Merge bb0600c7da into 1808ee0500
Thomas Antony
1 year ago
committed by
GitHub
commit
0d93538e5b
@ -0,0 +1,943 @@
|
||||
#include "llama.h"
|
||||
#include "ggml.h"
|
||||
|
||||
#include <cassert>
|
||||
#include <cmath>
|
||||
#include <cstdio>
|
||||
#include <cstring>
|
||||
#include <fstream>
|
||||
#include <map>
|
||||
#include <string>
|
||||
#include <vector>
|
||||
#include <regex>
|
||||
|
||||
// TODO: move somewhere else
|
||||
#define QK 32
|
||||
|
||||
|
||||
// determine number of model parts based on the dimension
|
||||
static const std::map<int, int> LLAMA_N_PARTS = {
|
||||
{ 4096, 1 },
|
||||
{ 5120, 2 },
|
||||
{ 6656, 4 },
|
||||
{ 8192, 8 },
|
||||
};
|
||||
|
||||
// load the model's weights from a file
|
||||
bool llama_model_load(const std::string & fname, llama_model & model, gpt_vocab & vocab, int n_ctx) {
|
||||
printf("%s: loading model from '%s' - please wait ...\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;
|
||||
}
|
||||
}
|
||||
|
||||
int n_ff = 0;
|
||||
int n_parts = 0;
|
||||
|
||||
// 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_mult, sizeof(hparams.n_mult));
|
||||
fin.read((char *) &hparams.n_head, sizeof(hparams.n_head));
|
||||
fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
|
||||
fin.read((char *) &hparams.n_rot, sizeof(hparams.n_rot));
|
||||
fin.read((char *) &hparams.f16, sizeof(hparams.f16));
|
||||
|
||||
hparams.n_ctx = n_ctx;
|
||||
|
||||
n_ff = ((2*(4*hparams.n_embd)/3 + hparams.n_mult - 1)/hparams.n_mult)*hparams.n_mult;
|
||||
n_parts = LLAMA_N_PARTS.at(hparams.n_embd);
|
||||
|
||||
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_mult = %d\n", __func__, hparams.n_mult);
|
||||
printf("%s: n_head = %d\n", __func__, hparams.n_head);
|
||||
printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
|
||||
printf("%s: n_rot = %d\n", __func__, hparams.n_rot);
|
||||
printf("%s: f16 = %d\n", __func__, hparams.f16);
|
||||
printf("%s: n_ff = %d\n", __func__, n_ff);
|
||||
printf("%s: n_parts = %d\n", __func__, n_parts);
|
||||
}
|
||||
|
||||
// load vocab
|
||||
{
|
||||
const int32_t n_vocab = model.hparams.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;
|
||||
|
||||
//if (i < 30000) {
|
||||
// printf("%s: vocab[%d] = '%s'\n", __func__, i, word.c_str());
|
||||
//}
|
||||
}
|
||||
}
|
||||
|
||||
// for the big tensors, we have the option to store the data in 16-bit floats or quantized
|
||||
// in order to save memory and also to speed up the computation
|
||||
ggml_type wtype = GGML_TYPE_COUNT;
|
||||
switch (model.hparams.f16) {
|
||||
case 0: wtype = GGML_TYPE_F32; break;
|
||||
case 1: wtype = GGML_TYPE_F16; break;
|
||||
case 2: wtype = GGML_TYPE_Q4_0; break;
|
||||
case 3: wtype = GGML_TYPE_Q4_1; break;
|
||||
default:
|
||||
{
|
||||
fprintf(stderr, "%s: invalid model file '%s' (bad f16 value %d)\n",
|
||||
__func__, fname.c_str(), model.hparams.f16);
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
const ggml_type wtype2 = 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*n_vocab*ggml_type_sizef(wtype); // tok_embeddings
|
||||
|
||||
ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // norm
|
||||
|
||||
ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // output
|
||||
|
||||
ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // attention_norm
|
||||
|
||||
ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wq
|
||||
ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wk
|
||||
ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wv
|
||||
ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wo
|
||||
|
||||
ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ffn_norm
|
||||
|
||||
ctx_size += n_layer*(n_ff*n_embd*ggml_type_sizef(wtype)); // w1
|
||||
ctx_size += n_layer*(n_ff*n_embd*ggml_type_sizef(wtype)); // w2
|
||||
ctx_size += n_layer*(n_ff*n_embd*ggml_type_sizef(wtype)); // w3
|
||||
|
||||
ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
