original token in case it was set to BOS
1012 tokens[seq_start] = token_org;
1013 }
1014
1015 if (llama_decode(ctx, batch)) {
1016 LOG_ERR("%s : failed to eval\n", __func__);
1017 llama_batch_free(batch);
1018 return false;
1019 }
1020
1021 if (params.compute_ppl && num_batches > 1) {
1022 const auto * batch_logits = llama_get_logits(ctx);
1023 logits.insert(position: logits.end(), first: batch_logits, last: batch_logits + batch_size * n_vocab);
1024 }
1025 }
1026
1027
1028 if (i == 0) {
1029 llama_synchronize(ctx);
1030 const auto t_end = std::chrono::high_resolution_clock::now();
1031 const float t_total = std::chrono::duration<float>(t_end - t_start).count();
1032 LOG_INF("%s: %.2f seconds per pass - ETA ", __func__, t_total);
1033 int total_seconds = (int)(t_total * n_chunk / n_seq);
1034 if (total_seconds >= 60*60) {
1035 LOG("%d hours ", total_seconds / (60*60));
1036 total_seconds = total_seconds % (60*60);
1037 }
1038 LOG("%.2f minutes\n", total_seconds / 60.0);
1039 }
1040
1041 if (params.compute_ppl) {
1042 const int first = n_ctx/2;
1043 for (int seq = 0; seq < n_seq_batch; seq++) {
1044 const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits_ith(ctx, i: seq*n_ctx);
1045
1046 llama_token * tokens_data = tokens.data() + start + seq*n_ctx + first;
1047
1048 process_logits(n_vocab, logits: all_logits + first*n_vocab,
1049 tokens: tokens_data, n_token: n_ctx - 1 - first,
1050 workers, nll, nll2,
1051 logit_history: logit_history.data() + start + seq*n_ctx + first,
1052 prob_history: prob_history.data() + start + seq*n_ctx + first);
1053
1054 count += n_ctx - first - 1;
1055
1056 LOG("[%d]%.4lf,", i + seq + 1, std::exp(nll / count));
1057 }
1058 fflush(stdout);
1059
1060 logits.clear();
1061 }
1062 }
1063
1064 LOG("\n");
1065
1066 if (params.compute_ppl) {
1067 nll2 /= count;
1068 nll /= count;
1069 const double ppl = exp(x: nll);
1070 nll2 -= nll * nll;
1071 if (nll2 > 0) {
1072 nll2 = sqrt(x: nll2/(count-1));
1073 LOG("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl);
1074 } else {
1075 LOG("Unexpected negative standard deviation of log(prob)\n");
1076 }
1077 }
1078
1079 llama_batch_free(batch);
1080
1081 return true;
1082}
1083
1084static bool show_statistics(const common_params & params) {
1085 std::vector<tensor_statistics> ts;
1086 if (params.in_files.empty() || params.in_files.size() > 1) {
1087 LOG_ERR("\nError: a single imatrix file is required to compute tensor statistics\n\n");
1088 return false;
1089 }
1090 if (g_collector.load_imatrix(file_name: params.in_files[0].c_str())) {
1091 for (const auto & [name, stats] :g_collector.get_mstats()) {
1092 compute_statistics(tstats&: ts, name, e: stats);
1093 }
1094 } else {
1095 LOG_ERR("\nError: %s is not a valid imatrix file\n\n", params.in_files[0].c_str());
1096 return false;
1097 }
1098 if (!ts.empty()) {
1099 compute_cossim(tstats&: ts);
1100 } else {
1101 LOG_ERR("Error: cannot compute statistics for %s\n\n", params.in_files[0].c_str());
1102 return false;
1103 }
1104
1105 struct tensor_comparer {
1106 bool operator()(const tensor_statistics & a, const tensor_statistics & b) const {
1107 std::string layer, name_a, name_b;
1108 ;
1109 process_tensor_name(input: a.tensor, layer, tensor&: name_a);
1110 process_tensor_name(input: b.tensor, layer, tensor&: name_b);
1111 return name_a < name_b || (name_a == name_b && a.total_sqract > b.total_sqract);
1112 }
1113 };
1114 std::sort(first: ts.begin(), last: ts.end(), comp: tensor_comparer());
1115
1116 struct weighted_stats {
1117 float weighted_bias = 0.0f;
1118 float weighted_zd = 0.0f;
1119 float weighted_cossim = 0.0f;
1120 int total_elements = 0;
1121 };
1122 std::map<int, weighted_stats> ws;
1123
1124 LOG_INF("\nComputing statistics for %s (%d tensors)\n", params.in_files[0].c_str(), static_cast<int>(ts.size()));
1125 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)",
1126 " Min", " Max", " μ", " σ", " % Active", "N", " Entropy", "E (norm)", "ZD",
1127 " CosSim");
1128 LOG_INF(
1129 "=============================================================================================================="
1130 "===========================================================\n");
1131 for (const auto & tstat : ts) {
1132 std::string layer, name;
1133 process_tensor_name(input: tstat.tensor, layer, tensor&: name);
1134
1135 int blk;
1136 try {
1137 blk = std::stoi(str: layer);
1138 } catch (const std::exception & e) {
1139 blk = -1; // not a block layer
1140 }
1141
1142 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",
1143 layer.c_str(), name.c_str(), tstat.total_sqract, tstat.min_sqract, tstat.max_sqract, tstat.mean_sqract,
1144 tstat.stddev, tstat.active * 100.0f, tstat.elements, tstat.entropy,
