File size: 23,847 Bytes
f8c5b0d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
#include "ggml_v1.h"
#include "otherarch.h"

#include "utils.h"

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



// load the model's weights from a file
ModelLoadResult legacy_gpt2_model_load(const std::string & fname, gpt2_v1_model & model, gpt_vocab & vocab, FileFormat file_format) {
    printf("%s: loading model from '%s'\n", __func__, fname.c_str());

    auto fin = std::ifstream(fname, std::ios::binary);
    if (!fin) {
        fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
        return ModelLoadResult::FAIL;
    }

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

    auto desiredMaxCtx = model.hparams.n_ctx;

    // load hparams
    {
        auto & hparams = model.hparams;

        
        fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
        fin.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
        fin.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
        fin.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
        fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
        fin.read((char *) &hparams.ftype,     sizeof(hparams.ftype));

        //used to expand KV size if needed
        desiredMaxCtx = std::max(hparams.n_ctx,desiredMaxCtx);

        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_head  = %d\n", __func__, hparams.n_head);
        printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
        printf("%s: f16     = %d\n", __func__, hparams.ftype);
    }

    // load vocab
    {
        int32_t n_vocab = 0;
        fin.read((char *) &n_vocab, sizeof(n_vocab));

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

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

            word.resize(len);
            fin.read((char *) word.data(), len);

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

    // for the big tensors, we have the option to store the data in 16-bit floats
    // in order to save memory and also to speed up the computation
    const ggml_v1_type wtype = GGML_V1_TYPE_F16;

    auto & ctx = model.ctx;

    auto memory_type = GGML_V1_TYPE_F16;

    size_t ctx_size = 0;

    {
        const auto & hparams = model.hparams;

        const int n_embd  = hparams.n_embd;
        const int n_layer = hparams.n_layer;
        const int n_ctx   = desiredMaxCtx;
        const int n_vocab = hparams.n_vocab;

        ctx_size += n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32); // ln_f_g
        ctx_size += n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32); // ln_f_b

        ctx_size += n_vocab*n_embd*ggml_v1_type_size(wtype);         // wte
        ctx_size +=   n_ctx*n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32); // wpe

        ctx_size += n_layer*(n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32)); // ln_1_g
        ctx_size += n_layer*(n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32)); // ln_1_b

        ctx_size += n_layer*(n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32)); // ln_2_g
        ctx_size += n_layer*(n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32)); // ln_2_b

        ctx_size += n_layer*(3*n_embd*n_embd*ggml_v1_type_size(wtype));         // c_attn_attn_w
        ctx_size += n_layer*(       3*n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32)); // c_attn_attn_b

        ctx_size += n_layer*(n_embd*n_embd*ggml_v1_type_size(wtype));           // c_attn_proj_w
        ctx_size += n_layer*(       n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32));   // c_attn_proj_b

        ctx_size += n_layer*(4*n_embd*n_embd*ggml_v1_type_size(wtype));         // c_mlp_fc_w
        ctx_size += n_layer*(       4*n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32)); // c_mlp_fc_b

        ctx_size += n_layer*(4*n_embd*n_embd*ggml_v1_type_size(wtype));         // c_mlp_proj_w
        ctx_size += n_layer*(         n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32)); // c_mlp_proj_b

        ctx_size += n_ctx*n_layer*n_embd*ggml_v1_type_size(memory_type); // memory_k
        ctx_size += n_ctx*n_layer*n_embd*ggml_v1_type_size(memory_type); // memory_v

        ctx_size += (6 + 12*n_layer)*256; // object overhead

        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
    }

    // create the ggml context
    {
        struct ggml_v1_init_params params;        
        params.mem_size   = ctx_size,
        params.mem_buffer = NULL,
        

        model.ctx = ggml_v1_init(params);
        if (!model.ctx) {
            fprintf(stderr, "%s: ggml_v1_init() failed\n", __func__);
            return ModelLoadResult::FAIL;
        }
    }

    // prepare memory for the weights
    {
        const auto & hparams = model.hparams;

        const int n_embd  = hparams.n_embd;
        const int n_layer = hparams.n_layer;
        const int n_ctx   = hparams.n_ctx;
        const int n_vocab = hparams.n_vocab;

        model.layers.resize(n_layer);

        model.ln_f_g = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32, n_embd);
        model.ln_f_b = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32, n_embd);

        model.wte = ggml_v1_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
        model.wpe = ggml_v1_new_tensor_2d(ctx, GGML_V1_TYPE_F32, n_embd, n_ctx);