|
||||
ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v
|
||||
|
||||
ctx_size += (5 + 10*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.tok_embeddings = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
|
||||
|
||||
model.norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
||||
model.output = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
|
||||
|
||||
// map by name
|
||||
model.tensors["tok_embeddings.weight"] = model.tok_embeddings;
|
||||
|
||||
model.tensors["norm.weight"] = model.norm;
|
||||
model.tensors["output.weight"] = model.output;
|
||||
|
||||
for (int i = 0; i < n_layer; ++i) {
|
||||
auto & layer = model.layers[i];
|
||||
|
||||
layer.attention_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
||||
|
||||
layer.wq = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
|
||||
layer.wk = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
|
||||
layer.wv = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
|
||||
layer.wo = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
|
||||
|
||||
layer.ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
||||
|
||||
layer.w1 = ggml_new_tensor_2d(ctx, wtype, n_embd, n_ff);
|
||||
layer.w2 = ggml_new_tensor_2d(ctx, wtype, n_ff, n_embd);
|
||||
layer.w3 = ggml_new_tensor_2d(ctx, wtype, n_embd, n_ff);
|
||||
|
||||
// map by name
|
||||
model.tensors["layers." + std::to_string(i) + ".attention_norm.weight"] = layer.attention_norm;
|
||||
|
||||
model.tensors["layers." + std::to_string(i) + ".attention.wq.weight"] = layer.wq;
|
||||
model.tensors["layers." + std::to_string(i) + ".attention.wk.weight"] = layer.wk;
|
||||
model.tensors["layers." + std::to_string(i) + ".attention.wv.weight"] = layer.wv;
|
||||
model.tensors["layers." + std::to_string(i) + ".attention.wo.weight"] = layer.wo;
|
||||
|
||||
model.tensors["layers." + std::to_string(i) + ".ffn_norm.weight"] = layer.ffn_norm;
|
||||
|
||||
model.tensors["layers." + std::to_string(i) + ".feed_forward.w1.weight"] = layer.w1;
|
||||
model.tensors["layers." + std::to_string(i) + ".feed_forward.w2.weight"] = layer.w2;
|
||||
model.tensors["layers." + std::to_string(i) + ".feed_forward.w3.weight"] = layer.w3;
|
||||
}
|
||||
}
|
||||
|
||||
// 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);
|
||||
}
|
||||
|
||||
const size_t file_offset = fin.tellg();
|
||||
|
||||
fin.close();
|
||||
|
||||
std::vector<uint8_t> tmp;
|
||||
|
||||
for (int i = 0; i < n_parts; ++i) {
|
||||
const int part_id = i;
|
||||
//const int part_id = n_parts - i - 1;
|
||||
|
||||
std::string fname_part = fname;
|
||||
if (i > 0) {
|
||||
fname_part += "." + std::to_string(i);
|
||||
}
|
||||
|
||||
printf("%s: loading model part %d/%d from '%s'\n", __func__, i+1, n_parts, fname_part.c_str());
|
||||
|
||||
fin = std::ifstream(fname_part, std::ios::binary);
|
||||
fin.seekg(file_offset);
|
||||
|
||||
// load weights
|
||||
{
|
||||
int n_tensors = 0;
|
||||
size_t total_size = 0;
|
||||
|
||||
printf("%s: ", __func__);
|
||||
|
||||
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;
|
||||
}
|
||||
|
||||
// split_type = 0: split by columns
|
||||
// split_type = 1: split by rows
|
||||
int split_type = 0;
|
||||
|
||||
// split_type = 0:
|
||||
// regex:
|
||||
// - tok_embeddings.*
|
||||
// - layers.*.attention.wo.weight
|
||||
// - layers.*.feed_forward.w2.weight
|
||||
|
||||
// split_type = 1:
|
||||
// regex:
|
||||
// - output.*
|
||||
// - layers.*.attention.wq.weight
|
||||
// - layers.*.attention.wk.weight
|
||||
// - layers.*.attention.wv.weight
|
||||
// - layers.*.feed_forward.w1.weight
|
||||
// - layers.*.feed_forward.w3.weight
|
||||
if (name.find("tok_embeddings") != std::string::npos) {
|
||||
split_type = 0;
|
||||
} else if (name.find("layers") != std::string::npos) {
|
||||
if (name.find("attention.wo.weight") != std::string::npos) {
|
||||
split_type = 0;
|
||||
