halfway refactoring, wip adding other model types

This commit is contained in:
Concedo 2023-04-01 01:13:05 +08:00
parent 56949197fe
commit 6b86f5ea22
19 changed files with 13135 additions and 319 deletions

View file

@ -211,7 +211,7 @@ endif
BLAS_BUILD =
ifeq ($(OS),Windows_NT)
BLAS_BUILD = $(CXX) $(CXXFLAGS) expose.cpp ggml_blas.o common.o extra.o libopenblas.lib -shared -o llamacpp_blas.dll $(LDFLAGS)
BLAS_BUILD = $(CXX) $(CXXFLAGS) ggml_blas.o common.o extra.o expose.o llama_adapter.o libopenblas.lib -shared -o llamacpp_blas.dll $(LDFLAGS)
else
BLAS_BUILD = @echo 'Your OS is $(OS) and does not appear to be Windows. If you want to use openblas, please link it manually with LLAMA_OPENBLAS=1'
endif
@ -243,6 +243,9 @@ ggml.o: ggml.c ggml.h
ggml_blas.o: ggml.c ggml.h
$(CC) $(CFLAGS) -DGGML_USE_OPENBLAS -c ggml.c -o ggml_blas.o
ggml_old_v1.o: otherarch/ggml_old.c otherarch/ggml_old.h
$(CC) $(CFLAGS) -c otherarch/ggml_old.c -o ggml_old_v1.o
llama.o: llama.cpp llama.h
$(CXX) $(CXXFLAGS) -c llama.cpp -o llama.o
@ -252,8 +255,14 @@ common.o: examples/common.cpp examples/common.h
extra.o: extra.cpp extra.h
$(CXX) $(CXXFLAGS) -c extra.cpp -o extra.o
expose.o: expose.cpp expose.h
$(CXX) $(CXXFLAGS) -c expose.cpp -o expose.o
llama_adapter.o: llama_adapter.cpp llama_adapter.h
$(CXX) $(CXXFLAGS) -c llama_adapter.cpp -o llama_adapter.o
clean:
rm -vf *.o main quantize perplexity embedding main.exe quantize.exe llamacpp.dll llamacpp_blas.dll
rm -vf *.o main quantize perplexity embedding main.exe quantize.exe llamacpp.dll llamacpp_blas.dll gpt2.exe gptj.exe
main: examples/main/main.cpp ggml.o llama.o common.o
$(CXX) $(CXXFLAGS) examples/main/main.cpp ggml.o llama.o common.o -o main $(LDFLAGS)
@ -261,10 +270,14 @@ main: examples/main/main.cpp ggml.o llama.o common.o
@echo '==== Run ./main -h for help. ===='
@echo
llamalib: expose.cpp ggml.o common.o extra.o
$(CXX) $(CXXFLAGS) expose.cpp ggml.o common.o extra.o -shared -o llamacpp.dll $(LDFLAGS)
gptj: ggml_old_v1.o
$(CXX) $(CXXFLAGS) otherarch/gptj_old.cpp otherarch/utils.cpp ggml_old_v1.o -o gptj $(LDFLAGS)
llamalib_blas: expose.cpp ggml_blas.o common.o extra.o
llamalib: ggml.o common.o extra.o expose.o llama_adapter.o
$(CXX) $(CXXFLAGS) ggml.o common.o extra.o expose.o llama_adapter.o -shared -o llamacpp.dll $(LDFLAGS)
llamalib_blas: ggml_blas.o common.o extra.o expose.o llama_adapter.o
$(BLAS_BUILD)
quantize: examples/quantize/quantize.cpp ggml.o llama.o

