katsu560
3 years ago
committed by
GitHub
commit
fc0e984846
File diff suppressed because one or more lines are too long
File diff suppressed because it is too large
Load Diff
@ -0,0 +1,925 @@
|
||||
#include "ggml.h"
|
||||
#include "gpt-2.h"
|
||||
|
||||
#include <cmath>
|
||||
#include <cstdio>
|
||||
#include <cstring>
|
||||
#include <fstream>
|
||||
#include <map>
|
||||
#include <string>
|
||||
#include <thread>
|
||||
#include <vector>
|
||||
#include <regex>
|
||||
#include <random>
|
||||
|
||||
/////////////////////// GPT-2 BEGIN /////////////////////////
|
||||
|
||||
//
|
||||
// Vocab utils
|
||||
//
|
||||
|
||||
std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::string & text) {
|
||||
std::vector<std::string> words;
|
||||
|
||||
// first split the text into words
|
||||
{
|
||||
std::string str = text;
|
||||
std::string pat = R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)";
|
||||
|
||||
std::regex re(pat);
|
||||
std::smatch m;
|
||||
|
||||
while (std::regex_search(str, m, re)) {
|
||||
for (auto x : m) {
|
||||
words.push_back(x);
|
||||
}
|
||||
str = m.suffix();
|
||||
}
|
||||
}
|
||||
|
||||
// find the longest tokens that form the words:
|
||||
std::vector<gpt_vocab::id> tokens;
|
||||
for (const auto & word : words) {
|
||||
if (word.size() == 0) continue;
|
||||
|
||||
int i = 0;
|
||||
int n = word.size();
|
||||
while (i < n) {
|
||||
int j = n;
|
||||
while (j > i) {
|
||||
auto it = vocab.token_to_id.find(word.substr(i, j-i));
|
||||
if (it != vocab.token_to_id.end()) {
|
||||
tokens.push_back(it->second);
|
||||
i = j;
|
||||
break;
|
||||
}
|
||||
--j;
|
||||
}
|
||||
if (i == n) {
|
||||
break;
|
||||
}
|
||||
if (j == i) {
|
||||
auto sub = word.substr(i, 1);
|
||||
if (vocab.token_to_id.find(sub) != vocab.token_to_id.end()) {
|
||||
tokens.push_back(vocab.token_to_id.at(sub));
|
||||
} else {
|
||||
fprintf(stderr, "%s: unknown token '%s'\n", __func__, sub.data());
|
||||
}
|
||||
++i;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
return tokens;
|
||||
}
|
||||
|
||||
gpt_vocab::id gpt_sample_top_k_top_p(
|
||||
const gpt_vocab & vocab,
|
||||
const float * logits,
|
||||
int top_k,
|
||||
double top_p,
|
||||
double temp,
|
||||
std::mt19937 & rng) {
|
||||
int n_logits = vocab.id_to_token.size();
|
||||
|
||||
std::vector<std::pair<double, gpt_vocab::id>> logits_id;
|
||||
logits_id.reserve(n_logits);
|
||||
|
||||
for (int i = 0; i < n_logits; i++) {
|
||||
logits_id.push_back(std::make_pair(logits[i], i));
|
||||
}
|
||||
|
||||
// find the top K tokens
|
||||
std::partial_sort(
|
||||
logits_id.begin(),
|
||||
logits_id.begin() + top_k, logits_id.end(),
|
||||
[](const std::pair<double, gpt_vocab::id> & a, const std::pair<double, gpt_vocab::id> & b) {
|
||||
return a.first > b.first;
|
||||
});
|
||||
|
||||
logits_id.resize(top_k);
|
||||
|
||||
// normalize
|
||||
{
|
||||
double sum = 0.0f;
|
||||
for (int i = 0; i < (int)logits_id.size(); i++) {
|
||||
sum += logits_id[i].first;
|
||||
}
|
||||
|
||||
sum = 1.0/sum;
|
||||
for (int i = 0; i < (int)logits_id.size(); i++) {
|
||||
logits_id[i].first *= sum;
|
||||
}
|
||||
}
|
||||
|
||||
if (top_p < 1.0f) {
|
||||
{
|
||||
double cumsum = 0.0f;
|
||||
for (int i = 0; i < top_k; i++) {
|
||||
cumsum += logits_id[i].first;
|
||||
if (cumsum >= top_p) {
|
||||
logits_id.resize(i+1);
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// normalize again
|
||||
{
|
||||
double sum = 0.0f;
|
||||
for (int i = 0; i < (int)logits_id.size(); i++) {
|
||||
sum += logits_id[i].first;
|
||||
}
|
||||
|
||||
sum = 1.0/sum;
|
||||
for (int i = 0; i < (int)logits_id.size(); i++) {
|
||||
logits_id[i].first *= sum;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
//printf("\n");
|
||||
