File size: 3,430 Bytes
1485644 |
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 |
#pragma once
//#include <stdbool.h>
#define PRINT(C) fputc((char)C, stdout), fflush(stdout)
typedef enum {false, true} bool;
typedef struct Sampler Sampler;
struct Sampler {
Mamba *model;
Tokenizer *tokenizer;
uint64_t rng_seed;
fp32_t temperature;
bool verbose;
bool (*generate) (Sampler *, char *, uint64_t);
uint64_t (*sample) (Sampler *, fp32_t *);
};
static void softmax(fp32_t* x, uint64_t size) {
fp32_t max_val = x[0];
for (uint64_t i = 1; i < size; ++i)
if (x[i] > max_val) max_val = x[i];
fp32_t sum = 0.0f;
for (uint64_t i = 0; i < size; ++i) {
x[i] = expf(x[i] - max_val);
sum += x[i];
}
for (uint64_t i = 0; i < size; ++i)
x[i] /= sum;
}
static uint64_t random_u32(uint64_t *rng_seed) {
*rng_seed ^= *rng_seed >> 12;
*rng_seed ^= *rng_seed << 25;
*rng_seed ^= *rng_seed >> 27;
*rng_seed = (*rng_seed * 0x2545F4914F6CDD1Dull) >> 32;
return *rng_seed;
}
static inline fp32_t random_f32(uint64_t *rng_seed) { return (random_u32(rng_seed) >> 8) / 16777216.0f; }
static uint64_t time_in_ms() {
struct timeval tv;
gettimeofday(&tv, NULL);
return tv.tv_sec * 1000 + tv.tv_usec / 1000;
}
static inline uint64_t sample_argmax(fp32_t* probabilities, uint64_t n) {
uint64_t max_i = 0;
fp32_t max_p = probabilities[0];
for (uint64_t i = 1; i < n; ++i)
if (probabilities[i] > max_p)
max_i = i, max_p = probabilities[i];
return max_i;
}
static inline uint64_t sample_mult(fp32_t* probabilities, uint64_t n, fp32_t coin) {
fp32_t cdf = 0.0f;
for (uint64_t i = 0; i < n; ++i) {
cdf += probabilities[i];
if (coin < cdf) return i;
}
return n - 1;
}
static uint64_t SamplerSample(Sampler *sampler, fp32_t* logits) {
uint64_t next,
vocab_size = sampler->tokenizer->vocab_size,
*rng_seed = &sampler->rng_seed;
//printf("Vocab size: %llu\n", vocab_size);
fp32_t temperature = sampler->temperature;
if (temperature == 0.0f) next = sample_argmax(logits, vocab_size);
else {
for (uint64_t q = 0; q < vocab_size; ++q)
logits[q] /= temperature;
softmax(logits, vocab_size);
fp32_t coin = random_f32(rng_seed);
next = sample_mult(logits, vocab_size, coin);
}
return next;
}
static bool SamplerGenerate(Sampler *sampler, char *seed_text, uint64_t n_predict) {
Mamba *model = sampler->model;
Tokenizer *tokenizer = sampler->tokenizer;
uint64_t vocab_size = tokenizer->vocab_size;
fp32_t temperature = sampler->temperature;
bool verbose = sampler->verbose;
uint64_t token;
fp32_t *logits;
char *text;
if (seed_text == NULL) return EXIT_FAILURE;
for (; *seed_text; ) {
token = tokenizer->encode(tokenizer, (uint8_t **) &seed_text);
text = tokenizer->decode(tokenizer, token);
fputs(text, stdout);
fflush(stdout);
logits = model->forward(model, token);
}
uint64_t time_start;
if (verbose) time_start = time_in_ms();
for (uint64_t i = 0; i < n_predict; ++i) {
token = sampler->sample(sampler, logits);
text = tokenizer->decode(tokenizer, token);
fputs(text, stdout);
fflush(stdout);
logits = model->forward(model, token);
}
CLOG(verbose, "\nachieved tok/s: %f\n", n_predict / (double)(time_in_ms() - time_start) * 1000);
return EXIT_SUCCESS;
}
Sampler sampler = {
.model = &mamba,
.tokenizer = &tokenizer,
.rng_seed = 42,
.temperature = 0.0f,
.verbose = false,
.generate = SamplerGenerate,
.sample = SamplerSample
};
|