1#include "ggml.h"
2#include "ggml-cpu.h"
3#include "llama.h"
4#include "common.h"
5
6#include "../src/llama-model.h"
7
8#include <algorithm>
9#include <cassert>
10#include <cinttypes>
11#include <cmath>
12#include <cstdio>
13#include <cstring>
14#include <numeric>
15#include <regex>
16#include <string>
17#include <vector>
18#include <thread>
19#include <mutex>
20
21#if defined(_MSC_VER)
22#pragma warning(disable: 4244 4267) // possible loss of data
23#endif
24
25struct quantize_stats_params {
26 std::string model = DEFAULT_MODEL_PATH;
27 bool verbose = false;
28 bool per_layer_stats = false;
29 bool print_histogram = false;
30 bool reference = false;
31 std::vector<std::string> include_layers;
32 std::vector<std::string> exclude_layers;
33 std::vector<enum ggml_type> include_types;
34};
35
36constexpr size_t HISTOGRAM_BUCKETS = 150;
37constexpr double HISTOGRAM_RANGE = 0.03;
38
39struct error_stats {
40 size_t num_samples;
41 double total_error;
42 double max_error;
43 uint64_t error_histogram[HISTOGRAM_BUCKETS];
44};
45
46static void quantize_stats_print_usage(int /*argc*/, char ** argv) {
47 quantize_stats_params params;
48 fprintf(stderr, format: "usage: %s [options]\n", argv[0]);
49 fprintf(stderr, format: "\n");
50 fprintf(stderr, format: "options:\n");
51 fprintf(stderr, format: " -h, --help show this help message and exit\n");
52 fprintf(stderr, format: " -m FNAME, --model FNAME\n");
53 fprintf(stderr, format: " model path (default: %s)\n", params.model.c_str());
54 fprintf(stderr, format: " -r, --reference\n");
55 fprintf(stderr, format: " use reference implementation (default: false)\n");
56 fprintf(stderr, format: " -v, --verbose\n");
57 fprintf(stderr, format: " verbose output (default: false)\n");
58 fprintf(stderr, format: " -p, --per-layer-stats\n");
59 fprintf(stderr, format: " print stats per layer (default: false)\n");
60 fprintf(stderr, format: " --histogram\n");
61 fprintf(stderr, format: " print error histogram (default: false)\n");
62 fprintf(stderr, format: " -l LAYER, --include-layer LAYER\n");
63 fprintf(stderr, format: " only test layers matching pattern\n");
64 fprintf(stderr, format: " -L LAYER, --exclude-layer LAYER\n");
65 fprintf(stderr, format: " exclude layers matching pattern\n");
66 fprintf(stderr, format: " -t TYPE, --type TYPE\n");
67 fprintf(stderr, format: " only test given type (q4_0, q4_1)\n");
68 fprintf(stderr, format: "\n");
69}
70
71// Check if a layer is included/excluded by command line
72static bool layer_included(const quantize_stats_params & params, const std::string & layer) {
73 for (const auto& excluded : params.exclude_layers) {
74 if (std::regex_search(s: layer, e: std::regex(excluded))) {
75 return false;
76 }
77 }
78 for (const auto& included : params.include_layers) {
79 if (std::regex_search(s: layer, e: std::regex(included))) {
80 return true;
81 }
82 }
83 return params.include_layers.empty();
84}
85
86// Update error statistics given vectors with the before/after result of quantization
87static void update_error_stats(int64_t nelements, const float * input, const float * output, error_stats & stats) {
88 for (int64_t i = 0; i < nelements; i++) {
89 double diff = input[i] - output[i];
90 stats.total_error += diff * diff;
91 stats.max_error = fmax(x: fabs(x: diff), y: stats.max_error);
92 stats.error_histogram[std::max(a: std::min(a: (size_t) floor(x: fabs(x: diff) / HISTOGRAM_RANGE * HISTOGRAM_BUCKETS), b: HISTOGRAM_BUCKETS-1), b: (size_t) 0)]++;
93 }
94 stats.num_samples += nelements;
95}
96
97static void combine_error_stats(error_stats & into, const error_stats & from) {
98 into.num_samples += from.num_samples;
99 into.total_error += from.total_error;
100 if (from.max_error > into.max_error) into.max_error = from.max_error;
101 for (size_t i=0; i<HISTOGRAM_BUCKETS; ++i) into.error_histogram[i] += from.error_histogram[i];
102}
103
