KBaba7/llama.cpp

1#include "ggml.h"2#include "llama.h"3#include "llama-context.h"4#include "common.h"5 6#include <algorithm>7#include <cassert>8#include <cinttypes>9#include <cmath>10#include <cstdio>11#include <cstring>12#include <numeric>13#include <regex>14#include <string>15#include <vector>16#include <thread>17#include <mutex>18 19#if defined(_MSC_VER)20#pragma warning(disable: 4244 4267) // possible loss of data21#endif22 23struct quantize_stats_params {24    std::string model = DEFAULT_MODEL_PATH;25    bool verbose = false;26    bool per_layer_stats = false;27    bool print_histogram = false;28    bool reference = false;29    std::vector<std::string> include_layers;30    std::vector<std::string> exclude_layers;31    std::vector<enum ggml_type> include_types;32};33 34constexpr size_t HISTOGRAM_BUCKETS = 150;35constexpr double HISTOGRAM_RANGE = 0.03;36 37struct error_stats {38    size_t num_samples;39    double total_error;40    double max_error;41    uint64_t error_histogram[HISTOGRAM_BUCKETS];42};43 44static void quantize_stats_print_usage(int /*argc*/, char ** argv) {45    quantize_stats_params params;46    fprintf(stderr, "usage: %s [options]\n", argv[0]);47    fprintf(stderr, "\n");48    fprintf(stderr, "options:\n");49    fprintf(stderr, "  -h, --help            show this help message and exit\n");50    fprintf(stderr, "  -m FNAME, --model FNAME\n");51    fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());52    fprintf(stderr, "  -r, --reference\n");53    fprintf(stderr, "                        use reference implementation (default: false)\n");54    fprintf(stderr, "  -v, --verbose\n");55    fprintf(stderr, "                        verbose output (default: false)\n");56    fprintf(stderr, "  -p, --per-layer-stats\n");57    fprintf(stderr, "                        print stats per layer (default: false)\n");58    fprintf(stderr, "  --histogram\n");59    fprintf(stderr, "                        print error histogram (default: false)\n");60    fprintf(stderr, "  -l LAYER, --include-layer LAYER\n");61    fprintf(stderr, "                        only test layers matching pattern\n");62    fprintf(stderr, "  -L LAYER, --exclude-layer LAYER\n");63    fprintf(stderr, "                        exclude layers matching pattern\n");64    fprintf(stderr, "  -t TYPE, --type TYPE\n");65    fprintf(stderr, "                        only test given type (q4_0, q4_1)\n");66    fprintf(stderr, "\n");67}68 69// Check if a layer is included/excluded by command line70static bool layer_included(const quantize_stats_params & params, const std::string & layer) {71    for (const auto& excluded : params.exclude_layers) {72        if (std::regex_search(layer, std::regex(excluded))) {73            return false;74        }75    }76    for (const auto& included : params.include_layers) {77        if (std::regex_search(layer, std::regex(included))) {78            return true;79        }80    }81    return params.include_layers.empty();82}83 84// Update error statistics given vectors with the before/after result of quantization85static void update_error_stats(int64_t nelements, const float * input, const float * output, error_stats & stats) {86    for (int64_t i = 0; i < nelements; i++) {87        double diff = input[i] - output[i];88        stats.total_error += diff * diff;89        stats.max_error = fmax(fabs(diff), stats.max_error);90        stats.error_histogram[std::max(std::min((size_t) floor(fabs(diff) / HISTOGRAM_RANGE * HISTOGRAM_BUCKETS), HISTOGRAM_BUCKETS-1), (size_t) 0)]++;91    }92    stats.num_samples += nelements;93}94 95static void combine_error_stats(error_stats & into, const error_stats & from) {96    into.num_samples += from.num_samples;97    into.total_error += from.total_error;98    if (from.max_error > into.max_error) into.max_error = from.max_error;99    for (size_t i=0; i<HISTOGRAM_BUCKETS; ++i) into.error_histogram[i] += from.error_histogram[i];100}101 102static double find_quantile(const error_stats & stats, double quantile) {103    double sum = std::accumulate(std::begin(stats.error_histogram), std::end(stats.error_histogram), 0.0);104 105    double accum = 0;106    for (size_t