//for (int i = 0; i < (int)logits_id.size(); i++) {
|
||||
// printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), logits_id[i].first);
|
||||
//}
|
||||
//exit(0);
|
||||
|
||||
// sample from the obtained distribution
|
||||
std::vector<double> probs;
|
||||
probs.reserve(logits_id.size());
|
||||
|
||||
for (int i = 0; i < (int) logits_id.size(); i++) {
|
||||
probs.push_back(logits_id[i].first);
|
||||
}
|
||||
|
||||
std::discrete_distribution<> dist(probs.begin(), probs.end());
|
||||
int idx = dist(rng);
|
||||
|
||||
return logits_id[idx].second;
|
||||
}
|
||||
|
||||
// default hparams (GPT-2 117M)
|
||||
struct gpt2_hparams {
|
||||
int32_t n_vocab = 50257;
|
||||
int32_t n_ctx = 1024;
|
||||
int32_t n_embd = 768;
|
||||
int32_t n_head = 12;
|
||||
int32_t n_layer = 12;
|
||||
int32_t f16 = 1;
|
||||
};
|
||||
|
||||
struct gpt2_layer {
|
||||
// normalization
|
||||
struct ggml_tensor * ln_1_g;
|
||||
struct ggml_tensor * ln_1_b;
|
||||
|
||||
struct ggml_tensor * ln_2_g;
|
||||
struct ggml_tensor * ln_2_b;
|
||||
|
||||
// attention
|
||||
struct ggml_tensor * c_attn_attn_w;
|
||||
struct ggml_tensor * c_attn_attn_b;
|
||||
|
||||
struct ggml_tensor * c_attn_proj_w;
|
||||
struct ggml_tensor * c_attn_proj_b;
|
||||
|
||||
// mlp
|
||||
struct ggml_tensor * c_mlp_fc_w;
|
||||
struct ggml_tensor * c_mlp_fc_b;
|
||||
|
||||
struct ggml_tensor * c_mlp_proj_w_trans; // transposed for efficiency
|
||||
struct ggml_tensor * c_mlp_proj_b;
|
||||
};
|
||||
|
||||
struct gpt2_model {
|
||||
gpt2_hparams hparams;
|
||||
|
||||
// normalization
|
||||
struct ggml_tensor * ln_f_g;
|
||||
struct ggml_tensor * ln_f_b;
|
||||
|
||||
struct ggml_tensor * wte; // position embedding
|
||||
struct ggml_tensor * wpe; // token embedding
|
||||
|
||||
std::vector<gpt2_layer> layers;
|
||||
|
||||
// key + value memory
|
||||
struct ggml_tensor * memory_k;
|
||||
struct ggml_tensor * memory_v;
|
||||
|
||||
//
|
||||
struct ggml_context * ctx;
|
||||
std::map<std::string, struct ggml_tensor *> tensors;
|
||||
};
|
||||
|
||||
// load the model's weights from a file
|
||||
bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab & vocab) {
|
||||
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 false;
|
||||
}
|
||||
|
||||
// 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 false;
|
||||
}
|
||||
}
|
||||
|
||||
// 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.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_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);
|
||||
}
|
||||
|
||||
// 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 false;
|
||||
}
|
||||
|
||||
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_type wtype = model.hparams.f16 ? GGML_TYPE_F16 : GGML_TYPE_F32;
|
||||
|
||||
auto & ctx = model.ctx;
|
||||
|
||||
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 = hparams.n_ctx;
|
||||
const int n_vocab = hparams.n_vocab;
|
||||
|
||||
ctx_size += n_embd*ggml_type_size(GGML_TYPE_F32); // ln_f_g
|
||||
ctx_size += n_embd*ggml_type_size(GGML_TYPE_F32); // ln_f_b
|
||||
|
||||
ctx_size += n_vocab*n_embd*ggml_type_size(wtype); // wte
|
||||
ctx_size += n_ctx*n_embd*ggml_type_size(GGML_TYPE_F32); // wpe
|
||||
|
||||
ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_1_g
|
||||
ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_1_b
|
||||
|
||||
ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_2_g
|
||||
ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_2_b
|
||||
|
||||
ctx_size += n_layer*(3*n_embd*n_embd*ggml_type_size(wtype)); // c_attn_attn_w
|
||||
ctx_size += n_layer*( 3*n_embd*ggml_type_size(GGML_TYPE_F32)); // c_attn_attn_b
|
||||
|
||||
ctx_size += n_layer*(n_embd*n_embd*ggml_type_size(wtype)); // c_attn_proj_w
|
||||
ctx_size += n_layer*( n_embd*ggml_type_size(GGML_TYPE_F32)); // c_attn_proj_b