        // map by name
        model.tensors["model/ln_f/g"] = model.ln_f_g;
        model.tensors["model/ln_f/b"] = model.ln_f_b;

        model.tensors["model/wte"] = model.wte;
        model.tensors["model/wpe"] = model.wpe;

        for (int i = 0; i < n_layer; ++i) {
            auto & layer = model.layers[i];

            layer.ln_1_g             = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32,   n_embd);
            layer.ln_1_b             = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32,   n_embd);

            layer.ln_2_g             = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32,   n_embd);
            layer.ln_2_b             = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32,   n_embd);

            layer.c_attn_attn_w      = ggml_v1_new_tensor_2d(ctx, wtype,         3*n_embd, n_embd);
            layer.c_attn_attn_b      = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32, 3*n_embd);

            layer.c_attn_proj_w      = ggml_v1_new_tensor_2d(ctx, wtype,           n_embd, n_embd);
            layer.c_attn_proj_b      = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32,   n_embd);

            layer.c_mlp_fc_w         = ggml_v1_new_tensor_2d(ctx, wtype,         4*n_embd, n_embd);
            layer.c_mlp_fc_b         = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32, 4*n_embd);

            layer.c_mlp_proj_w_trans = ggml_v1_new_tensor_2d(ctx, wtype,         4*n_embd, n_embd);
            layer.c_mlp_proj_b       = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32,   n_embd);

            // map by name
            model.tensors["model/h" + std::to_string(i) + "/ln_1/g"]        = layer.ln_1_g;
            model.tensors["model/h" + std::to_string(i) + "/ln_1/b"]        = layer.ln_1_b;

            model.tensors["model/h" + std::to_string(i) + "/ln_2/g"]        = layer.ln_2_g;
            model.tensors["model/h" + std::to_string(i) + "/ln_2/b"]        = layer.ln_2_b;

            model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/w"] = layer.c_attn_attn_w;
            model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/b"] = layer.c_attn_attn_b;

            model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/w"] = layer.c_attn_proj_w;
            model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/b"] = layer.c_attn_proj_b;

            model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/w"]    = layer.c_mlp_fc_w;
            model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/b"]    = layer.c_mlp_fc_b;

            model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/w"]  = layer.c_mlp_proj_w_trans;
            model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/b"]  = layer.c_mlp_proj_b;
        }
    }

    // key + value memory
    {
        const auto & hparams = model.hparams;

        const int n_embd  = hparams.n_embd;
        const int n_layer = hparams.n_layer;
        const int n_ctx   = desiredMaxCtx;

        const int n_mem      = n_layer*n_ctx;
        const int n_elements = n_embd*n_mem;

        model.memory_k = ggml_v1_new_tensor_1d(ctx, memory_type, n_elements);
        model.memory_v = ggml_v1_new_tensor_1d(ctx, memory_type, n_elements);

        const size_t memory_size = ggml_v1_nbytes(model.memory_k) + ggml_v1_nbytes(model.memory_v);

        printf("%s: memory size = %8.2f MB, n_mem = %d\n", __func__, memory_size/1024.0/1024.0, n_mem);
    }

    // load weights
    {
        size_t total_size = 0;

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

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

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

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

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

            if (model.tensors.find(name.data()) == model.tensors.end()) {
                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
                return ModelLoadResult::FAIL;
            }

            auto tensor = model.tensors[name.data()];
            if (ggml_v1_nelements(tensor) != nelements) {
                fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
                return ModelLoadResult::FAIL;
            }

            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
                 //test for transposition and retry older loader
                if(tensor->ne[0]==ne[1] && tensor->ne[1]==ne[0] && should_transpose_layer(name))
                {
                    printf("\nFound a transposed tensor. This could be an older or newer model. Retrying load...");
                    ggml_v1_free(ctx);
                    return ModelLoadResult::RETRY_LOAD;
                }
                else
                {
                    fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
                            __func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]);
                    return ModelLoadResult::FAIL;
                }
            }

            const size_t bpe = (ftype == 0) ? sizeof(float) : sizeof(ggml_v1_fp16_t);

            if (nelements*bpe != ggml_v1_nbytes(tensor)) {
                fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
                        __func__, name.data(), ggml_v1_nbytes(tensor), nelements*bpe);
                return ModelLoadResult::FAIL;
            }

            fin.read(reinterpret_cast<char *>(tensor->data), ggml_v1_nbytes(tensor));