1145 100.0f * (tstat.entropy / std::log2(tstat.elements)), 100.0f * tstat.zd, tstat.cossim);
1146
1147 const float weighted_bias = tstat.elements * tstat.total_sqract;
1148 const float weighted_zd = tstat.elements * tstat.zd;
1149 const float weighted_cossim = tstat.elements * tstat.cossim;
1150
1151 if (ws.find(x: blk) != ws.end()) {
1152 ws[blk].weighted_bias += weighted_bias;
1153 ws[blk].weighted_zd += weighted_zd;
1154 ws[blk].weighted_cossim += weighted_cossim;
1155 ws[blk].total_elements += tstat.elements;
1156 } else {
1157 weighted_stats temp_ws;
1158 temp_ws.weighted_bias = weighted_bias;
1159 temp_ws.weighted_zd = weighted_zd;
1160 temp_ws.weighted_cossim = weighted_cossim;
1161 temp_ws.total_elements = tstat.elements;
1162 ws[blk] = temp_ws;
1163 }
1164 }
1165
1166 const int layers = std::count_if(first: ws.begin(), last: ws.end(), pred: [](const auto & kv) { return kv.first >= 0; });
1167 LOG_INF("\nComputing weighted average statistics per layer (%d layers)\n", layers);
1168 LOG_INF("\n%s\t%s\t%s\t%s\n", " Layer", " μΣ(Act²)", " μZD", "μCosSim");
1169 LOG_INF("================================================\n");
1170 for (const auto & [first, second] : ws) {
1171 const auto & layer = first;
1172 const auto & stats = second;
1173
1174 if (stats.total_elements == 0) {
1175 continue;
1176 }
1177
1178 if (layer >= 0) {
1179 const float bias = stats.weighted_bias / stats.total_elements;
1180 const float zd = stats.weighted_zd / stats.total_elements;
1181 const float cossim = stats.weighted_cossim / stats.total_elements;
1182
1183 LOG_INF("%5d\t%14.2f\t%10.4f%%\t%6.4f\n", layer, bias, 100.0f * zd, cossim);
1184 }
1185 }
1186 LOG_INF("\n");
1187
1188 return true;
1189}
1190
1191int main(int argc, char ** argv) {
1192 common_params params;
1193
1194 params.out_file = "imatrix.gguf";
1195
1196 params.n_ctx = 512;
1197 params.escape = false;
1198
1199 if (!common_params_parse(argc, argv, params, ex: LLAMA_EXAMPLE_IMATRIX, print_usage)) {
1200 return 1;
1201 }
1202
1203 if (params.show_statistics) {
1204 if (!show_statistics(params)) {
1205 return 1;
1206 }
1207 return 0;
1208 }
1209
1210 common_init();
1211
1212 const int32_t n_ctx = params.n_ctx;
1213
1214 if (n_ctx <= 0) {
1215 LOG_ERR("%s: imatrix tool requires '--ctx-size' > 0\n", __func__);
1216 return 1;
1217 }
1218
1219 {
1220 const int32_t n_seq = std::max(a: 1, b: params.n_batch / n_ctx);
1221 const int32_t n_kv = n_seq * n_ctx;
1222
1223 params.n_parallel = n_seq;
1224 params.n_ctx = n_kv;
1225
1226 params.n_batch = std::min(a: params.n_batch, b: n_kv);
1227 }
1228
1229 g_collector.set_params(params);
1230
1231 for (const auto & in_file : params.in_files) {
1232 LOG_INF("%s : loading imatrix from '%s'\n", __func__, in_file.c_str());
1233 if (!g_collector.load_imatrix(file_name: in_file.c_str())) {
1234 LOG_ERR("%s : failed to load %s\n", __func__, in_file.c_str());
1235 return 1;
1236 }
1237 }
1238
1239 if (params.prompt.empty()) {
1240 LOG_INF("No prompt provided; combining precomputed matrices only.\n");
1241
1242 if (params.in_files.empty()) {
1243 LOG_ERR("Error: No prompt provided and no precomputed matrices (--in-file) to combine.\n");
1244 return 1;
1245 }
1246
1247 if (params.in_files.size() == 1) {
1248 LOG_INF("%s : saving imatrix to '%s'\n", __func__, params.out_file.c_str());
1249 } else if (params.in_files.size() > 1) {
1250 LOG_INF("%s : saving combined imatrix to '%s'\n", __func__, params.out_file.c_str());
1251 }
1252
1253 g_collector.save_imatrix();
1254
1255 return 0;
1256 }
1257
1258 llama_backend_init();
1259 llama_numa_init(numa: params.numa);
1260
1261 // pass the callback to the backend scheduler
1262 // it will be executed for each node during the graph computation
1263 params.cb_eval = ik_collect_imatrix;
1264 params.cb_eval_user_data = NULL;
1265 params.warmup = false;
1266
1267 // init
1268 common_init_result llama_init = common_init_from_params(params);
1269
1270 llama_model * model = llama_init.model.get();
1271 llama_context * ctx = llama_init.context.get();
1272
1273 if (model == nullptr || ctx == nullptr) {
1274 LOG_ERR("%s : failed to init\n", __func__);
1275 return 1;
1276 }
1277
1278 const int n_ctx_train = llama_model_n_ctx_train(model);
1279 if (params.n_ctx > n_ctx_train) {
1280 LOG_WRN("%s: model was trained on only %d context tokens (%d specified)\n",
1281 __func__, n_ctx_train, params.n_ctx);
1282 }
1283
1284 // print system information
1285 {
1286 LOG_INF("\n");
1287 LOG_INF("%s\n", common_params_get_system_info(params).c_str());
1288 }
1289
1290 if (!compute_imatrix(ctx, params, n_ctx)) {
1291 return 1;
1292 }
1293
1294 g_collector.save_imatrix();
1295
1296 LOG("\n");
1297 llama_perf_context_print(ctx);
1298
1299 llama_backend_free();
1300
1301 return 0;
1302}
1303