104static double find_quantile(const error_stats & stats, double quantile) {
105 double sum = std::accumulate(first: std::begin(arr: stats.error_histogram), last: std::end(arr: stats.error_histogram), init: 0.0);
106
107 double accum = 0;
108 for (size_t i = 0; i < HISTOGRAM_BUCKETS; i++) {
109 accum += stats.error_histogram[i];
110 if (accum >= sum*quantile) {
111 return (i+1) * HISTOGRAM_RANGE / HISTOGRAM_BUCKETS;
112 }
113 }
114 return INFINITY;
115}
116
117static void print_error_stats(const std::string & name, const error_stats & stats, bool print_histogram) {
118 double rmse = sqrt(x: stats.total_error / (double) stats.num_samples);
119 double median = find_quantile(stats, quantile: .5);
120 double pct95 = find_quantile(stats, quantile: .95);
121 printf(format: "%-50s: rmse %.8f, maxerr %.8f, 95pct<%.4f, median<%.4f\n", name.c_str(), rmse, stats.max_error, pct95, median);
122 if (print_histogram) {
123 printf(format: "Error distribution:\n");
124 for (size_t i = 0; i < HISTOGRAM_BUCKETS; i++) {
125 double lower = i * HISTOGRAM_RANGE / HISTOGRAM_BUCKETS;
126 double upper = (i+1) * HISTOGRAM_RANGE / HISTOGRAM_BUCKETS;
127 if (i == HISTOGRAM_BUCKETS -1) upper = INFINITY;
128 printf(format: "[%3.4f, %3.4f): %11" PRIu64 "\n", lower, upper, stats.error_histogram[i]);
129 }
130 }
131}
132
133// copied from ggml.h - verify that we can access this as a flat array
134static bool tensor_is_contiguous(const struct ggml_tensor * tensor) {
135 static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
136
137 return
138 tensor->nb[0] == ggml_type_size(type: tensor->type) &&
139 tensor->nb[1] == (tensor->nb[0]*tensor->ne[0])/ggml_blck_size(type: tensor->type) &&
140 tensor->nb[2] == tensor->nb[1]*tensor->ne[1] &&
141 tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
142}
143
144static void test_roundtrip_on_chunk(
145 const ggml_tensor * layer, int64_t offset, int64_t chunk_size, const ggml_type_traits & qfns, const ggml_type_traits_cpu & qfns_cpu, bool use_reference,
146 float * input_scratch, char * quantized_scratch, float * output_scratch, error_stats & stats
147) {
148 if (layer->type == GGML_TYPE_F16) {
149 for (int i = 0; i < chunk_size; i++) {
150 input_scratch[i] = ggml_get_f32_1d(tensor: layer, i: i + offset);
151 }
152 } else {
153 input_scratch = ggml_get_data_f32(tensor: layer) + offset;
154 }
155
156 if (use_reference) {
157 qfns.from_float_ref(input_scratch, quantized_scratch, chunk_size);
158 } else {
159 qfns_cpu.from_float(input_scratch, quantized_scratch, chunk_size);
160 }
161 qfns.to_float(quantized_scratch, output_scratch, chunk_size);
162
163 update_error_stats(nelements: chunk_size, input: input_scratch, output: output_scratch, stats);
164}
165
166
167// Run quantization function for a single layer and update error stats
168static void test_roundtrip_on_layer(
169 std::string & name, bool print_layer_stats, const ggml_type_traits & qfns, const ggml_type_traits_cpu & qfns_cpu, bool use_reference,
170 const ggml_tensor * layer, std::vector<float> & input_scratch, std::vector<char> & quantized_scratch,
171 std::vector<float> & output_scratch, error_stats & total_error, int max_thread = 0
172) {
173 assert(tensor_is_contiguous(layer));
174 error_stats layer_error {};
175 uint64_t nelements = ggml_nelements(tensor: layer);
176
177 float* input_scratch_ptr = nullptr;
178 if (layer->type == GGML_TYPE_F16) {
179 if (input_scratch.size() < nelements) input_scratch.resize(new_size: nelements);
180 input_scratch_ptr = input_scratch.data();
181 }
182 if (quantized_scratch.size() < 4*nelements) quantized_scratch.resize(new_size: 4*nelements);
183 if (output_scratch.size() < nelements) output_scratch.resize(new_size: nelements);
184
185 if (max_thread < 1) max_thread = std::thread::hardware_concurrency();
186 int chunk_size = 32*512;
187 int num_chunks = (nelements + chunk_size - 1)/chunk_size;
188
189 if (num_chunks < 2 || max_thread < 2) {