} else if (name.find("feed_forward.w2.weight") != std::string::npos) {
|
||||
split_type = 0;
|
||||
} else {
|
||||
split_type = 1;
|
||||
}
|
||||
} else if (name.find("output") != std::string::npos) {
|
||||
split_type = 1;
|
||||
}
|
||||
|
||||
auto tensor = model.tensors[name.data()];
|
||||
|
||||
if (n_dims == 1) {
|
||||
if (ggml_nelements(tensor) != nelements) {
|
||||
fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
|
||||
return false;
|
||||
}
|
||||
} else {
|
||||
if (ggml_nelements(tensor)/n_parts != nelements) {
|
||||
fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
if (n_dims == 1) {
|
||||
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;
|
||||
}
|
||||
} else {
|
||||
if (split_type == 0) {
|
||||
if (tensor->ne[0]/n_parts != 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]/n_parts, tensor->ne[1], ne[0], ne[1]);
|
||||
return false;
|
||||
}
|
||||
} else {
|
||||
if (tensor->ne[0] != ne[0] || tensor->ne[1]/n_parts != 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]/n_parts, ne[0], ne[1]);
|
||||
return false;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if (0) {
|
||||
static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", };
|
||||
printf("%24s - [%5d, %5d], type = %6s, split = %d\n", name.data(), ne[0], ne[1], ftype_str[ftype], split_type);
|
||||
}
|
||||
|
||||
size_t bpe = 0;
|
||||
|
||||
switch (ftype) {
|
||||
case 0: bpe = ggml_type_size(GGML_TYPE_F32); break;
|
||||
case 1: bpe = ggml_type_size(GGML_TYPE_F16); break;
|
||||
case 2: bpe = ggml_type_size(GGML_TYPE_Q4_0); assert(ne[0] % 64 == 0); break;
|
||||
case 3: bpe = ggml_type_size(GGML_TYPE_Q4_1); assert(ne[0] % 64 == 0); break;
|
||||
default:
|
||||
{
|
||||
fprintf(stderr, "%s: unknown ftype %d in model file\n", __func__, ftype);
|
||||
return false;
|
||||
}
|
||||
};
|
||||
|
||||
if (n_dims == 1 || n_parts == 1) {
|
||||
if ((nelements*bpe)/ggml_blck_size(tensor->type) != 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;
|
||||
}
|
||||
|
||||
if (part_id == 0) {
|
||||
fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
|
||||
} else {
|
||||
fin.seekg(ggml_nbytes(tensor), std::ios::cur);
|
||||
}
|
||||
|
||||
total_size += ggml_nbytes(tensor);
|
||||
} else {
|
||||
if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)/n_parts) {
|
||||
fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
|
||||
__func__, name.data(), ggml_nbytes(tensor)/n_parts, nelements*bpe);
|
||||
return false;
|
||||
}
|
||||
|
||||
if (split_type == 0) {
|
||||
const int np0 = ne[0];
|
||||
|
||||
const size_t row_size = (tensor->ne[0]/ggml_blck_size(tensor->type))*ggml_type_size(tensor->type);
|
||||
assert(row_size == tensor->nb[1]);
|
||||
|
||||
for (int i1 = 0; i1 < ne[1]; ++i1) {
|
||||
const size_t offset_row = i1*row_size;
|
||||
const size_t offset = offset_row + ((part_id*np0)/ggml_blck_size(tensor->type))*ggml_type_size(tensor->type);
|
||||
fin.read(reinterpret_cast<char *>(tensor->data) + offset, row_size/n_parts);
|
||||
}
|
||||
} else {
|
||||
const int np1 = ne[1];
|
||||
|
||||
const size_t row_size = (tensor->ne[0]/ggml_blck_size(tensor->type))*ggml_type_size(tensor->type);
|
||||
|
||||
for (int i1 = 0; i1 < ne[1]; ++i1) {
|
||||
const size_t offset_row = (i1 + part_id*np1)*row_size;
|
||||
fin.read(reinterpret_cast<char *>(tensor->data) + offset_row, row_size);
|
||||
}
|
||||
}
|
||||
|
||||
total_size += ggml_nbytes(tensor)/n_parts;
|
||||
}
|
||||
|
||||
//printf("%42s - [%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);
|
||||
if (++n_tensors % 8 == 0) {
|
||||
printf(".");
|
||||
fflush(stdout);
|
||||
}
|
||||
}
|
||||
|
||||
printf(" done\n");
|
||||
|
||||
printf("%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, n_tensors);
|
||||
}
|
||||
|
||||
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 logits for the next token
|
||||
//
|
||||
// The GPT-J model requires about 16MB of memory per input token.