View file

@ -7,323 +7,27 @@
//No dynamic memory allocation! Setup structs with FIXED (known) shapes and sizes for ALL output fields
//Python will ALWAYS provide the memory, we just write to it.
#include <time.h>
#include "./examples/main/main.cpp"
#include "llama_adapter.h"
#include "expose.h"
#include "extra.h"
#include "ggml.h"
clock_t bench_timer = 0;
void timer_start()
extern "C"
{
bench_timer = clock();
}
double timer_check()
{
double ticks = clock() - bench_timer;
double time_taken = ((double)ticks)/CLOCKS_PER_SEC;
return time_taken;
}
void print_tok_vec(std::vector<llama_token> & embd)
{
std::cout << "[";
bool first = true;
for (auto i: embd) {
if(!first)
{
std::cout << ',';
}
first = false;
std::cout << i;
}
std::cout << "]";
}
extern "C" {
struct load_model_inputs
{
const int threads;
const int max_context_length;
const int batch_size;
const bool f16_kv;
const char * model_filename;
const int n_parts_overwrite = -1;
};
struct generation_inputs
{
const int seed;
const char * prompt;
const int max_context_length;
const int max_length;
const float temperature;
const int top_k;
const float top_p;
const float rep_pen;
const int rep_pen_range;
};
struct generation_outputs
{
int status = -1;
char text[16384]; //16kb should be enough for any response
};
//return val: 0=fail, 1=(original ggml, alpaca), 2=(ggmf), 3=(ggjt)
int file_format = 0;
llama_context_params ctx_params;
gpt_params params;
int n_past = 0;
int n_threads = 4;
int n_batch = 8;
std::string model;
llama_context * ctx;
std::vector<llama_token> last_n_tokens;
std::vector<llama_token> current_context_tokens;
//return val: 0=fail, 1=(original ggml, alpaca), 2=(ggmf), 3=(ggjt)
static FileFormat file_format = FAIL;
bool load_model(const load_model_inputs inputs)
{
printf("System Info: %s\n", llama_print_system_info());
ctx_params = llama_context_default_params();
n_threads = inputs.threads;
n_batch = inputs.batch_size;
model = inputs.model_filename;
ctx_params.n_ctx = inputs.max_context_length;
ctx_params.n_parts = inputs.n_parts_overwrite;
ctx_params.seed = -1;
ctx_params.f16_kv = inputs.f16_kv;
ctx_params.logits_all = false;
std::string model = inputs.model_filename;
file_format = check_file_format(model.c_str());
printf("\nFile format detected: (ver %d)\n",file_format);
if(file_format==1 || file_format==2)
{
ctx = legacy_llama_init_from_file(model.c_str(), ctx_params);
}
else
{
ctx = llama_init_from_file(model.c_str(), ctx_params);
}
if (ctx == NULL) {
fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, model.c_str());
return false;
}
if(file_format<3)
{
printf("\n---\nWarning: Your model has an INVALID or OUTDATED format (ver %d). Please reconvert it for better results!\n---\n",file_format);
}
//determine mem per token
const std::vector<llama_token> tmp = { 0, 1, 2, 3 };
llama_eval(ctx, tmp.data(), tmp.size(), 0, params.n_threads);
return true;
printf("\nIdentified as LLAMA model: (ver %d)\n", file_format);
return llama_load_model(inputs, file_format);
}
generation_outputs generate(const generation_inputs inputs, generation_outputs & output)
generation_outputs generate(const generation_inputs inputs, generation_outputs &output)
{
params.prompt = inputs.prompt;
params.seed = inputs.seed;
params.n_predict = inputs.max_length;
params.top_k = inputs.top_k;
params.top_p = inputs.top_p;
params.temp = inputs.temperature;
params.repeat_last_n = inputs.rep_pen_range;
params.repeat_penalty = inputs.rep_pen;
params.n_ctx = inputs.max_context_length;
params.n_batch = n_batch;
params.n_threads = n_threads;
if(params.repeat_last_n<1)
{
params.repeat_last_n = 1;
}
if(params.top_k<1)
{
params.top_k = 300; //to disable top_k we actually need to increase this value to a very high number
}
if (params.seed <= 0)
{
params.seed = time(NULL);
}
params.prompt.insert(0, 1, ' ');
// tokenize the prompt
std::vector<llama_token> embd_inp;
if(file_format==1)
{
embd_inp = ::legacy_llama_tokenize(ctx, params.prompt, true);
}else{
embd_inp = ::llama_tokenize(ctx, params.prompt, true);
}
//params.n_predict = std::min(params.n_predict, params.n_ctx - (int) embd_inp.size());
