aboutsummaryrefslogtreecommitdiff
path: root/quantize.cpp
blob: 0ae537339ecf3a48ee408057d98c5470ca9a7db3 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
#include "ggml.h"

#include "utils.h"

#include <cassert>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <fstream>
#include <map>
#include <string>
#include <vector>
#include <regex>

// TODO: move somewhere else
#define QK 32

// default hparams (LLaMA76B)
struct llama_hparams {
    int32_t n_vocab = 32000;
    int32_t n_ctx   = 512;   // this is provided as user input?
    int32_t n_embd  = 4096;
    int32_t n_mult  = 256;
    int32_t n_head  = 32;
    int32_t n_layer = 32;
    int32_t n_rot   = 64;
    int32_t f16     = 1;
};


// quantize a model
bool llama_model_quantize(const std::string & fname_inp, const std::string & fname_out, int itype) {
    ggml_type type = GGML_TYPE_Q4_1;

    switch (itype) {
        case 2: type = GGML_TYPE_Q4_0; break;
        case 3: type = GGML_TYPE_Q4_1; break;
        default: fprintf(stderr, "%s: invalid quantization type %d\n", __func__, itype); return 1;
    };

    if (type != GGML_TYPE_Q4_0 && type != GGML_TYPE_Q4_1) {
        fprintf(stderr, "%s: invalid quantization type %d\n", __func__, type);
        return false;
    }

    gpt_vocab vocab;

    printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str());

    auto finp = std::ifstream(fname_inp, std::ios::binary);
    if (!finp) {
        fprintf(stderr, "%s: failed to open '%s' for reading\n", __func__, fname_inp.c_str());
        return false;
    }

    auto fout = std::ofstream(fname_out, std::ios::binary);
    if (!fout) {
        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname_out.c_str());
        return false;
    }

    // verify magic
    {
        uint32_t magic;
        finp.read((char *) &magic, sizeof(magic));
        if (magic != 0x67676d6c) {
            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname_inp.c_str());
            return false;
        }

        fout.write((char *) &magic, sizeof(magic));
    }

    llama_hparams hparams;

    // load hparams
    {
        finp.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
        //finp.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
        finp.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
        finp.read((char *) &hparams.n_mult,  sizeof(hparams.n_mult));
        finp.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
        finp.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
        finp.read((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
        finp.read((char *) &hparams.f16,     sizeof(hparams.f16));

        printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
        printf("%s: n_ctx   = %d\n", __func__, hparams.n_ctx);
        printf("%s: n_embd  = %d\n", __func__, hparams.n_embd);
        printf("%s: n_mult  = %d\n", __func__, hparams.n_mult);
        printf("%s: n_head  = %d\n", __func__, hparams.n_head);
        printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
        printf("%s: f16     = %d\n", __func__, hparams.f16);

        fout.write((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
        //fout.write((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
        fout.write((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
        fout.write((char *) &hparams.n_mult,  sizeof(hparams.n_mult));
        fout.write((char *) &hparams.n_head,  sizeof(hparams.n_head));
        fout.write((char *) &hparams.n_layer, sizeof(hparams.n_layer));
        fout.write((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
        fout.write((char *) &itype,           sizeof(hparams.f16));
    }

    // load vocab
    {
        const int32_t n_vocab = hparams.n_vocab;

        if (n_vocab != hparams.n_vocab) {
            fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
                    __func__, fname_inp.c_str(), n_vocab, hparams.n_vocab);
            return false;
        }

        std::string word;
        for (int i = 0; i < n_vocab; i++) {
            uint32_t len;
            finp.read ((char *) &len, sizeof(len));
            fout.write((char *) &len, sizeof(len));

            word.resize(len);
            finp.read ((char *) word.data(), len);
            fout.write((char *) word.data(), len);

            vocab.token_to_id[word] = i;
            vocab.id_to_token[i] = word;
        }
    }

    // load weights
    {
        size_t total_size_org = 0;
        size_t total_size_new = 0;

        std::vector<float> work;

        std::vector<uint8_t>     data_u8;
        std::vector<ggml_fp16_t> data_f16;
        std::vector<float>       data_f32;

        std::vector<int64_t> hist_all(1 << 4, 0);

        while (true) {
            int32_t n_dims;
            int32_t length;
            int32_t ftype;

            finp.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
            finp.read(reinterpret_cast<char *>(&length), sizeof(length));
            finp.read(reinterpret_cast<char *>(&ftype),  sizeof(ftype));

            if (finp.eof()) {
                break;
            }

            int32_t nelements = 1;
            int32_t ne[2] = { 1, 1 };
            for (int i = 0; i < n_dims; ++i) {
                finp.read (reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
                nelements *= ne[i];
            }

            std::string name(length, 0);
            finp.read (&name[0], length);

            {
                static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", };
                printf("%48s - [%5d, %5d], type = %6s ", name.data(), ne[0], ne[1], ftype_str[ftype]);
            }