|
||||
//
|
||||
bool llama_eval(
|
||||
const llama_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;
|
||||
const int n_rot = hparams.n_embd/hparams.n_head;
|
||||
|
||||
const int d_key = n_embd/n_head;
|
||||
|
||||
static size_t buf_size = 512u*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);
|
||||
ggml_cgraph gf = {};
|
||||
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 * inpL = ggml_get_rows(ctx0, model.tok_embeddings, embd);
|
||||
|
||||
for (int il = 0; il < n_layer; ++il) {
|
||||
struct ggml_tensor * inpSA = inpL;
|
||||
|
||||
struct ggml_tensor * cur;
|
||||
|
||||
// norm
|
||||
{
|
||||
cur = ggml_norm(ctx0, inpL);
|
||||
|
||||
// cur = attention_norm*cur
|
||||
cur = ggml_mul(ctx0,
|
||||
ggml_repeat(ctx0, model.layers[il].attention_norm, cur),
|
||||
cur);
|
||||
}
|
||||
|
||||
// self-attention
|
||||
{
|
||||
struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
|
||||
struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
|
||||
struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
|
||||
|
||||
// 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)
|
||||
struct ggml_tensor * Q =
|
||||
ggml_permute(ctx0,
|
||||
ggml_rope(ctx0,
|
||||
ggml_cpy(ctx0,
|
||||
Qcur,
|
||||
ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)),
|
||||
n_past, n_rot, 0),
|
||||
0, 2, 1, 3);
|
||||
|
||||
// K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
|
||||
struct ggml_tensor * K =
|
||||
ggml_permute(ctx0,
|
||||
ggml_rope(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),
|
||||
n_past, n_rot, 1),
|
||||
0, 2, 1, 3);
|
||||
|
||||
// K * Q
|
||||
struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
|
||||
|
||||
// KQ_scaled = KQ / sqrt(n_embd/n_head)
|
||||
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)
|
||||
struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
|
||||
|
||||
// KQ = soft_max(KQ_masked)
|
||||
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()
|
||||
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
|
||||
struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
|
||||
|
||||
// KQV_merged = KQV.permute(0, 2, 1, 3)
|
||||
struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
|
||||
|
||||
// cur = KQV_merged.contiguous().view(n_embd, N)
|
||||
cur = ggml_cpy(ctx0,
|
||||
KQV_merged,
|
||||
ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
|
||||
|
||||
// projection (no bias)
|
||||
cur = ggml_mul_mat(ctx0,
|
||||
model.layers[il].wo,
|
||||
cur);
|
||||
}
|
||||
|
||||
struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpSA);
|
||||
|
||||
// feed-forward network
|
||||
{
|
||||
// norm
|
||||
{
|
||||
cur = ggml_norm(ctx0, inpFF);
|
||||
|
||||
// cur = ffn_norm*cur
|
||||
cur = ggml_mul(ctx0,
|
||||
ggml_repeat(ctx0, model.layers[il].ffn_norm, cur),
|
||||
cur);
|
||||
}
|
||||
|
||||
struct ggml_tensor * tmp = ggml_mul_mat(ctx0,
|
||||
model.layers[il].w3,
|
||||
cur);
|
||||
|
||||
|
||||
cur = ggml_mul_mat(ctx0,
|
||||
model.layers[il].w1,
|
||||
cur);
|
||||
|
||||
// SILU activation
|
||||
cur = ggml_silu(ctx0, cur);
|
||||
|
||||
cur = ggml_mul(ctx0, cur, tmp);
|
||||
|
||||
cur = ggml_mul_mat(ctx0,
|
||||
model.layers[il].w2,
|
||||
cur);
|
||||
}
|
||||
|
||||