//truncate to front of the prompt if its too long
if (embd_inp.size() + params.n_predict > params.n_ctx) {
int offset = embd_inp.size() - params.n_ctx + params.n_predict;
embd_inp = std::vector<llama_token>(embd_inp.begin() + offset, embd_inp.end());
}
//determine how much npast we have to rewind from the current state
std::vector<llama_token> embd;
int last_n_size = params.repeat_last_n;
last_n_tokens.resize(last_n_size);
//display usage
// std::string tst = " ";
// char * tst2 = (char*)tst.c_str();
// gpt_print_usage(1,&tst2,params);
std::fill(last_n_tokens.begin(), last_n_tokens.end(), 0);
n_past = 0;
//fast forward the past based on identical tokens, stop once a divergence is noted
int embd_inp_len = embd_inp.size();
for(int i=0;i<current_context_tokens.size();++i)
{
if(current_context_tokens[i]==embd_inp[i])
{
n_past += 1;
last_n_tokens.push_back(current_context_tokens[i]);
}
else
{
break;
}
if((i+2)>=embd_inp_len)
{
break;
}
}
last_n_tokens.erase(last_n_tokens.begin(),last_n_tokens.begin()+n_past);
embd_inp.erase(embd_inp.begin(),embd_inp.begin()+n_past);
//if using BLAS and prompt is big enough, switch to single thread and use a huge batch
bool blasmode = (embd_inp.size() >= 32 && ggml_cpu_has_blas());
int original_batch = params.n_batch;
int original_threads = params.n_threads;
if(blasmode)
{
params.n_batch = 512;
params.n_threads = 1;
}
current_context_tokens.resize(n_past);
int remaining_tokens = params.n_predict;
int input_consumed = 0;
std::mt19937 rng(params.seed);
std::string concat_output = "";
bool startedsampling = false;
timer_start();
double time1=0,time2=0;
unsigned int embd_inp_size = embd_inp.size();
printf("\n");
while (remaining_tokens > 0)
{
llama_token id = 0;
// predict
unsigned int embdsize = embd.size();
if (embdsize > 0)
{
//print progress
if(!startedsampling)
{
printf("\rProcessing Prompt%s (%d / %d tokens)",(blasmode?" [BLAS]":""), input_consumed,embd_inp_size);
}
else
{
printf("\rGenerating (%d / %d tokens)",(1+params.n_predict-remaining_tokens),params.n_predict);
}
//printf("\nnp:%d embd:%d txt:%s",n_past,embd.size(),llama_token_to_str(ctx, embd[0]));
if (llama_eval(ctx, embd.data(), embdsize, n_past, params.n_threads))
{
fprintf(stderr, "Failed to predict\n");
snprintf(output.text, sizeof(output.text), "%s", "");
output.status = 0;
return output;
}
}
n_past += embd.size();
embd.clear();
if ((int) embd_inp.size() <= input_consumed)
{
// out of user input, sample next token
const float top_k = params.top_k;
const float top_p = params.top_p;
const float temp = params.temp;
const float repeat_penalty = params.repeat_penalty;
if(!startedsampling)
{
startedsampling = true;
params.n_batch = original_batch;
params.n_threads = original_threads;
time1 = timer_check();
timer_start();
printf("\n");
}
{
auto logits = llama_get_logits(ctx);
// set the logit of the eos token (2) to zero to avoid sampling it
logits[llama_token_eos()] = 0;
//set logits of opening square bracket to zero.
logits[518] = 0;
logits[29961] = 0;
id = llama_sample_top_p_top_k(ctx, last_n_tokens.data(), last_n_tokens.size(), top_k, top_p, temp, repeat_penalty);
last_n_tokens.erase(last_n_tokens.begin());
last_n_tokens.push_back(id);
current_context_tokens.push_back(id);
}
// add it to the context
embd.push_back(id);
// decrement remaining sampling budget
--remaining_tokens;
//printf("\nid:%d word:%s\n",id,llama_token_to_str(ctx, id));
concat_output += llama_token_to_str(ctx, id);
}
else
{
// some user input remains from prompt or interaction, forward it to processing
while ((int) embd_inp.size() > input_consumed)
{
embd.push_back(embd_inp[input_consumed]);
last_n_tokens.erase(last_n_tokens.begin());
last_n_tokens.push_back(embd_inp[input_consumed]);
current_context_tokens.push_back(embd_inp[input_consumed]);
++input_consumed;
if ((int) embd.size() >= params.n_batch)
{
break;
}
}
}
}
time2 = timer_check();
printf("\nTime Taken - Processing:%.1fs, Generation:%.1fs, Total:%.1fs",time1,time2,(time1+time2));
output.status = 1;
snprintf(output.text, sizeof(output.text), "%s", concat_output.c_str());
return output;
return llama_generate(inputs, output);
}
}