|
||||
|
||||
ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_size(wtype)); // c_mlp_fc_w
|
||||
ctx_size += n_layer*( 4*n_embd*ggml_type_size(GGML_TYPE_F32)); // c_mlp_fc_b
|
||||
|
||||
ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_size(wtype)); // c_mlp_proj_w
|
||||
ctx_size += n_layer*( n_embd*ggml_type_size(GGML_TYPE_F32)); // c_mlp_proj_b
|
||||
|
||||
ctx_size += n_ctx*n_layer*n_embd*ggml_type_size(GGML_TYPE_F32); // memory_k
|
||||
ctx_size += n_ctx*n_layer*n_embd*ggml_type_size(GGML_TYPE_F32); // 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_init_params params = {
|
||||
.mem_size = ctx_size,
|
||||
.mem_buffer = NULL,
|
||||
};
|
||||
|
||||
model.ctx = ggml_init(params);
|
||||
if (!model.ctx) {
|
||||
fprintf(stderr, "%s: ggml_init() failed\n", __func__);
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
// 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_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
||||
model.ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
||||
|
||||
model.wte = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
|
||||
model.wpe = ggml_new_tensor_2d(ctx, GGML_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_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
||||
layer.ln_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
||||
|
||||
layer.ln_2_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
||||
layer.ln_2_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
||||
|
||||
layer.c_attn_attn_w = ggml_new_tensor_2d(ctx, wtype, 3*n_embd, n_embd);
|
||||
layer.c_attn_attn_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 3*n_embd);
|
||||
|
||||
layer.c_attn_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
|
||||
layer.c_attn_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
||||
|
||||
layer.c_mlp_fc_w = ggml_new_tensor_2d(ctx, wtype, 4*n_embd, n_embd);
|
||||
layer.c_mlp_fc_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
|
||||
|
||||
layer.c_mlp_proj_w_trans = ggml_new_tensor_2d(ctx, wtype, 4*n_embd, n_embd);
|
||||
layer.c_mlp_proj_b = ggml_new_tensor_1d(ctx, GGML_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 = hparams.n_ctx;
|
||||
|
||||
const int n_mem = n_layer*n_ctx;
|
||||
const int n_elements = n_embd*n_mem;
|
||||
|
||||
model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
|
||||
model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
|
||||
|
||||
const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_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 false;
|
||||
}
|
||||
|
||||
auto tensor = model.tensors[name.data()];
|
||||
if (ggml_nelements(tensor) != nelements) {
|
||||
fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
|
||||
return false;
|
||||
}
|
||||
|
||||
if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
|
||||
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 false;
|
||||
}
|
||||
|
||||
const size_t bpe = (ftype == 0) ? sizeof(float) : sizeof(ggml_fp16_t);
|
||||
|
||||
if (nelements*bpe != ggml_nbytes(tensor)) {
|
||||
fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
|
||||
__func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
|
||||
return false;
|
||||
}
|
||||
|
||||
fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
|
||||
|
||||
//printf("%24s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ftype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0);
|
||||
total_size += ggml_nbytes(tensor);
|
||||
}
|
||||
|
||||
printf("%s: model size = %8.2f MB\n", __func__, total_size/1024.0/1024.0);
|
||||
}
|
||||
|
||||
fin.close();
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
// 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 probabilities of the next token
|
||||
//
|
||||
