forked from karpathy/llm.c
-
Notifications
You must be signed in to change notification settings - Fork 0
/
trainGPT2.c
385 lines (319 loc) · 14.9 KB
/
trainGPT2.c
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
/** @file trainGPT2.c
* @brief Source file for the functions to train GPT 2.
* @date Created on 2024/04/09
*/
/* Includes ---------------------------------------------------------------------------------- */
#include "trainGPT2.h"
/* Private define ---------------------------------------------------------------------------- */
#define MODEL_HEADER_SIZE (256)
#define MAX_SEQ_LEN_INDEX (2)
#define VOCAB_SIZE_INDEX (3)
#define NUM_LAYERS_INDEX (4)
#define NUM_HEADS_INDEX (5)
#define CHANNELS_INDEX (6)
/* Private typedef --------------------------------------------------------------------------- */
/* Private macro ----------------------------------------------------------------------------- */
/* Private enum ------------------------------------------------------------------------------ */
/* Private struct ---------------------------------------------------------------------------- */
/* Private variables ------------------------------------------------------------------------- */
/* Private function prototypes --------------------------------------------------------------- */
static float *prv_pfMallocAndPointParameters(xParameterTensors_t *xParams, size_t *pulParamSizes);
static void prv_vGpt2BuildFromCheckpoint(xGPT2_t *pxModel, char *pstrCheckpointPath);
static void prv_vDataloaderInit(xDataLoader_t *xpLoader, char *pstrFileName, uint8_t ucB, uint8_t ucT);
static void prv_vDataLoaderReset(xDataLoader_t *pxLoader);
/* Private functions ------------------------------------------------------------------------- */
static float *prv_pfMallocAndPointParameters(xParameterTensors_t *xParams, size_t *pulParamSizes)
{
size_t ulNumParameters = 0;
for (size_t ucI = 0; ucI < NUM_PARAMETER_TENSORS; ucI++)
{
ulNumParameters += pulParamSizes[ucI];
}
/* Malloc all parameters all at once */
float *pfParamsMemory = malloc(ulNumParameters * sizeof(float));
/* Assign all the tensors */
float **ppPtrs[] =
{
&xParams->pfWte, &xParams->pfWpe, &xParams->pfLn1w, &xParams->pfLn1b, &xParams->pfQkvw, &xParams->pfQkvb,
&xParams->pfAttprojw, &xParams->pfAttprojb, &xParams->pfLn2w, &xParams->pfLn2b, &xParams->pfFcw, &xParams->pfFcb,
&xParams->pfFcprojw, &xParams->pfFcprojb, &xParams->pfLnfw, &xParams->pfLnfb
};
float *pfParamsMemoryIterator = pfParamsMemory;
for (size_t ucI = 0; ucI < NUM_PARAMETER_TENSORS; ucI++)
{
*(ppPtrs[ucI]) = pfParamsMemoryIterator;
pfParamsMemoryIterator += pulParamSizes[ucI];
}
return pfParamsMemory;
}
static void prv_vGpt2BuildFromCheckpoint(xGPT2_t *pxModel, char *pstrCheckpointPath)
{
/* Read in model from a checkpoint file */
FILE *pxModelFile = fopen(pstrCheckpointPath, "rb");
int32_t slModelHeader[MODEL_HEADER_SIZE] = { 0 };
int32_t slMaxT = 0;
int32_t slV = 0;
int32_t slL = 0;
int32_t slNH = 0;
int32_t slC = 0;
size_t ulNumParameters = 0;
if (!pxModelFile)
{
printf("Error opening model file\n");
exit(1);
}
fread(slModelHeader, sizeof(int), MODEL_HEADER_SIZE, pxModelFile);
if (slModelHeader[0] != 20240326)
{
printf("Bad magic model file");
exit(1);
}
if (slModelHeader[1] != 1)
{
printf("Bad version in model file");
exit(1);
}
/* Read in hyperparameters */
pxModel->xConfig.usMaxSeqLen = (uint8_t)slModelHeader[MAX_SEQ_LEN_INDEX];
slMaxT = slModelHeader[MAX_SEQ_LEN_INDEX];