28
expose.h Normal file
View file

@ -0,0 +1,28 @@
#pragma once
struct load_model_inputs
{
const int threads;
const int max_context_length;
const int batch_size;
const bool f16_kv;
const char *model_filename;
const int n_parts_overwrite = -1;
};
struct generation_inputs
{
const int seed;
const char *prompt;
const int max_context_length;
const int max_length;
const float temperature;
const int top_k;
const float top_p;
const float rep_pen;
const int rep_pen_range;
};
struct generation_outputs
{
int status = -1;
char text[16384]; //16kb should be enough for any response
};

View file

@ -19,7 +19,7 @@
#endif
//return val: 0=fail, 1=(original ggml, alpaca), 2=(ggmf), 3=(ggjt)
int check_file_format(const std::string & fname)
FileFormat check_file_format(const std::string & fname)
{
std::vector<char> f_buf(1024*1024);
@ -27,22 +27,22 @@
fin.rdbuf()->pubsetbuf(f_buf.data(), f_buf.size());
if (!fin) {
fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
return false;
return FileFormat::FAIL;
}
int fileformat = 0;
FileFormat fileformat = FileFormat::FAIL;
uint32_t magic;
fin.read((char *) &magic, sizeof(magic));
if (magic == 0x67676d6c) { //v1 format ggml, alpaca
fileformat = 1;
fileformat = FileFormat::GGML;
}
else if(magic == 0x67676d66) //v2 format ggmf
{
fileformat = 2;
fileformat = FileFormat::GGHF;
}
else if(magic == 0x67676a74) //v3 format ggjt
{
fileformat = 3; //ggjt by default
fileformat = FileFormat::GGJT; //ggjt by default
}
fin.close();

11
extra.h
View file

@ -1,3 +1,4 @@
#pragma once
#include "common.h"
#include <cassert>
@ -14,8 +15,16 @@
#include "llama.h"
#include "ggml.h"
//return val: 0=fail, 1=(original ggml, alpaca), 2=(ggmf), 3=(ggjt)
enum FileFormat
{
FAIL=0,
GGML=1,
GGHF=2,
GGJT=3
};
int check_file_format(const std::string & fname);
FileFormat check_file_format(const std::string & fname);
std::vector<llama_token> legacy_llama_tokenize(struct llama_context * ctx, const std::string & text, bool add_bos);
static bool legacy_llama_model_load(const std::string & fname, llama_context & lctx, int n_ctx, int n_parts, ggml_type memory_type, bool vocab_only, llama_progress_callback progress_callback, void *progress_callback_user_data);