190 test_roundtrip_on_chunk(layer, offset: 0, chunk_size: nelements, qfns, qfns_cpu, use_reference, input_scratch: input_scratch_ptr, quantized_scratch: quantized_scratch.data(),
191 output_scratch: output_scratch.data(), stats&: print_layer_stats ? layer_error : total_error);
192 } else {
193 auto & stats = print_layer_stats ? layer_error : total_error;
194 std::mutex mutex;
195 uint64_t counter = 0;
196 auto compute = [&mutex, &counter, &stats, &qfns, &qfns_cpu, nelements, layer, use_reference, input_scratch_ptr,
197 &quantized_scratch, &output_scratch, chunk_size] () {
198 error_stats local_stats {};
199 while (true) {
200 std::unique_lock<std::mutex> lock(mutex);
201 uint64_t offset = counter; counter += chunk_size;
202 if (offset >= nelements) {
203 combine_error_stats(into&: stats, from: local_stats);
204 break;
205 }
206 lock.unlock();
207 uint64_t chunk = offset + chunk_size < nelements ? chunk_size : nelements - offset;
208 test_roundtrip_on_chunk(layer, offset, chunk_size: chunk, qfns, qfns_cpu, use_reference, input_scratch: input_scratch_ptr + offset,
209 quantized_scratch: quantized_scratch.data() + 4*offset, output_scratch: output_scratch.data() + offset, stats&: local_stats);
210 }
211 };
212 int nthread = std::min(a: num_chunks, b: max_thread);
213 std::vector<std::thread> workers(nthread-1);
214 for (auto& w : workers) w = std::thread(compute);
215 compute();
216 for (auto& w : workers) w.join();
217 }
218
219 if (print_layer_stats) {
220 print_error_stats(name, stats: layer_error, print_histogram: false);
221 combine_error_stats(into&: total_error, from: layer_error);
222 }
223}
224
225int main(int argc, char ** argv) {
226 ggml_time_init();
227
228 quantize_stats_params params;
229
230 // read command line
231
232 int max_thread = 0;
233 bool invalid_param = false;
234 std::string arg;
235 for (int i = 1; i < argc; i++) {
236 arg = argv[i];
237
238 if (arg == "-h" || arg == "--help") {
239 quantize_stats_print_usage(argc, argv);
240 exit(status: 0);
241 } else if (arg == "-r" || arg == "--reference") {
242 params.reference = true;
243 } else if (arg == "-v") {
244 params.verbose = true;
245 } else if (arg == "-p" || arg == "--per-layer-stats") {
246 params.per_layer_stats = true;
247 } else if (arg == "--histogram") {
248 params.print_histogram = true;
249 } else if (arg == "-m" || arg == "--model") {
250 if (++i >= argc) {
251 invalid_param = true;
252 break;
253 }
254 params.model = argv[i];
255 } else if (arg == "-l" || arg == "--include-layer") {
256 if (++i >= argc) {
257 invalid_param = true;
258 break;
259 }
260 params.include_layers.emplace_back(args&: argv[i]);
261 } else if (arg == "-L" || arg == "--exclude-layer") {
262 if (++i >= argc) {
263 invalid_param = true;
264 break;
265 }
266 params.exclude_layers.emplace_back(args&: argv[i]);
267 } else if (arg == "-t" || arg == "--type") {
268 if (++i >= argc) {
269 invalid_param = true;
270 break;
271 }
272 int j;
273 for (j = 0; j < GGML_TYPE_COUNT; ++j) {
274 const auto * name = ggml_type_name(type: (ggml_type) j);
275 if (name && strcmp(s1: argv[i], s2: name) == 0) break;
276 }
277 if (j < GGML_TYPE_COUNT) {
278 params.include_types.push_back(x: (ggml_type) j);
279 } else {
280 fprintf(stderr, format: "error: %s not in list of types\n", argv[i]);
281 invalid_param = true;
282 }
283 } else if (arg == "-n" || arg == "--num-threads") {
284 if (++i >= argc) {
285 invalid_param = true;
286 break;
287 }
288 max_thread = atoi(nptr: argv[i]);
289 } else {
290 fprintf(stderr, format: "error: unknown argument: %s\n", arg.c_str());
291 quantize_stats_print_usage(argc, argv);
292 return 1;
293 }
294 }
295 if (invalid_param) {
296 fprintf(stderr, format: "error: invalid parameter for argument: %s\n", arg.c_str());
297 quantize_stats_print_usage(argc, argv);
298 return 1;
299 }
300
301 print_build_info();
302
303 // load the model