            //printf("%24s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ftype == 0 ? "float" : "f16", ggml_v1_nbytes(tensor)/1024.0/1024.0);
            total_size += ggml_v1_nbytes(tensor);
        }

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

    fin.close();

    return ModelLoadResult::SUCCESS;
}

// evaluate the transformer
//
//   - model:     the model
//   - n_threads: number of threads to use
//   - n_past:    the context size so far
//   - embd_inp:  the embeddings of the tokens in the context
//   - embd_w:    the predicted logits for the next token
//
bool legacy_gpt2_eval(
        const gpt2_v1_model & model,
        const int n_threads,
        const int n_past,
        const std::vector<gpt_vocab::id> & embd_inp,
              std::vector<float>         & embd_w,
              size_t                     & mem_per_token,
              FileFormat file_format) {
    const int N = embd_inp.size();

    const auto & hparams = model.hparams;

    const int n_embd  = hparams.n_embd;
    const int n_layer = hparams.n_layer;
    const int n_ctx   = hparams.n_ctx;
    const int n_head  = hparams.n_head;
    const int n_vocab = hparams.n_vocab;

    static size_t buf_size = 256u*1024*1024;
    static void * buf = malloc(buf_size);

    if (mem_per_token > 0 && mem_per_token*N > buf_size) {
        const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead
        //printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);

        // reallocate
        buf_size = buf_size_new;
        buf = realloc(buf, buf_size);
        if (buf == nullptr) {
            fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size);
            return false;
        }
    }

    struct ggml_v1_init_params params;
    params.mem_size   = buf_size;
    params.mem_buffer = buf;
    

    struct ggml_v1_context * ctx0 = ggml_v1_init(params);
    struct ggml_v1_cgraph gf = {};
    gf.n_threads = n_threads;

    struct ggml_v1_tensor * embd = ggml_v1_new_tensor_1d(ctx0, GGML_V1_TYPE_I32, N);
    memcpy(embd->data, embd_inp.data(), N*ggml_v1_element_size(embd));

    struct ggml_v1_tensor * position = ggml_v1_new_tensor_1d(ctx0, GGML_V1_TYPE_I32, N);
    for (int i = 0; i < N; ++i) {
        ((int32_t *) position->data)[i] = n_past + i;
    }

    // wte + wpe
    struct ggml_v1_tensor * inpL =
        ggml_v1_add(ctx0,
                ggml_v1_get_rows(ctx0, model.wte, embd),
                ggml_v1_get_rows(ctx0, model.wpe, position));

    for (int il = 0; il < n_layer; ++il) {
        struct ggml_v1_tensor * cur;

        // norm
        {
            // [ 768, N]
            cur = ggml_v1_norm(ctx0, inpL);

            // cur = ln_1_g*cur + ln_1_b
            // [ 768, N]
            cur = ggml_v1_add(ctx0,
                    ggml_v1_mul(ctx0,
                        ggml_v1_repeat(ctx0, model.layers[il].ln_1_g, cur),
                        cur),
                    ggml_v1_repeat(ctx0, model.layers[il].ln_1_b, cur));
        }

        // attn
        // [2304, 768] - model.layers[il].c_attn_attn_w
        // [2304,   1] - model.layers[il].c_attn_attn_b
        // [ 768,   N] - cur (in)
        // [2304,   N] - cur (out)
        //
        // cur = attn_w*cur + attn_b
        // [2304, N]
        {
            cur = ggml_v1_mul_mat(ctx0,
                    ggml_v1_transpose(ctx0, model.layers[il].c_attn_attn_w),
                    cur);

            cur = ggml_v1_add(ctx0,
                    ggml_v1_repeat(ctx0, model.layers[il].c_attn_attn_b, cur),
                    cur);
        }