301
llama_adapter.cpp Normal file
View file

@ -0,0 +1,301 @@
//This is Concedo's shitty adapter for adding python bindings for llama
//Considerations:
//Don't want to use pybind11 due to dependencies on MSVCC
//ZERO or MINIMAL changes as possible to main.cpp - do not move their function declarations here!
//Leave main.cpp UNTOUCHED, We want to be able to update the repo and pull any changes automatically.
//No dynamic memory allocation! Setup structs with FIXED (known) shapes and sizes for ALL output fields
//Python will ALWAYS provide the memory, we just write to it.
#include <time.h>
#include "./examples/main/main.cpp"
#include "extra.h"
#include "ggml.h"
#include "llama_adapter.h"
//return val: 0=fail, 1=(original ggml, alpaca), 2=(ggmf), 3=(ggjt)
static int file_format = 0;
static llama_context_params ctx_params;
static gpt_params params;
static int n_past = 0;
static int n_threads = 4;
static int n_batch = 8;
static std::string model;
static llama_context *ctx;
static std::vector<llama_token> last_n_tokens;
static std::vector<llama_token> current_context_tokens;
static clock_t bench_timer = 0;
void timer_start()
{
bench_timer = clock();
}
double timer_check()
{
double ticks = clock() - bench_timer;
double time_taken = ((double)ticks) / CLOCKS_PER_SEC;
return time_taken;
}
void print_tok_vec(std::vector<llama_token> &embd)
{
std::cout << "[";
bool first = true;
for (auto i : embd)
{
if (!first)
{
std::cout << ',';
}
first = false;
std::cout << i;
}
std::cout << "]";
}
bool llama_load_model(const load_model_inputs inputs, FileFormat file_format)
{
printf("System Info: %s\n", llama_print_system_info());
ctx_params = llama_context_default_params();
n_threads = inputs.threads;
n_batch = inputs.batch_size;
model = inputs.model_filename;
ctx_params.n_ctx = inputs.max_context_length;
ctx_params.n_parts = inputs.n_parts_overwrite;
ctx_params.seed = -1;
ctx_params.f16_kv = inputs.f16_kv;
ctx_params.logits_all = false;
if (file_format == FileFormat::GGML || file_format == FileFormat::GGHF)
{
ctx = legacy_llama_init_from_file(model.c_str(), ctx_params);
}
else
{
ctx = llama_init_from_file(model.c_str(), ctx_params);
}
if (ctx == NULL)
{
fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, model.c_str());
return false;
}
if (file_format < FileFormat::GGJT)
{
printf("\n---\nWarning: Your model has an INVALID or OUTDATED format (ver %d). Please reconvert it for better results!\n---\n", file_format);
}
//determine mem per token
const std::vector<llama_token> tmp = {0, 1, 2, 3};
llama_eval(ctx, tmp.data(), tmp.size(), 0, params.n_threads);
return true;
}
generation_outputs llama_generate(const generation_inputs inputs, generation_outputs &output)
{
params.prompt = inputs.prompt;
params.seed = inputs.seed;
params.n_predict = inputs.max_length;
params.top_k = inputs.top_k;
params.top_p = inputs.top_p;
params.temp = inputs.temperature;
params.repeat_last_n = inputs.rep_pen_range;
params.repeat_penalty = inputs.rep_pen;
params.n_ctx = inputs.max_context_length;
params.n_batch = n_batch;
params.n_threads = n_threads;
if (params.repeat_last_n < 1)
{
params.repeat_last_n = 1;
}
if (params.top_k < 1)
{
params.top_k = 300; //to disable top_k we actually need to increase this value to a very high number
}
if (params.seed <= 0)
{
params.seed = time(NULL);
}
params.prompt.insert(0, 1, ' ');
// tokenize the prompt
std::vector<llama_token> embd_inp;
if (file_format == 1)
{
embd_inp = ::legacy_llama_tokenize(ctx, params.prompt, true);
}
else
{
embd_inp = ::llama_tokenize(ctx, params.prompt, true);