cur = ggml_add(ctx0, cur, inpFF);
|
||||
|
||||
// input for next layer
|
||||
inpL = cur;
|
||||
}
|
||||
|
||||
// norm
|
||||
{
|
||||
inpL = ggml_norm(ctx0, inpL);
|
||||
|
||||
// inpL = norm*inpL
|
||||
inpL = ggml_mul(ctx0,
|
||||
ggml_repeat(ctx0, model.norm, inpL),
|
||||
inpL);
|
||||
}
|
||||
|
||||
// lm_head
|
||||
{
|
||||
inpL = ggml_mul_mat(ctx0, model.output, 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;
|
||||
}
|
||||
bool llama_model_quantize(const std::string & fname_inp, const std::string & fname_out, int itype) {
|
||||
ggml_type type = GGML_TYPE_Q4_1;
|
||||
|
||||
switch (itype) {
|
||||
case 2: type = GGML_TYPE_Q4_0; break;
|
||||
case 3: type = GGML_TYPE_Q4_1; break;
|
||||
default: fprintf(stderr, "%s: invalid quantization type %d\n", __func__, itype); return 1;
|
||||
};
|
||||
|
||||
if (type != GGML_TYPE_Q4_0 && type != GGML_TYPE_Q4_1) {
|
||||
fprintf(stderr, "%s: invalid quantization type %d\n", __func__, type);
|
||||
return false;
|
||||
}
|
||||
|
||||
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));
|
||||
if (magic != 0x67676d6c) {
|
||||
fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname_inp.c_str());
|
||||
return false;
|
||||
}
|
||||
|
||||
fout.write((char *) &magic, sizeof(magic));
|
||||
}
|
||||
|
||||
llama_hparams hparams;
|
||||
|
||||
// load hparams
|
||||
{
|
||||
finp.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
|
||||
//finp.read((char *) &hparams.n_ctx, sizeof(hparams.n_ctx));
|
||||
finp.read((char *) &hparams.n_embd, sizeof(hparams.n_embd));
|
||||
finp.read((char *) &hparams.n_mult, sizeof(hparams.n_mult));
|
||||
finp.read((char *) &hparams.n_head, sizeof(hparams.n_head));
|
||||
finp.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
|
||||
finp.read((char *) &hparams.n_rot, sizeof(hparams.n_rot));
|
||||
finp.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_mult = %d\n", __func__, hparams.n_mult);
|
||||
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);
|
||||
|
||||
fout.write((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
|
||||
//fout.write((char *) &hparams.n_ctx, sizeof(hparams.n_ctx));
|
||||
fout.write((char *) &hparams.n_embd, sizeof(hparams.n_embd));
|
||||
fout.write((char *) &hparams.n_mult, sizeof(hparams.n_mult));
|
||||
fout.write((char *) &hparams.n_head, sizeof(hparams.n_head));
|
||||
fout.write((char *) &hparams.n_layer, sizeof(hparams.n_layer));
|
||||
fout.write((char *) &hparams.n_rot, sizeof(hparams.n_rot));
|
||||
fout.write((char *) &itype, sizeof(hparams.f16));
|
||||
}
|
||||
|
||||
// load vocab
|
||||
{
|
||||
const int32_t n_vocab = hparams.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;
|
||||
}
|
||||
|
||||
std::string word;
|
||||
for (int i = 0; i < n_vocab; i++) {
|
||||
uint32_t len;
|
||||
finp.read ((char *) &len, sizeof(len));
|
||||
fout.write((char *) &len, sizeof(len));
|
||||
|
||||
word.resize(len);
|
||||
finp.read ((char *) word.data(), len);
|
||||
fout.write((char *) word.data(), len);
|
||||
|
||||
vocab.token_to_id[word] = i;
|
||||
vocab.id_to_token[i] = word;
|
||||
}
|
||||
}
|
||||
|
||||
// load weights
|
||||
{
|
||||
size_t total_size_org = 0;
|
||||
size_t total_size_new = 0;
|
||||
|
||||
std::vector<float> work;
|
||||
|
||||