bool gpt2_eval(
|
||||
const gpt2_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) {
|
||||
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 = 640u*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_init_params params = {
|
||||
.mem_size = buf_size,
|
||||
.mem_buffer = buf,
|
||||
};
|
||||
|
||||
struct ggml_context * ctx0 = ggml_init(params);
|
||||
struct ggml_cgraph gf = { .n_threads = n_threads };
|
||||
|
||||
struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
|
||||
memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
|
||||
|
||||
struct ggml_tensor * position = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
|
||||
for (int i = 0; i < N; ++i) {
|
||||
((int32_t *) position->data)[i] = n_past + i;
|
||||
}
|
||||
|
||||
// wte + wpe
|
||||
struct ggml_tensor * inpL =
|
||||
ggml_add(ctx0,
|
||||
ggml_get_rows(ctx0, model.wte, embd),
|
||||
ggml_get_rows(ctx0, model.wpe, position));
|
||||
|
||||
for (int il = 0; il < n_layer; ++il) {
|
||||
struct ggml_tensor * cur;
|
||||
|
||||
// norm
|
||||
{
|
||||
// [ 768, N]
|
||||
cur = ggml_norm(ctx0, inpL);
|
||||
|
||||
// cur = ln_1_g*cur + ln_1_b
|
||||
// [ 768, N]
|
||||
cur = ggml_add(ctx0,
|
||||
ggml_mul(ctx0,
|
||||
ggml_repeat(ctx0, model.layers[il].ln_1_g, cur),
|
||||
cur),
|
||||
ggml_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_mul_mat(ctx0,
|
||||
ggml_transpose(ctx0, model.layers[il].c_attn_attn_w),
|
||||
cur);
|
||||
|
||||
cur = ggml_add(ctx0,
|
||||
ggml_repeat(ctx0, model.layers[il].c_attn_attn_b, cur),
|
||||
cur);
|
||||
}
|
||||
|
||||
// self-attention
|
||||
{
|
||||
struct ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd);
|
||||
struct ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd);
|
||||
struct ggml_tensor * Vcur = ggml_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_tensor * k = ggml_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past));
|
||||
struct ggml_tensor * v = ggml_view_1d(ctx0, model.memory_v, N*n_embd, (ggml_element_size(model.memory_v)*n_embd)*(il*n_ctx + n_past));
|
||||
|
||||
ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
|
||||
ggml_build_forward_expand(&gf, ggml_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_tensor * Q =
|
||||
ggml_permute(ctx0,
|
||||
ggml_cpy(ctx0,
|
||||
Qcur,
|
||||
ggml_new_tensor_3d(ctx0, GGML_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_tensor * K =
|
||||
ggml_permute(ctx0,
|
||||
ggml_reshape_3d(ctx0,
|
||||
ggml_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_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_tensor * V =
|
||||
// ggml_cpy(ctx0,
|
||||
// ggml_permute(ctx0,
|
||||
// ggml_reshape_3d(ctx0,
|
||||
// ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
|
||||
// n_embd/n_head, n_head, n_past + N),
|
||||
// 1, 2, 0, 3),
|
||||
// ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_past + N, n_embd/n_head, n_head));
|
||||
|
||||
//struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, true);
|
||||
|
||||
// K * Q
|
||||
// [n_past + N, N, 12]
|
||||
struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
|
||||
|
||||
// KQ_scaled = KQ / sqrt(n_embd/n_head)
|
||||
// [n_past + N, N, 12]
|
||||
struct ggml_tensor * KQ_scaled =
|
||||
ggml_scale(ctx0,
|
||||
KQ,
|
||||
ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
|
||||
);
|
||||
|
||||
// KQ_masked = mask_past(KQ_scaled)
|
||||
// [n_past + N, N, 12]
|
||||
struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
|
||||
|
||||
// KQ = soft_max(KQ_masked)
|
||||
// [n_past + N, N, 12]
|
||||
struct ggml_tensor * KQ_soft_max = ggml_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_tensor * V_trans =
|
||||
ggml_permute(ctx0,
|
||||
ggml_reshape_3d(ctx0,
|
||||
ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_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_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
|
||||
|
||||
// KQV_merged = KQV.permute(0, 2, 1, 3)
|
||||
// [64, 12, N]
|
||||
struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
|
||||
|
||||
// cur = KQV_merged.contiguous().view(n_embd, N)
|
||||
// [768, N]
|
||||
cur = ggml_cpy(ctx0,
|
||||
KQV_merged,
|
||||
ggml_new_tensor_2d(ctx0, GGML_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_mul_mat(ctx0,
|
||||
ggml_transpose(ctx0, model.layers[il].c_attn_proj_w),
|
||||
cur);
|
||||
|
||||
cur = ggml_add(ctx0,
|
||||
ggml_repeat(ctx0, model.layers[il].c_attn_proj_b, cur),
|
||||
cur);
|
||||
}
|
||||
|
||||
// add the input
|
||||
cur = ggml_add(ctx0, cur, inpL);
|
||||
|
||||
struct ggml_tensor * inpFF = cur;
|
||||
|
||||
// feed-forward network
|
||||
{
|
||||
// norm
|
||||
{
|
||||
cur = ggml_norm(ctx0, inpFF);
|
||||
|
||||
// cur = ln_2_g*cur + ln_2_b
|
||||
// [ 768, N]
|
||||
cur = ggml_add(ctx0,
|
||||
ggml_mul(ctx0,
|
||||
ggml_repeat(ctx0, model.layers[il].ln_2_g, cur),
|
||||
cur),
|
||||
ggml_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_mul_mat(ctx0,
|
||||
ggml_transpose(ctx0, model.layers[il].c_mlp_fc_w),
|
||||
cur);
|
||||
|
||||
cur = ggml_add(ctx0,
|
||||
ggml_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur),
|
||||
cur);
|
||||
|
||||
// GELU activation
|
||||
// [3072, N]
|
||||
cur = ggml_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_mul_mat(ctx0,
|
||||
model.layers[il].c_mlp_proj_w_trans,
|
||||
cur);
|
||||
|
||||
cur = ggml_add(ctx0,
|
||||
ggml_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur),
|
||||
cur);
|
||||
}
|
||||
|
||||
// input for next layer
|
||||
inpL = ggml_add(ctx0, cur, inpFF);
|
||||
}
|
||||
|
||||
// norm
|
||||
{
|
||||
// [ 768, N]
|
||||
inpL = ggml_norm(ctx0, inpL);
|
||||
|
||||
// inpL = ln_f_g*inpL + ln_f_b
|
||||
// [ 768, N]
|
||||
inpL = ggml_add(ctx0,
|
||||
ggml_mul(ctx0,
|
||||
ggml_repeat(ctx0, model.ln_f_g, inpL),
|
||||
inpL),
|
||||
ggml_repeat(ctx0, model.ln_f_b, inpL));
|
||||
}
|
||||
|
||||
// inpL = WTE * inpL
|
||||
// [ 768, 50257] - model.wte
|
||||
// [ 768, N] - inpL
|
||||
inpL = ggml_mul_mat(ctx0, model.wte, inpL);
|
||||
|
||||
// logits -> probs
|
||||
inpL = ggml_soft_max(ctx0, inpL);
|
||||
|
||||
// run the computation
|
||||
ggml_build_forward_expand(&gf, inpL);
|
||||
ggml_graph_compute (ctx0, &gf);
|
||||
|
||||
//if (n_past%100 == 0) {
|
||||
// ggml_graph_print (&gf);
|
||||
// ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
|
||||
//}
|
||||
|
||||
//embd_w.resize(n_vocab*N);
|
||||
//memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
|
||||
|
||||
// return result for just the last token
|
||||
embd_w.resize(n_vocab);
|
||||
memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
|
||||
|
||||
if (mem_per_token == 0) {
|
||||
mem_per_token = ggml_used_mem(ctx0)/N;
|
||||
}
|
||||
//printf("used_mem = %zu\n", ggml_used_mem(ctx0));
|
||||
|
||||
ggml_free(ctx0);
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
/////////////////////////////// GPT-2 END ////////////////////////////////
|
||||
|
||||
constexpr int N_THREAD = 8;
|
||||
|
||||
struct gpt2_context {
|
||||
std::string prompt_base = R"(Hello, how are you?
|
||||
I'm fine, thanks. How are you?
|
||||
Thanks, I'm fine too. What are you doing?
|
||||
I'm just sitting here.
|
||||
It's a lovely day, isn't it?
|
||||
Yes, it is.
|
||||
Did you know that I'm a robot?
|
||||
I wasn't aware of that.
|
||||
)";
|
||||
|
||||
std::mt19937 rng;
|
||||
|
||||
gpt_vocab vocab;
|
||||
gpt2_model model;
|
||||
|
||||
int32_t n_threads = std::min(N_THREAD, (int) std::thread::hardware_concurrency());
|
||||
|
||||
// sampling parameters
|
||||