pxModel->xConfig.usVocabSize = (uint16_t)slModelHeader[VOCAB_SIZE_INDEX];
slV = slModelHeader[VOCAB_SIZE_INDEX];
pxModel->xConfig.ucNumLayers = (uint8_t)slModelHeader[NUM_LAYERS_INDEX];
slL = slModelHeader[NUM_LAYERS_INDEX];
pxModel->xConfig.ucNumHeads = (uint8_t)slModelHeader[NUM_HEADS_INDEX];
slNH = slModelHeader[NUM_HEADS_INDEX];
pxModel->xConfig.usChannels = (uint16_t)slModelHeader[CHANNELS_INDEX];
slC = slModelHeader[CHANNELS_INDEX];
printf("[GPT-2]\n");
printf("max_seq_len: %u\n", slMaxT);
printf("vocab_size: %u\n", slV);
printf("num_layers: %u\n", slL);
printf("num_heads: %u\n", slNH);
printf("channels: %u\n", slC);
/* Allocate space for all the parameters and read them in */
pxModel->aulParamSizes[0] = slV * slC;
pxModel->aulParamSizes[1] = slMaxT * slC;
pxModel->aulParamSizes[2] = slL * slC;
pxModel->aulParamSizes[3] = slL * slC;
pxModel->aulParamSizes[4] = slL * (3 * slC) * slC;
pxModel->aulParamSizes[5] = slL * (3 * slC);
pxModel->aulParamSizes[6] = slL * slC * slC;
pxModel->aulParamSizes[7] = slL * slC;
pxModel->aulParamSizes[8] = slL * slC;
pxModel->aulParamSizes[9] = slL * slC;
pxModel->aulParamSizes[10] = slL * (4 * slC) * slC;
pxModel->aulParamSizes[11] = slL * (4 * slC);
pxModel->aulParamSizes[12] = slL * slC * (4 * slC);
pxModel->aulParamSizes[13] = slL * slC;
pxModel->aulParamSizes[14] = slC;
pxModel->aulParamSizes[15] = slC;
/* Count the number of paramaters */
for (size_t ucI = 0; ucI < NUM_PARAMETER_TENSORS; ucI++)
{
ulNumParameters += pxModel->aulParamSizes[ucI];
}
printf("num_parameters: %zu\n", ulNumParameters);
pxModel->ulNumParameters = ulNumParameters;
/* Read in all the parameters from file */
pxModel->pfParamsMemory = prv_pfMallocAndPointParameters(&pxModel->xParams, pxModel->aulParamSizes);
fread(pxModel->pfParamsMemory, sizeof(float), ulNumParameters, pxModelFile);
fclose(pxModelFile);
/* Other inits */
pxModel->pfActsMemory = NULL;
pxModel->pfGradsMemory = NULL;
pxModel->pfMemoryM = NULL;
pxModel->pfMemoryV = NULL;
pxModel->pfGradsActsMemory = NULL;
pxModel->pulInputs = NULL;
pxModel->pulTargets = NULL;
pxModel->ulBatchSize = 0;
pxModel->ulSeqLen = 0;
pxModel->fMeanLoss = -1.0f; /* -1.0f will designate no loss */
}
/*---------------------------------------------------------------------------------------------*/
static void prv_vDataloaderInit(xDataLoader_t *xpLoader, char *pstrFileName, uint8_t ucB, uint8_t ucT)
{
xpLoader->ucB = ucB;
xpLoader->ucT = ucT;
/* Open the input file for reading */
xpLoader->pxTokensFile = fopen(pstrFileName, "rb");
if(xpLoader->pxTokensFile ==NULL)
{
printf("Error opening tokens file\n");
exit(1);
}
fseek(xpLoader->pxTokensFile, 0, SEEK_END);
xpLoader->ullFileSize = ftell(xpLoader->pxTokensFile);
fseek(xpLoader->pxTokensFile, 0, SEEK_SET);
if (xpLoader->ullFileSize < (ucB * ucT + 1) * sizeof(int)) {
printf("Error: file size is too small for the batch size and sequence length\n");
exit(1);
}
/* start at the beginning */
xpLoader->ullCurrentPosition = 0;
/* allocate space for B*T + 1 integers to store the inputs and targets */
xpLoader->pulBatch = malloc((ucB * ucT + 1) * sizeof(int));
xpLoader->pulInputs = xpLoader->pulBatch;
xpLoader->pulTargets = xpLoader->pulBatch + 1; // targets are shifted by one
xpLoader->ulNumBatches = xpLoader->ullFileSize / (ucB * ucT * sizeof(int));
}
/*---------------------------------------------------------------------------------------------*/