std::vector<uint8_t> data_u8;
|
||||
std::vector<ggml_fp16_t> data_f16;
|
||||
std::vector<float> data_f32;
|
||||
|
||||
std::vector<int64_t> hist_all(1 << 4, 0);
|
||||
|
||||
while (true) {
|
||||
int32_t n_dims;
|
||||
int32_t length;
|
||||
int32_t ftype;
|
||||
|
||||
finp.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
|
||||
finp.read(reinterpret_cast<char *>(&length), sizeof(length));
|
||||
finp.read(reinterpret_cast<char *>(&ftype), sizeof(ftype));
|
||||
|
||||
if (finp.eof()) {
|
||||
break;
|
||||
}
|
||||
|
||||
int32_t nelements = 1;
|
||||
int32_t ne[2] = { 1, 1 };
|
||||
for (int i = 0; i < n_dims; ++i) {
|
||||
finp.read (reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
|
||||
nelements *= ne[i];
|
||||
}
|
||||
|
||||
std::string name(length, 0);
|
||||
finp.read (&name[0], length);
|
||||
|
||||
{
|
||||
static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", };
|
||||
printf("%48s - [%5d, %5d], type = %6s ", name.data(), ne[0], ne[1], ftype_str[ftype]);
|
||||
}
|
||||
|
||||
// regexes of tensor names to be quantized
|
||||
const std::vector<std::string> k_names = {
|
||||
".*weight",
|
||||
};
|
||||
|
||||
bool quantize = false;
|
||||
for (const auto & s : k_names) {
|
||||
if (std::regex_match(name, std::regex(s))) {
|
||||
quantize = true;
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
// quantize only 2D tensors
|
||||
quantize &= (n_dims == 2);
|
||||
|
||||
if (quantize) {
|
||||
if (ftype != 0 && ftype != 1) {
|
||||
fprintf(stderr, "%s: unsupported ftype %d for integer quantization\n", __func__, ftype);
|
||||
return false;
|
||||
}
|
||||
|
||||
if (ftype == 1) {
|
||||
data_f16.resize(nelements);
|
||||
finp.read(reinterpret_cast<char *>(data_f16.data()), nelements * sizeof(ggml_fp16_t));
|
||||
data_f32.resize(nelements);
|
||||
for (int i = 0; i < nelements; ++i) {
|
||||
data_f32[i] = ggml_fp16_to_fp32(data_f16[i]);
|
||||
}
|
||||
} else {
|
||||
data_f32.resize(nelements);
|
||||
finp.read(reinterpret_cast<char *>(data_f32.data()), nelements * sizeof(float));
|
||||
}
|
||||
|
||||
ftype = itype;
|
||||
} else {
|
||||
const int bpe = (ftype == 0) ? sizeof(float) : sizeof(uint16_t);
|
||||
|
||||
data_u8.resize(nelements*bpe);
|
||||
finp.read(reinterpret_cast<char *>(data_u8.data()), nelements * bpe);
|
||||
}
|
||||
|
||||
fout.write(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
|
||||
fout.write(reinterpret_cast<char *>(&length), sizeof(length));
|
||||
fout.write(reinterpret_cast<char *>(&ftype), sizeof(ftype));
|
||||
for (int i = 0; i < n_dims; ++i) {
|
||||
fout.write(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
|
||||
}
|
||||
fout.write(&name[0], length);
|
||||
|
||||
if (quantize) {
|
||||
printf("quantizing .. ");
|
||||
work.resize(nelements); // for quantization
|
||||
|
||||
size_t cur_size = 0;
|
||||
std::vector<int64_t> hist_cur(1 << 4, 0);
|
||||
|
||||
switch (type) {
|
||||
case GGML_TYPE_Q4_0:
|
||||
{
|
||||
cur_size = ggml_quantize_q4_0(data_f32.data(), work.data(), nelements, ne[0], QK, hist_cur.data());
|
||||
} break;
|
||||
case GGML_TYPE_Q4_1:
|
||||
{
|
||||
cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], QK, hist_cur.data());
|
||||
} break;
|
||||
default:
|
||||
{
|
||||