            // regexes of tensor names to be quantized
            const std::vector<std::string> k_names = {
                ".*weight",
            };

            bool quantize = false;
            for (const auto & s : k_names) {
                if (std::regex_match(name, std::regex(s))) {
                    quantize = true;
                    break;
                }
            }

            // quantize only 2D tensors
            quantize &= (n_dims == 2);

            if (quantize) {
                if (ftype != 0 && ftype != 1) {
                    fprintf(stderr, "%s: unsupported ftype %d for integer quantization\n", __func__, ftype);
                    return false;
                }

                if (ftype == 1) {
                    data_f16.resize(nelements);
                    finp.read(reinterpret_cast<char *>(data_f16.data()), nelements * sizeof(ggml_fp16_t));
                    data_f32.resize(nelements);
                    for (int i = 0; i < nelements; ++i) {
                        data_f32[i] = ggml_fp16_to_fp32(data_f16[i]);
                    }
                } else {
                    data_f32.resize(nelements);
                    finp.read(reinterpret_cast<char *>(data_f32.data()), nelements * sizeof(float));
                }

                ftype = itype;
            } else {
                const int bpe = (ftype == 0) ? sizeof(float) : sizeof(uint16_t);

                data_u8.resize(nelements*bpe);
                finp.read(reinterpret_cast<char *>(data_u8.data()), nelements * bpe);
            }

            fout.write(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
            fout.write(reinterpret_cast<char *>(&length), sizeof(length));
            fout.write(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
            for (int i = 0; i < n_dims; ++i) {
                fout.write(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
            }
            fout.write(&name[0], length);

            if (quantize) {
                printf("quantizing .. ");
                work.resize(nelements); // for quantization

                size_t cur_size = 0;
                std::vector<int64_t> hist_cur(1 << 4, 0);

                switch (type) {
                    case GGML_TYPE_Q4_0:
                        {
                            cur_size = ggml_quantize_q4_0(data_f32.data(), work.data(), nelements, ne[0], QK, hist_cur.data());
                        } break;
                    case GGML_TYPE_Q4_1:
                        {
                            cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], QK, hist_cur.data());
                        } break;
                    default:
                        {
                            fprintf(stderr, "%s: unsupported quantization type %d\n", __func__, type);
                            return false;
                        }
                }

                fout.write(reinterpret_cast<char *>(work.data()), cur_size);
                total_size_new += cur_size;

                printf("size = %8.2f MB -> %8.2f MB | hist: ", nelements * sizeof(float)/1024.0/1024.0, cur_size/1024.0/1024.0);
                for (int i = 0; i < hist_cur.size(); ++i) {
                    hist_all[i] += hist_cur[i];
                }

                for (int i = 0; i < hist_cur.size(); ++i) {
                    printf("%5.3f ", hist_cur[i] / (float)nelements);
                }
                printf("\n");
            } else {
                printf("size = %8.3f MB\n", data_u8.size()/1024.0/1024.0);
                fout.write(reinterpret_cast<char *>(data_u8.data()), data_u8.size());
                total_size_new += data_u8.size();
            }

            total_size_org += nelements * sizeof(float);
        }

        printf("%s: model size  = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
        printf("%s: quant size  = %8.2f MB\n", __func__, total_size_new/1024.0/1024.0);

        {
            int64_t sum_all = 0;
            for (int i = 0; i < hist_all.size(); ++i) {
                sum_all += hist_all[i];
            }

            printf("%s: hist: ", __func__);
            for (int i = 0; i < hist_all.size(); ++i) {
                printf("%5.3f ", hist_all[i] / (float)sum_all);
            }
            printf("\n");
        }
    }

    finp.close();
    fout.close();

    return true;
}

// usage:
//  ./llama-quantize models/llama/ggml-model.bin models/llama/ggml-model-quant.bin type
//
int main(int argc, char ** argv) {
    if (argc != 4) {
        fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
        fprintf(stderr, "  type = 2 - q4_0\n");
        fprintf(stderr, "  type = 3 - q4_1\n");
        return 1;
    }

    // needed to initialize f16 tables
    {
        struct ggml_init_params params = { 0, NULL };
        struct ggml_context * ctx = ggml_init(params);
        ggml_free(ctx);
    }

    const std::string fname_inp = argv[1];
    const std::string fname_out = argv[2];

    const int itype = atoi(argv[3]);

    const int64_t t_main_start_us = ggml_time_us();

    int64_t t_quantize_us = 0;

    // load the model
    {
        const int64_t t_start_us = ggml_time_us();

        if (!llama_model_quantize(fname_inp, fname_out, itype)) {
            fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
            return 1;
        }

        t_quantize_us = ggml_time_us() - t_start_us;
    }

    // report timing
    {
        const int64_t t_main_end_us = ggml_time_us();

        printf("\n");
        printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0f);
        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
    }

    return 0;
}