static void prv_vDataLoaderReset(xDataLoader_t *pxLoader)
{
pxLoader->ullCurrentPosition = 0;
}
/*---------------------------------------------------------------------------------------------*/
static void prv_vDataloaderNextBatch(xDataLoader_t *pxLoader)
{
uint8_t ucB = pxLoader->ucB;
uint8_t ucT = pxLoader->ucT;
/* if we are at the end of the file, loop back to the beginning */
if (pxLoader->ullCurrentPosition + (ucB*ucT+1) * sizeof(int) > pxLoader->ullFileSize)
{
pxLoader->ullCurrentPosition = 0;
}
/* read the B*T+1 integers from the file into batch */
fseek(pxLoader->pxTokensFile, pxLoader->ullCurrentPosition, SEEK_SET);
fread(pxLoader->pulBatch, sizeof(int), ucB*ucT+1, pxLoader->pxTokensFile);
/* advance the current position by B*T integers */
pxLoader->ullCurrentPosition += ucB*ucT * sizeof(int);
}
static void prv_vGpt2Forward(xGPT2_t *pxModel, uint32_t *pulInputs, uint32_t *pulTargets, uint8_t ucB, uint8_t ucT)
{
/* convenience parameters */
uint16_t usV = pxModel->xConfig.usVocabSize;
uint8_t ucL = pxModel->xConfig.ucNumLayers;
uint8_t ucNH = pxModel->xConfig.ucNumHeads;
uint8_t ucC = pxModel->xConfig.usChannels;
float *pfResidual;
/* ensure the model was initialized or error out */
if (pxModel->pfParamsMemory == NULL)
{
printf("Error: model was not initialized properly.\n");
exit(1);
}
/* allocate space for all the activations if needed (done here, lazily) */
if(pxModel->pfActsMemory == NULL)
{
/* record the current B,T as well */
pxModel->ulBatchSize = ucB;
pxModel->ulSeqLen = ucT;
pxModel->ulActSizes[0] = ucB * ucT * ucC; // encoded
pxModel->ulActSizes[1] = ucL * ucB * ucT * ucC; // ln1
pxModel->ulActSizes[2] = ucL * ucB * ucT; // ln1_mean
pxModel->ulActSizes[3] = ucL * ucB * ucT; // ln1_rstd
pxModel->ulActSizes[4] = ucL * ucB * ucT * 3*ucC; // qkv
pxModel->ulActSizes[5] = ucL * ucB * ucT * ucC; // atty
pxModel->ulActSizes[6] = ucL * ucB * ucNH * ucT * ucT; // preatt
pxModel->ulActSizes[7] = ucL * ucB * ucNH * ucT * ucT; // att
pxModel->ulActSizes[8] = ucL * ucB * ucT * ucC; // attproj
pxModel->ulActSizes[9] = ucL * ucB * ucT * ucC; // residual2
pxModel->ulActSizes[10] = ucL * ucB * ucT * ucC; // ln2
pxModel->ulActSizes[11] = ucL * ucB * ucT; // ln2_mean
pxModel->ulActSizes[12] = ucL * ucB * ucT; // ln2_rstd
pxModel->ulActSizes[13] = ucL * ucB * ucT * 4 * ucC; // fch
pxModel->ulActSizes[14] = ucL * ucB * ucT * 4 * ucC; // fch_gelu
pxModel->ulActSizes[15] = ucL * ucB * ucT * ucC; // fcproj
pxModel->ulActSizes[16] = ucL * ucB * ucT * ucC; // residual3
pxModel->ulActSizes[17] = ucB * ucT * ucC; // lnf
pxModel->ulActSizes[18] = ucB * ucT; // lnf_mean
pxModel->ulActSizes[19] = ucB * ucT; // lnf_rstd
pxModel->ulActSizes[20] = ucB * ucT * usV; // logits
pxModel->ulActSizes[21] = ucB * ucT * usV; // probs
pxModel->ulActSizes[22] = ucB * ucT; // losses
size_t ulNumActivations = 0;
for (size_t i = 0; i < NUM_ACTIVATION_TENSORS; i++)
{
ulNumActivations += pxModel->ulActSizes[i];
}
printf("num_activations: %zu\n", ulNumActivations);
pxModel->ulNumActivations = ulNumActivations;
pxModel->pfActsMemory = malloc_and_point_activations(&pxModel->xActs, pxModel->ulActSizes);
/* also create memory for caching inputs and targets */
pxModel->pulInputs = malloc(ucB * ucT * sizeof(int));
/* might be unused if we never have targets but it's small */
pxModel->pulTargets = malloc(ucB * ucT * sizeof(int));
}
else
{
/* validate B,T is no larger than what was previously allocated */