i = 0; i < HISTOGRAM_BUCKETS; i++) {107        accum += stats.error_histogram[i];108        if (accum >= sum*quantile) {109            return (i+1) * HISTOGRAM_RANGE / HISTOGRAM_BUCKETS;110        }111    }112    return INFINITY;113}114 115static void print_error_stats(const std::string & name, const error_stats & stats, bool print_histogram) {116    double rmse = sqrt(stats.total_error / (double) stats.num_samples);117    double median = find_quantile(stats, .5);118    double pct95 = find_quantile(stats, .95);119    printf("%-50s: rmse %.8f, maxerr %.8f, 95pct<%.4f, median<%.4f\n", name.c_str(), rmse, stats.max_error, pct95, median);120    if (print_histogram) {121        printf("Error distribution:\n");122        for (size_t i = 0; i < HISTOGRAM_BUCKETS; i++) {123            double lower = i * HISTOGRAM_RANGE / HISTOGRAM_BUCKETS;124            double upper = (i+1) * HISTOGRAM_RANGE / HISTOGRAM_BUCKETS;125            if (i == HISTOGRAM_BUCKETS -1) upper = INFINITY;126            printf("[%3.4f, %3.4f): %11" PRIu64 "\n", lower, upper, stats.error_histogram[i]);127        }128    }129}130 131// copied from ggml.h - verify that we can access this as a flat array132static bool tensor_is_contiguous(const struct ggml_tensor * tensor) {133    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");134 135    return136        tensor->nb[0] == ggml_type_size(tensor->type) &&137        tensor->nb[1] == (tensor->nb[0]*tensor->ne[0])/ggml_blck_size(tensor->type) &&138        tensor->nb[2] == tensor->nb[1]*tensor->ne[1] &&139        tensor->nb[3] == tensor->nb[2]*tensor->ne[2];140}141 142static void test_roundtrip_on_chunk(143    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,144    float * input_scratch, char * quantized_scratch, float * output_scratch, error_stats & stats145) {146    if (layer->type == GGML_TYPE_F16) {147        for (int i = 0; i < chunk_size; i++) {148            input_scratch[i] = ggml_get_f32_1d(layer, i + offset);149        }150    } else {151        input_scratch = ggml_get_data_f32(layer) + offset;152    }153 154    if (use_reference) {155        qfns.from_float_ref(input_scratch, quantized_scratch, chunk_size);156    } else {157        qfns_cpu.from_float(input_scratch, quantized_scratch, chunk_size);158    }159    qfns.to_float(quantized_scratch, output_scratch, chunk_size);160 161    update_error_stats(chunk_size, input_scratch, output_scratch, stats);162}163 164 165// Run quantization function for a single layer and update error stats166static void test_roundtrip_on_layer(167    std::string & name, bool print_layer_stats, const ggml_type_traits & qfns, const ggml_type_traits_cpu & qfns_cpu, bool use_reference,168    const ggml_tensor * layer, std::vector<float> & input_scratch, std::vector<char> & quantized_scratch,169    std::vector<float> & output_scratch, error_stats & total_error, int max_thread = 0170) {171    assert(tensor_is_contiguous(layer));172    error_stats layer_error {};173    uint64_t nelements = ggml_nelements(layer);174 175    float* input_scratch_ptr = nullptr;176    if (layer->type == GGML_TYPE_F16) {177        if (input_scratch.size() < nelements) input_scratch.resize(nelements);178        input_scratch_ptr = input_scratch.data();179    }180    if (quantized_scratch.size() < 4*nelements) quantized_scratch.resize(4*nelements);181    if (output_scratch.size() < nelements) output_scratch.resize(nelements);182 183    if (max_thread < 1) max_thread = std::thread::hardware_concurrency();184    int chunk_size = 32*512;185    int num_chunks = (nelements + chunk_size - 1)/chunk_size;186 187    if (num_chunks < 2 || max_thread < 2) {188        test_roundtrip_on_chunk(layer, 0, nelements, qfns, qfns_cpu, use_reference, input_scratch_ptr, quantized_scratch.data(),189                output_scratch.data(), print_layer_stats ? layer_error : total_error);190    } else {191        auto & stats = print_layer_stats ? layer_error : total_error;192        std::mutex mutex;193        uint64_t counter = 0;194        auto compute = [&mutex, &counter, &stats, &qfns, &qfns_cpu, nelements, layer, use_reference, input_scratch_ptr,195             &quantized_scratch, &output_scratch, chunk_size] () {196            error_stats local_stats {};197            while (true) {198                std::unique_lock<std::mutex> lock(mutex);199                uint64_t offset = counter; counter += chunk_size;200                if (offset >= nelements) {201                    combine_error_stats(stats, local_stats);202                    break;203                }204                lock.unlock();205                uint64_t chunk = offset + chunk_size < nelements ? chunk_size : nelements - offset;206                test_roundtrip_on_chunk(layer, offset, chunk, qfns, qfns_cpu, use_reference, input_scratch_ptr + offset,207                        quantized_scratch.data() + 4*offset, output_scratch.data() + offset, local_stats);208            }209        };210        int nthread = std::min(num_chunks, max_thread);211        std::vector<std::thread> workers(nthread-1);212        for (auto& w : workers) w = std::thread(compute);213        compute();214        for (auto& w : workers) w.join();215    }216 217    if (print_layer_stats) {218        print_error_stats(name, layer_error, false);219        combine_error_stats(total_error, layer_error);220    }221}222 223int main(int argc, char ** argv) {224    ggml_time_init();225 226    quantize_stats_params params;227 228    // read command line229 230    int max_thread = 0;231    bool invalid_param = false;232    std::string arg;233    for (int i = 1; i < argc; i++) {234        arg = argv[i];235 236        if (arg == "-h" || arg == "--help") {237            quantize_stats_print_usage(argc, argv);238            exit(0);239        } else if (arg == "-r" || arg == "--reference") {240            params.reference = true;241        } else if (arg == "-v") {242            params.verbose = true;243        } else if (arg == "-p" || arg == "--per-layer-stats") {244            params.per_layer_stats = true;245        } else if (arg == "--histogram") {246            params.print_histogram = true;247        } else if (arg == "-m" || arg == "--model") {248            if (++i >= argc) {249                invalid_param = true;250                break;251            }252            params.model = argv[i];253        } else if (arg == "-l" || arg == "--include-layer") {254            if (++i >= argc) {255                invalid_param = true;256                break;257            }258            params.include_layers.emplace_back(argv[i]);259        } else if (arg == "-L" || arg == "--exclude-layer") {260            if (++i >= argc) {261                invalid_param = true;262                break;263            }264            params.exclude_layers.emplace_back(argv[i]);265        } else if (arg == "-t" || arg == "--type") {266            if (++i >= argc) {267                invalid_param = true;268                break;269            }270            int j;271            for (j = 0; j < GGML_TYPE_COUNT; ++j) {272               const auto * name = ggml_type_name((ggml_type) j);273               if (name && strcmp(argv[i], name) == 0) break;274            }275            if (j < GGML_TYPE_COUNT) {276                params.include_types.push_back((ggml_type) j);277            } else {278                fprintf(stderr, "error: %s not in list of types\n", argv[i]);279                invalid_param = true;280            }281        } else if (arg == "-n" || arg == "--num-threads") {282            if (++i >= argc) {283                invalid_param = true;284                break;285            }286            max_thread = atoi(argv[i]);287        } else {288            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());289            quantize_stats_print_usage(argc, argv);290            return 1;291        }292    }293    if (invalid_param) {294        fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());295        quantize_stats_print_usage(argc, argv);296        return 1;297    }298 299    print_build_info();300 301    // load the model302    fprintf(stderr, "Loading model\n");303 304    const int64_t t_main_start_us = ggml_time_us();305    llama_model * model;306    llama_context * ctx;307 308    {309        auto mparams = llama_model_default_params();310        