2023-04-30 15:51:57 +00:00
|
|
|
#include "ggml.h"
|
|
|
|
|
|
|
|
#include "common.h"
|
|
|
|
#include "common-ggml.h"
|
|
|
|
|
|
|
|
#include <cassert>
|
|
|
|
#include <cmath>
|
|
|
|
#include <cstdio>
|
|
|
|
#include <cstring>
|
|
|
|
#include <fstream>
|
|
|
|
#include <map>
|
|
|
|
#include <string>
|
|
|
|
#include <vector>
|
|
|
|
#include <regex>
|
|
|
|
|
|
|
|
// default hparams (Whisper tiny)
|
|
|
|
struct whisper_hparams {
|
|
|
|
int32_t n_vocab = 51864;
|
|
|
|
int32_t n_audio_ctx = 1500;
|
|
|
|
int32_t n_audio_state = 384;
|
|
|
|
int32_t n_audio_head = 6;
|
|
|
|
int32_t n_audio_layer = 4;
|
|
|
|
int32_t n_text_ctx = 448;
|
|
|
|
int32_t n_text_state = 384;
|
|
|
|
int32_t n_text_head = 6;
|
|
|
|
int32_t n_text_layer = 4;
|
|
|
|
int32_t n_mels = 80;
|
2023-05-14 15:04:23 +00:00
|
|
|
int32_t ftype = 1;
|
2023-04-30 15:51:57 +00:00
|
|
|
};
|
|
|
|
|
|
|
|
struct whisper_filters {
|
|
|
|
int32_t n_mel;
|
|
|
|
int32_t n_fft;
|
|
|
|
|
|
|
|
std::vector<float> data;
|
|
|
|
};
|
|
|
|
|
|
|
|
// quantize a model
|
|
|
|
bool whisper_model_quantize(const std::string & fname_inp, const std::string & fname_out, ggml_ftype ftype) {
|
|
|
|
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));
|
2023-07-02 18:45:27 +00:00
|
|
|
if (magic != GGML_FILE_MAGIC) {
|
2023-04-30 15:51:57 +00:00
|
|
|
fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname_inp.c_str());
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
fout.write((char *) &magic, sizeof(magic));
|
|
|
|
}
|
|
|
|
|
|
|
|
whisper_hparams hparams;
|
|
|
|
|
|
|
|
// load hparams
|
|
|
|
{
|
|
|
|
finp.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
|
|
|
|
finp.read((char *) &hparams.n_audio_ctx, sizeof(hparams.n_audio_ctx));
|
|
|
|
finp.read((char *) &hparams.n_audio_state, sizeof(hparams.n_audio_state));
|
|
|
|
finp.read((char *) &hparams.n_audio_head, sizeof(hparams.n_audio_head));
|
|
|
|
finp.read((char *) &hparams.n_audio_layer, sizeof(hparams.n_audio_layer));
|
|
|
|
finp.read((char *) &hparams.n_text_ctx, sizeof(hparams.n_text_ctx));
|
|
|
|
finp.read((char *) &hparams.n_text_state, sizeof(hparams.n_text_state));
|
|
|
|
finp.read((char *) &hparams.n_text_head, sizeof(hparams.n_text_head));
|
|
|
|
finp.read((char *) &hparams.n_text_layer, sizeof(hparams.n_text_layer));
|
|
|
|
finp.read((char *) &hparams.n_mels, sizeof(hparams.n_mels));
|
2023-05-14 15:04:23 +00:00
|
|
|
finp.read((char *) &hparams.ftype, sizeof(hparams.ftype));
|
|
|
|
|
|
|
|
const int32_t qntvr_src = hparams.ftype / GGML_QNT_VERSION_FACTOR;
|
|
|
|
const int32_t ftype_dst = GGML_QNT_VERSION * GGML_QNT_VERSION_FACTOR + ftype;
|
2023-04-30 15:51:57 +00:00
|
|
|
|
|
|
|
fprintf(stderr, "%s: n_vocab = %d\n", __func__, hparams.n_vocab);
|
|
|
|
fprintf(stderr, "%s: n_audio_ctx = %d\n", __func__, hparams.n_audio_ctx);
|
|
|
|
fprintf(stderr, "%s: n_audio_state = %d\n", __func__, hparams.n_audio_state);
|
|
|
|
fprintf(stderr, "%s: n_audio_head = %d\n", __func__, hparams.n_audio_head);
|
|
|
|
fprintf(stderr, "%s: n_audio_layer = %d\n", __func__, hparams.n_audio_layer);
|
|
|
|
fprintf(stderr, "%s: n_text_ctx = %d\n", __func__, hparams.n_text_ctx);
|
|
|
|
fprintf(stderr, "%s: n_text_state = %d\n", __func__, hparams.n_text_state);
|
|
|
|
fprintf(stderr, "%s: n_text_head = %d\n", __func__, hparams.n_text_head);
|
|
|
|
fprintf(stderr, "%s: n_text_layer = %d\n", __func__, hparams.n_text_layer);
|
|
|
|
fprintf(stderr, "%s: n_mels = %d\n", __func__, hparams.n_mels);
|
2023-05-14 15:04:23 +00:00
|
|
|
fprintf(stderr, "%s: ftype (src) = %d\n", __func__, hparams.ftype);
|
|
|
|
fprintf(stderr, "%s: qntvr (src) = %d\n", __func__, qntvr_src);
|
|
|
|
fprintf(stderr, "%s: ftype (dst) = %d\n", __func__, ftype_dst);
|
|
|
|