        // self-attention
        {
            struct ggml_v1_tensor * Qcur = ggml_v1_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd);
            struct ggml_v1_tensor * Kcur = ggml_v1_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd);
            struct ggml_v1_tensor * Vcur = ggml_v1_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 2*sizeof(float)*n_embd);

            // store key and value to memory
            if (N >= 1) {
                struct ggml_v1_tensor * k = ggml_v1_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_v1_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past));
                struct ggml_v1_tensor * v = ggml_v1_view_1d(ctx0, model.memory_v, N*n_embd, (ggml_v1_element_size(model.memory_v)*n_embd)*(il*n_ctx + n_past));

                ggml_v1_build_forward_expand(&gf, ggml_v1_cpy(ctx0, Kcur, k));
                ggml_v1_build_forward_expand(&gf, ggml_v1_cpy(ctx0, Vcur, v));
            }

            // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
            // [64, N, 12]
            struct ggml_v1_tensor * Q =
                ggml_v1_permute(ctx0,
                        ggml_v1_cpy(ctx0,
                            Qcur,
                            ggml_v1_new_tensor_3d(ctx0, GGML_V1_TYPE_F32, n_embd/n_head, n_head, N)),
                        0, 2, 1, 3);

            // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
            // [64, n_past + N, 12]
            struct ggml_v1_tensor * K =
                ggml_v1_permute(ctx0,
                        ggml_v1_reshape_3d(ctx0,
                            ggml_v1_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_v1_element_size(model.memory_k)*n_embd),
                            n_embd/n_head, n_head, n_past + N),
                        0, 2, 1, 3);

            // GG: flash attention
            //struct ggml_v1_tensor * V =
            //    ggml_v1_cpy(ctx0,
            //            ggml_v1_permute(ctx0,
            //                ggml_v1_reshape_3d(ctx0,
            //                    ggml_v1_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_v1_element_size(model.memory_v)*n_embd),
            //                    n_embd/n_head, n_head, n_past + N),
            //                1, 2, 0, 3),
            //            ggml_v1_new_tensor_3d(ctx0, GGML_V1_TYPE_F32, n_past + N, n_embd/n_head, n_head));

            //struct ggml_v1_tensor * KQV = ggml_v1_flash_attn(ctx0, Q, K, V, true);

            // K * Q
            // [n_past + N, N, 12]
            struct ggml_v1_tensor * KQ = ggml_v1_mul_mat(ctx0, K, Q);

            // KQ_scaled = KQ / sqrt(n_embd/n_head)
            // [n_past + N, N, 12]
            struct ggml_v1_tensor * KQ_scaled =
                ggml_v1_scale(ctx0,
                        KQ,
                        ggml_v1_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
                        );

            // KQ_masked = mask_past(KQ_scaled)
            // [n_past + N, N, 12]
            struct ggml_v1_tensor * KQ_masked = ggml_v1_diag_mask_inf(ctx0, KQ_scaled, n_past);

            // KQ = soft_max(KQ_masked)
            // [n_past + N, N, 12]
            struct ggml_v1_tensor * KQ_soft_max = ggml_v1_soft_max(ctx0, KQ_masked);

            // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
            // [n_past + N, 64, 12]
            struct ggml_v1_tensor * V_trans =
                ggml_v1_permute(ctx0,
                        ggml_v1_reshape_3d(ctx0,
                            ggml_v1_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_v1_element_size(model.memory_v)*n_embd),
                            n_embd/n_head, n_head, n_past + N),
                        1, 2, 0, 3);

            // KQV = transpose(V) * KQ_soft_max
            // [64, N, 12]
            struct ggml_v1_tensor * KQV = ggml_v1_mul_mat(ctx0, V_trans, KQ_soft_max);

            // KQV_merged = KQV.permute(0, 2, 1, 3)
            // [64, 12, N]
            struct ggml_v1_tensor * KQV_merged = ggml_v1_permute(ctx0, KQV, 0, 2, 1, 3);

            // cur = KQV_merged.contiguous().view(n_embd, N)
            // [768, N]
            cur = ggml_v1_cpy(ctx0,
                    KQV_merged,
                    ggml_v1_new_tensor_2d(ctx0, GGML_V1_TYPE_F32, n_embd, N));
        }