int32_t top_k = 40;
|
||||
float top_p = 0.9f;
|
||||
float temp = 1.0f;
|
||||
};
|
||||
|
||||
struct gpt2_context * gpt2_init(const char * path_model) {
|
||||
gpt2_context * ctx = new gpt2_context;
|
||||
|
||||
ctx->rng = std::mt19937(time(NULL));
|
||||
|
||||
// load the model
|
||||
{
|
||||
const int64_t t_start_us = ggml_time_us();
|
||||
|
||||
if (!gpt2_model_load(path_model, ctx->model, ctx->vocab)) {
|
||||
fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, "gpt-2.bin");
|
||||
return nullptr;
|
||||
}
|
||||
|
||||
const int64_t t_load_us = ggml_time_us() - t_start_us;
|
||||
|
||||
printf("gpt-2: model loaded in %d ms\n", (int) (t_load_us/1000));
|
||||
}
|
||||
|
||||
return ctx;
|
||||
}
|
||||
|
||||
void gpt2_free(struct gpt2_context * ctx) {
|
||||
delete ctx;
|
||||
}
|
||||
|
||||
const char * gpt2_get_prompt(struct gpt2_context * ctx) {
|
||||
return ctx->prompt_base.c_str();
|
||||
}
|
||||
|
||||
void gpt2_set_prompt(struct gpt2_context * ctx, const char * prompt) {
|
||||
ctx->prompt_base = prompt;
|
||||
}
|
||||
|
||||
std::vector<gpt_vocab::id> gpt2_tokenize(const gpt2_context * ctx, const char * text) {
|
||||
return ::gpt_tokenize(ctx->vocab, text);
|
||||
}
|
||||
|
||||
std::string gpt2_gen_text(gpt2_context * ctx, const char * text, int max_tokens) {
|
||||
int n_past = 0;
|
||||
|
||||
std::vector<float> embd_w;
|
||||
|
||||
// tokenize the prompt
|
||||
std::vector<gpt_vocab::id> embd_inp = ::gpt2_tokenize(ctx, text);
|
||||
|
||||
int n_predict = std::min(max_tokens, ctx->model.hparams.n_ctx - (int) embd_inp.size());
|
||||
|
||||
std::vector<gpt_vocab::id> embd = embd_inp;
|
||||
|
||||
size_t mem_per_token = 3000000;
|
||||
|
||||
std::string result;
|
||||
|
||||
for (int i = embd.size(); i < embd_inp.size() + n_predict; i++) {
|
||||
// predict
|
||||
if (embd.size() > 0) {
|
||||
if (!gpt2_eval(ctx->model, ctx->n_threads, n_past, embd, embd_w, mem_per_token)) {
|
||||
printf("gpt-2: failed to generate text\n");
|
||||
return "";
|
||||
}
|
||||
}
|
||||
|
||||
n_past += embd.size();
|
||||
embd.clear();
|
||||
|
||||
{
|
||||
// sample next token
|
||||
const int top_k = ctx->top_k;
|
||||
const float top_p = ctx->top_p;
|
||||
const float temp = ctx->temp;
|
||||
|
||||
const int n_vocab = ctx->model.hparams.n_vocab;
|
||||
|
||||
const gpt_vocab::id id = gpt_sample_top_k_top_p(ctx->vocab, embd_w.data() + (embd_w.size() - n_vocab), top_k, top_p, temp, ctx->rng);
|
||||
|
||||
// add it to the context
|
||||
embd.push_back(id);
|
||||
}
|
||||
|
||||
result += ctx->vocab.id_to_token[embd[0]];
|
||||
|
||||
// end of text token
|
||||
if (embd.back() == 50256 ||
|
||||
ctx->vocab.id_to_token[embd.back()] == "." ||
|
||||
ctx->vocab.id_to_token[embd.back()] == "!" ||
|
||||
ctx->vocab.id_to_token[embd.back()] == "?") {
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
return result;
|
||||
}
|
||||
@ -0,0 +1,27 @@
|
||||
#pragma once
|
||||
|
||||
// TODO: Change to C-style API and move to ./examples for easy reuse.
|
||||
|
||||
#include <vector>
|
||||
#include <map>
|
||||
#include <string>
|
||||
|
||||
struct gpt_vocab {
|
||||
using id = int32_t;
|
||||
using token = std::string;
|
||||
|
||||
std::map<token, id> token_to_id;
|
||||
std::map<id, token> id_to_token;
|
||||
};
|
||||
|
||||
struct gpt2_context;
|
||||
|
||||
struct gpt2_context * gpt2_init(const char * path_model);
|
||||
void gpt2_free(struct gpt2_context * ctx);
|
||||
|
||||
const char * gpt2_get_prompt(struct gpt2_context * ctx);
|
||||
void gpt2_set_prompt(struct gpt2_context * ctx, const char * prompt);
|
||||
|
||||
std::vector<gpt_vocab::id> gpt2_tokenize(const gpt2_context * ctx, const char * text);
|
||||
|
||||
std::string gpt2_gen_text(gpt2_context * ctx, const char * text, int max_tokens);