fprintf(stderr, "%s: unsupported quantization type %d\n", __func__, type);
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
fout.write(reinterpret_cast<char *>(work.data()), cur_size);
|
||||
total_size_new += cur_size;
|
||||
|
||||
printf("size = %8.2f MB -> %8.2f MB | hist: ", nelements * sizeof(float)/1024.0/1024.0, cur_size/1024.0/1024.0);
|
||||
for (int i = 0; i < hist_cur.size(); ++i) {
|
||||
hist_all[i] += hist_cur[i];
|
||||
}
|
||||
|
||||
for (int i = 0; i < hist_cur.size(); ++i) {
|
||||
printf("%5.3f ", hist_cur[i] / (float)nelements);
|
||||
}
|
||||
printf("\n");
|
||||
} else {
|
||||
printf("size = %8.3f MB\n", data_u8.size()/1024.0/1024.0);
|
||||
fout.write(reinterpret_cast<char *>(data_u8.data()), data_u8.size());
|
||||
total_size_new += data_u8.size();
|
||||
}
|
||||
|
||||
total_size_org += nelements * sizeof(float);
|
||||
}
|
||||
|
||||
printf("%s: model size = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
|
||||
printf("%s: quant size = %8.2f MB\n", __func__, total_size_new/1024.0/1024.0);
|
||||
|
||||
{
|
||||
int64_t sum_all = 0;
|
||||
for (int i = 0; i < hist_all.size(); ++i) {
|
||||
sum_all += hist_all[i];
|
||||
}
|
||||
|
||||
printf("%s: hist: ", __func__);
|
||||
for (int i = 0; i < hist_all.size(); ++i) {
|
||||
printf("%5.3f ", hist_all[i] / (float)sum_all);
|
||||
}
|
||||
printf("\n");
|
||||
}
|
||||
}
|
||||
|
||||
finp.close();
|
||||
fout.close();
|
||||
|
||||
return true;
|
||||
}
|
||||
@ -0,0 +1,67 @@
|
||||
#pragma once
|
||||
|
||||
#include <string>
|
||||
#include <vector>
|
||||
#include <map>
|
||||
|
||||
#include "utils.h"
|
||||
|
||||
// default hparams (LLaMA 7B)
|
||||
struct llama_hparams {
|
||||
int32_t n_vocab = 32000;
|
||||
int32_t n_ctx = 512; // this is provided as user input?
|
||||
int32_t n_embd = 4096;
|
||||
int32_t n_mult = 256;
|
||||
int32_t n_head = 32;
|
||||
int32_t n_layer = 32;
|
||||
int32_t n_rot = 64;
|
||||
int32_t f16 = 1;
|
||||
};
|
||||
|
||||
struct llama_layer {
|
||||
// normalization
|
||||
struct ggml_tensor * attention_norm;
|
||||
|
||||
// attention
|
||||
struct ggml_tensor * wq;
|
||||
struct ggml_tensor * wk;
|
||||
struct ggml_tensor * wv;
|
||||
struct ggml_tensor * wo;
|
||||
|
||||
// normalization
|
||||
struct ggml_tensor * ffn_norm;
|
||||
|
||||
// ff
|
||||
struct ggml_tensor * w1;
|
||||
struct ggml_tensor * w2;
|
||||
struct ggml_tensor * w3;
|
||||
};
|
||||
|
||||
struct llama_model {
|
||||
llama_hparams hparams;
|
||||
|
||||
struct ggml_tensor * tok_embeddings;
|
||||
|
||||
struct ggml_tensor * norm;
|
||||
struct ggml_tensor * output;
|
||||
|
||||
std::vector<llama_layer> layers;
|
||||
|
||||
// key + value memory
|
||||
struct ggml_tensor * memory_k;
|
||||
struct ggml_tensor * memory_v;
|
||||
|
||||
//
|
||||
struct ggml_context * ctx = nullptr;
|
||||
std::map<std::string, struct ggml_tensor *> tensors;
|
||||
};
|
||||
|
||||
bool llama_model_load(const std::string & fname, llama_model & model, gpt_vocab & vocab, int n_ctx);
|
||||
bool llama_eval(
|
||||
const llama_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);
|
||||
bool llama_model_quantize(const std::string & fname_inp, const std::string & fname_out, int itype);
|
||||
Loading…
Reference in new issue