        // projection
        // [ 768, 768] - model.layers[il].c_attn_proj_w
        // [ 768,   1] - model.layers[il].c_attn_proj_b
        // [ 768,   N] - cur (in)
        // [ 768,   N] - cur (out)
        //
        // cur = proj_w*cur + proj_b
        // [768, N]
        {
            cur = ggml_v1_mul_mat(ctx0,
                    ggml_v1_transpose(ctx0, model.layers[il].c_attn_proj_w),
                    cur);

            cur = ggml_v1_add(ctx0,
                    ggml_v1_repeat(ctx0, model.layers[il].c_attn_proj_b, cur),
                    cur);
        }

        // add the input
        cur = ggml_v1_add(ctx0, cur, inpL);

        struct ggml_v1_tensor * inpFF = cur;

        // feed-forward network
        {
            // norm
            {
                cur = ggml_v1_norm(ctx0, inpFF);

                // cur = ln_2_g*cur + ln_2_b
                // [ 768, N]
                cur = ggml_v1_add(ctx0,
                        ggml_v1_mul(ctx0,
                            ggml_v1_repeat(ctx0, model.layers[il].ln_2_g, cur),
                            cur),
                        ggml_v1_repeat(ctx0, model.layers[il].ln_2_b, cur));
            }

            // fully connected
            // [3072, 768] - model.layers[il].c_mlp_fc_w
            // [3072,   1] - model.layers[il].c_mlp_fc_b
            // [ 768,   N] - cur (in)
            // [3072,   N] - cur (out)
            //
            // cur = fc_w*cur + fc_b
            // [3072, N]
            cur = ggml_v1_mul_mat(ctx0,
                    ggml_v1_transpose(ctx0, model.layers[il].c_mlp_fc_w),
                    cur);

            cur = ggml_v1_add(ctx0,
                    ggml_v1_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur),
                    cur);

            // GELU activation
            // [3072, N]
            cur = ggml_v1_gelu(ctx0, cur);

            // projection
            // [ 768, 3072] - model.layers[il].c_mlp_proj_w
            // [ 768,    1] - model.layers[il].c_mlp_proj_b
            // [3072,    N] - cur (in)
            // [ 768,    N] - cur (out)
            //
            // cur = proj_w*cur + proj_b
            // [768, N]
            cur = ggml_v1_mul_mat(ctx0,
                    model.layers[il].c_mlp_proj_w_trans,
                    cur);

            cur = ggml_v1_add(ctx0,
                    ggml_v1_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur),
                    cur);
        }

        // input for next layer
        inpL = ggml_v1_add(ctx0, cur, inpFF);
    }

    // norm
    {
        // [ 768, N]
        inpL = ggml_v1_norm(ctx0, inpL);

        // inpL = ln_f_g*inpL + ln_f_b
        // [ 768, N]
        inpL = ggml_v1_add(ctx0,
                ggml_v1_mul(ctx0,
                    ggml_v1_repeat(ctx0, model.ln_f_g, inpL),
                    inpL),
                ggml_v1_repeat(ctx0, model.ln_f_b, inpL));
    }

    // inpL = WTE * inpL
    // [ 768, 50257] - model.wte
    // [ 768, N]     - inpL
    inpL = ggml_v1_mul_mat(ctx0, model.wte, inpL);

    // logits -> probs
    //inpL = ggml_v1_soft_max(ctx0, inpL);

    // run the computation
    ggml_v1_build_forward_expand(&gf, inpL);
    ggml_v1_graph_compute       (ctx0, &gf);

    //if (n_past%100 == 0) {
    //    ggml_v1_graph_print   (&gf);
    //    ggml_v1_graph_dump_dot(&gf, NULL, "gpt-2.dot");
    //}

    //embd_w.resize(n_vocab*N);
    //memcpy(embd_w.data(), ggml_v1_get_data(inpL), sizeof(float)*n_vocab*N);

    // return result just for the last token
    embd_w.resize(n_vocab);
    memcpy(embd_w.data(), (float *) ggml_v1_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);

    if (mem_per_token == 0) {
        mem_per_token = ggml_v1_used_mem(ctx0)/N;
    }
    //printf("used_mem = %zu\n", ggml_v1_used_mem(ctx0));

    ggml_v1_free(ctx0);

    return true;
}