|
||||
@ -0,0 +1,212 @@
|
||||
# Convert Hugging Face fine-tuned models to ggml format
|
||||
#
|
||||
# Usage:
|
||||
#
|
||||
# git clone https://github.com/openai/whisper
|
||||
# git clone https://github.com/ggerganov/whisper.cpp
|
||||
# git clone https://huggingface.co/openai/whisper-medium
|
||||
#
|
||||
# python3 ./whisper.cpp/models/convert-h5-to-ggml.py ./whisper-medium/ ./whisper .
|
||||
#
|
||||
# This script is similar to "convert-pt-to-ggml.py"
|
||||
#
|
||||
# For more info:
|
||||
#
|
||||
# https://github.com/ggerganov/whisper.cpp/issues/157
|
||||
#
|
||||
|
||||
import io
|
||||
import os
|
||||
import sys
|
||||
import struct
|
||||
import json
|
||||
import code
|
||||
import torch
|
||||
import numpy as np
|
||||
|
||||
from transformers import WhisperForConditionalGeneration
|
||||
|
||||
conv_map = {
|
||||
'self_attn.k_proj' : 'attn.key',
|
||||
'self_attn.q_proj' : 'attn.query',
|
||||
'self_attn.v_proj' : 'attn.value',
|
||||
'self_attn.out_proj' : 'attn.out',
|
||||
'self_attn_layer_norm' : 'attn_ln',
|
||||
'encoder_attn.q_proj' : 'cross_attn.query',
|
||||
'encoder_attn.v_proj' : 'cross_attn.value',
|
||||
'encoder_attn.out_proj' : 'cross_attn.out',
|
||||
'encoder_attn_layer_norm' : 'cross_attn_ln',
|
||||
'fc1' : 'mlp.0',
|
||||
'fc2' : 'mlp.2',
|
||||
'final_layer_norm' : 'mlp_ln',
|
||||
'encoder.layer_norm.bias' : 'encoder.ln_post.bias',
|
||||
'encoder.layer_norm.weight' : 'encoder.ln_post.weight',
|
||||
'encoder.embed_positions.weight': 'encoder.positional_embedding',
|
||||
'decoder.layer_norm.bias' : 'decoder.ln.bias',
|
||||
'decoder.layer_norm.weight' : 'decoder.ln.weight',
|
||||
'decoder.embed_positions.weight': 'decoder.positional_embedding',
|
||||
'decoder.embed_tokens.weight' : 'decoder.token_embedding.weight',
|
||||
'proj_out.weight' : 'decoder.proj.weight',
|
||||
}
|
||||
|
||||
# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
|
||||
def bytes_to_unicode():
|
||||
"""
|
||||
Returns list of utf-8 byte and a corresponding list of unicode strings.
|
||||
The reversible bpe codes work on unicode strings.
|
||||
This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
|
||||
When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
|
||||
This is a signficant percentage of your normal, say, 32K bpe vocab.
|
||||
To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
|
||||
And avoids mapping to whitespace/control characters the bpe code barfs on.
|
||||
"""
|
||||
bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
|
||||
cs = bs[:]
|
||||
n = 0
|
||||
for b in range(2**8):
|
||||
if b not in bs:
|
||||
bs.append(b)
|
||||
cs.append(2**8+n)
|
||||
n += 1
|
||||
cs = [chr(n) for n in cs]
|
||||
return dict(zip(bs, cs))
|
||||
|
||||
if len(sys.argv) < 4:
|
||||
print("Usage: convert-h5-to-ggml.py dir_model path-to-whisper-repo dir-output [use-f32]\n")
|
||||
sys.exit(1)
|
||||
|
||||
dir_model = sys.argv[1]
|
||||
dir_whisper = sys.argv[2]
|
||||
dir_out = sys.argv[3]
|
||||
|
||||
with open(dir_model + "/vocab.json", "r") as f:
|
||||
encoder = json.load(f)
|
||||
with open(dir_model + "/added_tokens.json", "r") as f:
|
||||
encoder_added = json.load(f)
|
||||
with open(dir_model + "/config.json", "r") as f:
|
||||
hparams = json.load(f)
|
||||
|
||||
model = WhisperForConditionalGeneration.from_pretrained(dir_model)
|
||||
|
||||
#code.interact(local=locals())
|
||||
|
||||
n_mels = hparams["num_mel_bins"]
|
||||
with np.load(os.path.join(dir_whisper, "whisper/assets", "mel_filters.npz")) as f:
|
||||
filters = torch.from_numpy(f[f"mel_{n_mels}"])
|
||||
|
||||
dir_tokenizer = dir_model
|
||||
|
||||
fname_out = dir_out + "/ggml-model.bin"
|
||||
|
||||
with open(dir_tokenizer + "/vocab.json", "r", encoding="utf8") as f:
|
||||
tokens = json.load(f)
|
||||
|
||||
# use 16-bit or 32-bit floats
|
||||
use_f16 = True
|
||||
if len(sys.argv) > 4:
|
||||