mparams.use_mlock  = false;311 312        model = llama_model_load_from_file(params.model.c_str(), mparams);313 314        if (model == NULL) {315            fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());316            return 1;317        }318 319        auto cparams = llama_context_default_params();320        cparams.n_ctx = 256;321 322        ctx = llama_init_from_model(model, cparams);323 324        if (ctx == NULL) {325            fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, params.model.c_str());326            llama_model_free(model);327            return 1;328        }329    }330 331    const auto & tensors = llama_internal_get_tensor_map(ctx);332 333    // check layer tensors334    int included_layers = 0;335    int64_t max_nelements = 0;336    bool is_f16 = false;337    for (const auto & kv_tensor : tensors) {338        if (!layer_included(params, kv_tensor.first)) {339            continue;340        }341        if (params.verbose) {342            printf("%s: type %s, size %" PRId64 "\n", kv_tensor.first.c_str(), ggml_type_name(kv_tensor.second->type), ggml_nelements(kv_tensor.second));343        }344        if (kv_tensor.second->type == GGML_TYPE_F16) {345            is_f16 = true;346        } else if (kv_tensor.second->type != GGML_TYPE_F32) {347            fprintf(stderr, "%s: error: Quantization should be tested with a float model, "348                "this model contains already quantized layers (%s is type %d)\n", __func__, kv_tensor.first.c_str(), kv_tensor.second->type);349            llama_free(ctx);350            llama_model_free(model);351            return 1;352        }353        included_layers++;354        max_nelements = std::max(max_nelements, ggml_nelements(kv_tensor.second));355    }356 357    if (is_f16) {358        printf("note: source model is f16\n");359    }360    printf("testing %d layers with max size %" PRId64 "\n", included_layers, max_nelements);361    // allocate scratch space362    std::vector<float> input_scratch;363    std::vector<char> quantized_scratch;364    std::vector<float> output_scratch;365 366    // loop throught quantization types367    for (int i = 0; i < GGML_TYPE_COUNT; i++) {368        const ggml_type type = (ggml_type) i;369        if (!params.include_types.empty() && std::find(params.include_types.begin(), params.include_types.end(), i) == params.include_types.end()) {370            continue;371        }372        const auto * qfns     = ggml_get_type_traits(type);373        const auto * qfns_cpu = ggml_get_type_traits_cpu(type);374        if (qfns_cpu->from_float && qfns->to_float) {375            if (params.verbose) {376                printf("testing %s ...\n",  ggml_type_name(type));377            }378 379            ggml_quantize_init(type);380 381            error_stats global_stats {};382 383            for (const auto & kv_tensor : tensors) {384                if (!layer_included(params, kv_tensor.first)) {385                    continue;386                }387                if (params.verbose) {388                    printf("  %s ...\n",  kv_tensor.first.c_str());389                }390                std::string layer_name { ggml_type_name(type) };391                layer_name += "::" + kv_tensor.first;392                test_roundtrip_on_layer(393                        layer_name,394                        params.per_layer_stats,395                        *qfns, *qfns_cpu,396                        params.reference,397                        kv_tensor.second,398                        input_scratch,399                        quantized_scratch,400                        output_scratch,401                        global_stats,402                        max_thread403                );404            }405 406            print_error_stats(ggml_type_name(type), global_stats, params.print_histogram);407        }408    }409 410 411    llama_free(ctx);412    llama_model_free(model);413    // report timing414    {415        const int64_t t_main_end_us = ggml_time_us();416 417        printf("\n");418        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0);419    }420 421    return 0;422}423