304 fprintf(stderr, format: "Loading model\n");
305
306 const int64_t t_main_start_us = ggml_time_us();
307 llama_model * model;
308 llama_context * ctx;
309
310 {
311 auto mparams = llama_model_default_params();
312 mparams.use_mlock = false;
313
314 model = llama_model_load_from_file(path_model: params.model.c_str(), params: mparams);
315
316 if (model == NULL) {
317 fprintf(stderr, format: "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
318 return 1;
319 }
320
321 auto cparams = llama_context_default_params();
322 cparams.n_ctx = 256;
323
324 ctx = llama_init_from_model(model, params: cparams);
325
326 if (ctx == NULL) {
327 fprintf(stderr, format: "%s: error: failed to create context with model '%s'\n", __func__, params.model.c_str());
328 llama_model_free(model);
329 return 1;
330 }
331 }
332
333 const auto & tensors = llama_internal_get_tensor_map(model);
334
335 // check layer tensors
336 int included_layers = 0;
337 int64_t max_nelements = 0;
338 bool is_f16 = false;
339 for (const auto & kv_tensor : tensors) {
340 if (!layer_included(params, layer: kv_tensor.first)) {
341 continue;
342 }
343 if (params.verbose) {
344 printf(format: "%s: type %s, size %" PRId64 "\n", kv_tensor.first.c_str(), ggml_type_name(type: kv_tensor.second->type), ggml_nelements(tensor: kv_tensor.second));
345 }
346 if (kv_tensor.second->type == GGML_TYPE_F16) {
347 is_f16 = true;
348 } else if (kv_tensor.second->type != GGML_TYPE_F32) {
349 fprintf(stderr, format: "%s: error: Quantization should be tested with a float model, "
350 "this model contains already quantized layers (%s is type %d)\n", __func__, kv_tensor.first.c_str(), kv_tensor.second->type);
351 llama_free(ctx);
352 llama_model_free(model);
353 return 1;
354 }
355 included_layers++;
356 max_nelements = std::max(a: max_nelements, b: ggml_nelements(tensor: kv_tensor.second));
357 }
358
359 if (is_f16) {
360 printf(format: "note: source model is f16\n");
361 }
362 printf(format: "testing %d layers with max size %" PRId64 "\n", included_layers, max_nelements);
363 // allocate scratch space
364 std::vector<float> input_scratch;
365 std::vector<char> quantized_scratch;
366 std::vector<float> output_scratch;
367
368 // loop throught quantization types
369 for (int i = 0; i < GGML_TYPE_COUNT; i++) {
370 const ggml_type type = (ggml_type) i;
371 if (!params.include_types.empty() && std::find(first: params.include_types.begin(), last: params.include_types.end(), val: i) == params.include_types.end()) {
372 continue;
373 }
374 const auto * qfns = ggml_get_type_traits(type);
375 const auto * qfns_cpu = ggml_get_type_traits_cpu(type);
376 if (qfns_cpu->from_float && qfns->to_float) {
377 if (params.verbose) {
378 printf(format: "testing %s ...\n", ggml_type_name(type));
379 }
380
381 ggml_quantize_init(type);
382
383 error_stats global_stats {};
384
385 for (const auto & kv_tensor : tensors) {
386 if (!layer_included(params, layer: kv_tensor.first)) {
387 continue;
388 }
389 if (params.verbose) {
390 printf(format: " %s ...\n", kv_tensor.first.c_str());
391 }
392 std::string layer_name { ggml_type_name(type) };
393 layer_name += "::" + kv_tensor.first;
394 test_roundtrip_on_layer(
395 name&: layer_name,
396 print_layer_stats: params.per_layer_stats,
397 qfns: *qfns, qfns_cpu: *qfns_cpu,
398 use_reference: params.reference,
399 layer: kv_tensor.second,
400 input_scratch,
401 quantized_scratch,
402 output_scratch,
403 total_error&: global_stats,
404 max_thread
405 );
406 }
407
408 print_error_stats(name: ggml_type_name(type), stats: global_stats, print_histogram: params.print_histogram);
409 }
410 }
411
412
413 llama_free(ctx);
414 llama_model_free(model);
415 // report timing
416 {
417 const int64_t t_main_end_us = ggml_time_us();
418
419 printf(format: "\n");
420 printf(format: "%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0);
421 }
422
423 return 0;
424}
425