use_f16 = False
|
||||
fname_out = dir_out + "/ggml-model-f32.bin"
|
||||
|
||||
fout = open(fname_out, "wb")
|
||||
|
||||
fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
|
||||
fout.write(struct.pack("i", hparams["vocab_size"]))
|
||||
fout.write(struct.pack("i", hparams["max_source_positions"]))
|
||||
fout.write(struct.pack("i", hparams["d_model"]))
|
||||
fout.write(struct.pack("i", hparams["encoder_attention_heads"]))
|
||||
fout.write(struct.pack("i", hparams["encoder_layers"]))
|
||||
fout.write(struct.pack("i", hparams["max_length"]))
|
||||
fout.write(struct.pack("i", hparams["d_model"]))
|
||||
fout.write(struct.pack("i", hparams["decoder_attention_heads"]))
|
||||
fout.write(struct.pack("i", hparams["decoder_layers"]))
|
||||
fout.write(struct.pack("i", hparams["num_mel_bins"]))
|
||||
fout.write(struct.pack("i", use_f16))
|
||||
|
||||
fout.write(struct.pack("i", filters.shape[0]))
|
||||
fout.write(struct.pack("i", filters.shape[1]))
|
||||
for i in range(filters.shape[0]):
|
||||
for j in range(filters.shape[1]):
|
||||
fout.write(struct.pack("f", filters[i][j]))
|
||||
|
||||
byte_encoder = bytes_to_unicode()
|
||||
byte_decoder = {v:k for k, v in byte_encoder.items()}
|
||||
|
||||
fout.write(struct.pack("i", len(tokens)))
|
||||
|
||||
tokens = sorted(tokens.items(), key=lambda x: x[1])
|
||||
for key in tokens:
|
||||
text = bytearray([byte_decoder[c] for c in key[0]])
|
||||
fout.write(struct.pack("i", len(text)))
|
||||
fout.write(text)
|
||||
|
||||
list_vars = model.state_dict()
|
||||
for name in list_vars.keys():
|
||||
# this seems to not be used
|
||||
# ref: https://github.com/huggingface/transformers/blob/9a5b84a0076a04fe9596da72e8668069d4f09ea0/src/transformers/models/whisper/modeling_whisper.py#L1099-L1106
|
||||
if name == "proj_out.weight":
|
||||
print('Skipping', name)
|
||||
continue
|
||||
|
||||
src = name
|
||||
|
||||
nn = name
|
||||
if name != "proj_out.weight":
|
||||
nn = nn.split(".")[1:]
|
||||
else:
|
||||
nn = nn.split(".")
|
||||
|
||||
if nn[1] == "layers":
|
||||
nn[1] = "blocks"
|
||||
if ".".join(nn[3:-1]) == "encoder_attn.k_proj":
|
||||
mapped = "attn.key" if nn[0] == "encoder" else "cross_attn.key"
|
||||
else:
|
||||
mapped = conv_map[".".join(nn[3:-1])]
|
||||
name = ".".join(nn[:3] + [mapped] + nn[-1:])
|
||||
else:
|
||||
name = ".".join(nn)
|
||||
name = conv_map[name] if name in conv_map else name
|
||||
|
||||
print(src, ' -> ', name)
|
||||
data = list_vars[src].squeeze().numpy()
|
||||
data = data.astype(np.float16)
|
||||
|
||||
# reshape conv bias from [n] to [n, 1]
|
||||
if name == "encoder.conv1.bias" or \
|
||||
name == "encoder.conv2.bias":
|
||||
data = data.reshape(data.shape[0], 1)
|
||||
print(" Reshaped variable: " + name + " to shape: ", data.shape)
|
||||
|
||||
n_dims = len(data.shape)
|
||||
print(name, n_dims, data.shape)
|
||||
|
||||
# looks like the whisper models are in f16 by default
|
||||
# so we need to convert the small tensors to f32 until we fully support f16 in ggml
|
||||
# ftype == 0 -> float32, ftype == 1 -> float16
|
||||
ftype = 1;
|
||||
if use_f16:
|
||||
if n_dims < 2 or \
|
||||
name == "encoder.conv1.bias" or \
|
||||
name == "encoder.conv2.bias" or \
|
||||
name == "encoder.positional_embedding" or \
|
||||
name == "decoder.positional_embedding":
|
||||
print(" Converting to float32")
|
||||
data = data.astype(np.float32)
|
||||
ftype = 0
|
||||
else:
|
||||
data = data.astype(np.float32)
|
||||
ftype = 0
|
||||
|
||||
# header
|
||||
str = name.encode('utf-8')
|
||||
fout.write(struct.pack("iii", n_dims, len(str), ftype))
|
||||
for i in range(n_dims):
|
||||
fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
|
||||
fout.write(str);
|
||||
|
||||
# data
|
||||
data.tofile(fout)
|
||||
|
||||
fout.close()
|
||||
|
||||
print("Done. Output file: " + fname_out)
|
||||
print("")
|
||||
Loading…
Reference in new issue