/* in principle, we could re-allocate a larger chunk of memory, for now we just error out */
if (ucB > pxModel->ulBatchSize || ucT > pxModel->ulSeqLen)
{
printf("Error: batch size or sequence length is inadequately large\n");
printf("Model: B=%d T=%d, Desired: B=%d T=%d\n", pxModel->ulBatchSize, pxModel->ulSeqLen, ucB, ucT);
exit(1);
}
}
/* cache the inputs/targets */
memcpy(pxModel->pulInputs, pulInputs, ucB * ucT * sizeof(int));
if (pulTargets != NULL)
{
memcpy(pxModel->pulTargets, pulTargets, ucB * ucT * sizeof(int));
}
/* forward pass */
xParameterTensors_t xParams = pxModel->xParams; /* for brevity */
xActivationTensors_t xActs = pxModel->xActs;
encoder_forward(xActs.pfEncoded, pulInputs, xParams.pfWte, xParams.pfWpe, ucB, ucT, ucC); /* encoding goes into residual[0] */
for (uint8_t ucl = 0; ucl < ucL; ucl++)
{
pfResidual = ucl == 0 ? xActs.pfEncoded : xActs.pfResidual3 + (ucl-1) * ucB * ucT * ucC;
// get the pointers of the weights for this layer
float *pfLln1w = xParams.pfLn1w + ucl * ucC;
float *pfLln1b = xParams.pfLn1b + ucl * ucC;
float *pfLqkvw = xParams.pfQkvw + ucl * 3 * ucC * ucC;
float *pfLqkvb = xParams.pfQkvb + ucl * 3 * ucC;
float *pfLattprojw = xParams.pfAttprojw + ucl * ucC * ucC;
float *pfLattprojb = xParams.pfAttprojb + ucl * ucC;
float *pfLln2w = xParams.pfLn2w + ucl * ucC;
float *pfLln2b = xParams.pfLn2b + ucl * ucC;
float *pfLfcw = xParams.pfFcw + ucl * 4 * ucC * ucC;
float *pfLfcb = xParams.pfFcb + ucl * 4 * ucC;
float *pfLfcprojw = xParams.pfFcprojw + ucl * ucC * 4 * ucC;
float *pfLfcprojb = xParams.pfFcprojb + ucl * ucC;
}
}
/* Exported functions ------------------------------------------------------------------------ */
int main(void)
{
/* Build the GPT-2 model from a checkpoint */
xGPT2_t xModel;
prv_vGpt2BuildFromCheckpoint(&xModel, "gpt2_124M.bin");
/* Build the DataLoaders from tokens files, for now use tiny_shakespeare if available, else tiny_stories */
char *pstrTinyStoriesTrain = "data/TinyStories_train.bin";
char *pstrTinyStoriesVal = "data/TinyStories_val.bin";
char *pstrTinyShakespeareTrain = "data/tiny_shakespeare_train.bin";
char *pstrTinyShakespeareVal = "data/tiny_shakespeare_val.bin";
char *pstrTrainTokens = access(pstrTinyShakespeareTrain, F_OK) != -1 ? pstrTinyShakespeareTrain : pstrTinyStoriesTrain;
char *pstrTalTokens = access(pstrTinyShakespeareVal, F_OK) != -1 ? pstrTinyShakespeareVal : pstrTinyStoriesVal;
uint8_t ucB = 4;
uint8_t ucT = 64;
xDataLoader_t xTrainLoader;
xDataLoader_t xValLoader;
uint8_t ucValNumBatches = 10;
uint64_t ullRngState = 1337;
const uint8_t ucGenMaxLength = 64; /* during inference step we'll generate sequences of this many tokens */
uint32_t ulGenTokens[ucGenMaxLength];
struct timespec xStart, xEnd;
prv_vDataloaderInit(&xTrainLoader, pstrTrainTokens, ucB, ucT);
printf("train dataset num_batches: %d\n", xTrainLoader.ulNumBatches);
prv_vDataloaderInit(&xValLoader, pstrTalTokens, ucB, ucT);
printf("val dataset num_batches: %d\n", xValLoader.ulNumBatches);
/* some memory for generating samples from the model */
/* train */
for (uint8_t ucStep = 0; ucStep <= 20; ucStep++)
{
/* once in a while estimate the validation loss */
if (ucStep % 10 == 0)
{
float fValLoss = 0.0f;
prv_vDataLoaderReset(&xValLoader);
for (uint8_t ucI = 0; ucI < ucValNumBatches; ucI++)
{
prv_vDataloaderNextBatch(&xValLoader);
}
}
}
return 0;
}
/*---------------------------------------------------------------------------------------------*/