fprintf(stderr, "%s: qntvr (dst) = %d\n", __func__, GGML_QNT_VERSION);
|
2023-04-30 15:51:57 +00:00
|
|
|
|
2023-06-25 11:22:21 +00:00
|
|
|
fout.write((const char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
|
|
|
|
fout.write((const char *) &hparams.n_audio_ctx, sizeof(hparams.n_audio_ctx));
|
|
|
|
fout.write((const char *) &hparams.n_audio_state, sizeof(hparams.n_audio_state));
|
|
|
|
fout.write((const char *) &hparams.n_audio_head, sizeof(hparams.n_audio_head));
|
|
|
|
fout.write((const char *) &hparams.n_audio_layer, sizeof(hparams.n_audio_layer));
|
|
|
|
fout.write((const char *) &hparams.n_text_ctx, sizeof(hparams.n_text_ctx));
|
|
|
|
fout.write((const char *) &hparams.n_text_state, sizeof(hparams.n_text_state));
|
|
|
|
fout.write((const char *) &hparams.n_text_head, sizeof(hparams.n_text_head));
|
|
|
|
fout.write((const char *) &hparams.n_text_layer, sizeof(hparams.n_text_layer));
|
|
|
|
fout.write((const char *) &hparams.n_mels, sizeof(hparams.n_mels));
|
|
|
|
fout.write((const char *) &ftype_dst, sizeof(hparams.ftype));
|
2023-04-30 15:51:57 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// load mel filters
|
|
|
|
{
|
|
|
|
whisper_filters filters;
|
|
|
|
|
|
|
|
finp.read ((char *) &filters.n_mel, sizeof(filters.n_mel));
|
|
|
|
fout.write((char *) &filters.n_mel, sizeof(filters.n_mel));
|
|
|
|
finp.read ((char *) &filters.n_fft, sizeof(filters.n_fft));
|
|
|
|
fout.write((char *) &filters.n_fft, sizeof(filters.n_fft));
|
|
|
|
|
|
|
|
filters.data.resize(filters.n_mel * filters.n_fft);
|
|
|
|
finp.read ((char *) filters.data.data(), filters.data.size() * sizeof(float));
|
|
|
|
fout.write((char *) filters.data.data(), filters.data.size() * sizeof(float));
|
|
|
|
}
|
|
|
|
|
|
|
|
// load vocab
|
|
|
|
{
|
|
|
|
int32_t n_vocab = 0;
|
|
|
|
finp.read ((char *) &n_vocab, sizeof(n_vocab));
|
|
|
|
fout.write((char *) &n_vocab, sizeof(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;
|
|
|
|
//}
|
|
|
|
|
2023-08-06 08:04:42 +00:00
|
|
|
char word[129];
|
2023-06-25 11:22:21 +00:00
|
|
|
|
2023-04-30 15:51:57 +00:00
|
|
|
for (int i = 0; i < n_vocab; i++) {
|
|
|
|
uint32_t len;
|
|
|
|
finp.read ((char *) &len, sizeof(len));
|
|
|
|
fout.write((char *) &len, sizeof(len));
|
|
|
|
|
2023-06-25 11:22:21 +00:00
|
|
|
word[len] = '\0';
|
|
|
|
|
|
|
|
finp.read ((char *) word, len);
|
|
|
|
fout.write((char *) word, len);
|
2023-04-30 15:51:57 +00:00
|
|
|
|
|
|
|
vocab.token_to_id[word] = i;
|
|
|
|
vocab.id_to_token[i] = word;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// regexes of tensor names to not be quantized
|
|
|
|
const std::vector<std::string> to_skip = {
|
|
|
|
//"encoder.*",
|
|
|
|
"encoder.conv1.bias",
|
|
|
|
"encoder.conv2.bias",
|
|
|
|
"encoder.positional_embedding",
|
|
|
|
"decoder.positional_embedding",
|
|
|
|
};
|
|
|
|
|
|
|
|
if (!ggml_common_quantize_0(finp, fout, ftype, { ".*" }, to_skip)) {
|
|
|
|
fprintf(stderr, "%s: failed to quantize model '%s'\n", __func__, fname_inp.c_str());
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
finp.close();
|
|
|
|
fout.close();
|
|
|
|
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
int main(int argc, char ** argv) {
|
|
|
|
if (argc != 4) {
|
|
|
|
fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
|
|
|
|
ggml_print_ftypes(stderr);
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
// needed to initialize f16 tables
|
|
|
|
{
|
|
|
|
struct ggml_init_params params = { 0, NULL, false };
|
|
|
|
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 ggml_ftype ftype = ggml_parse_ftype(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 (!whisper_model_quantize(fname_inp, fname_out, ggml_ftype(ftype))) {
|
|
|
|
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;
|
|
|
|
}
|