}
//params.n_predict = std::min(params.n_predict, params.n_ctx - (int) embd_inp.size());
//truncate to front of the prompt if its too long
if (embd_inp.size() + params.n_predict > params.n_ctx)
{
int offset = embd_inp.size() - params.n_ctx + params.n_predict;
embd_inp = std::vector<llama_token>(embd_inp.begin() + offset, embd_inp.end());
}
//determine how much npast we have to rewind from the current state
std::vector<llama_token> embd;
int last_n_size = params.repeat_last_n;
last_n_tokens.resize(last_n_size);
//display usage
// std::string tst = " ";
// char * tst2 = (char*)tst.c_str();
// gpt_print_usage(1,&tst2,params);
std::fill(last_n_tokens.begin(), last_n_tokens.end(), 0);
n_past = 0;
//fast forward the past based on identical tokens, stop once a divergence is noted
int embd_inp_len = embd_inp.size();
for (int i = 0; i < current_context_tokens.size(); ++i)
{
if (current_context_tokens[i] == embd_inp[i])
{
n_past += 1;
last_n_tokens.push_back(current_context_tokens[i]);
}
else
{
break;
}
if ((i + 2) >= embd_inp_len)
{
break;
}
}
last_n_tokens.erase(last_n_tokens.begin(), last_n_tokens.begin() + n_past);
embd_inp.erase(embd_inp.begin(), embd_inp.begin() + n_past);
//if using BLAS and prompt is big enough, switch to single thread and use a huge batch
bool blasmode = (embd_inp.size() >= 32 && ggml_cpu_has_blas());
int original_batch = params.n_batch;
int original_threads = params.n_threads;
if (blasmode)
{
params.n_batch = 512;
params.n_threads = 1;
}
current_context_tokens.resize(n_past);
int remaining_tokens = params.n_predict;
int input_consumed = 0;
std::mt19937 rng(params.seed);
std::string concat_output = "";
bool startedsampling = false;
timer_start();
double time1 = 0, time2 = 0;
unsigned int embd_inp_size = embd_inp.size();
printf("\n");
while (remaining_tokens > 0)
{
llama_token id = 0;
// predict
unsigned int embdsize = embd.size();
if (embdsize > 0)
{
//print progress
if (!startedsampling)
{
printf("\rProcessing Prompt%s (%d / %d tokens)", (blasmode ? " [BLAS]" : ""), input_consumed, embd_inp_size);
}
else
{
printf("\rGenerating (%d / %d tokens)", (1 + params.n_predict - remaining_tokens), params.n_predict);
}
//printf("\nnp:%d embd:%d txt:%s",n_past,embd.size(),llama_token_to_str(ctx, embd[0]));
if (llama_eval(ctx, embd.data(), embdsize, n_past, params.n_threads))
{
fprintf(stderr, "Failed to predict\n");
snprintf(output.text, sizeof(output.text), "%s", "");
output.status = 0;
return output;
}
}
n_past += embd.size();
embd.clear();
if ((int)embd_inp.size() <= input_consumed)
{
// out of user input, sample next token
const float top_k = params.top_k;
const float top_p = params.top_p;
const float temp = params.temp;
const float repeat_penalty = params.repeat_penalty;
if (!startedsampling)
{
startedsampling = true;
params.n_batch = original_batch;
params.n_threads = original_threads;
time1 = timer_check();
timer_start();
printf("\n");
}
{
auto logits = llama_get_logits(ctx);
// set the logit of the eos token (2) to zero to avoid sampling it
logits[llama_token_eos()] = 0;
//set logits of opening square bracket to zero.
logits[518] = 0;
logits[29961] = 0;
id = llama_sample_top_p_top_k(ctx, last_n_tokens.data(), last_n_tokens.size(), top_k, top_p, temp, repeat_penalty);
last_n_tokens.erase(last_n_tokens.begin());
last_n_tokens.push_back(id);
current_context_tokens.push_back(id);
}
// add it to the context
embd.push_back(id);
// decrement remaining sampling budget
--remaining_tokens;
//printf("\nid:%d word:%s\n",id,llama_token_to_str(ctx, id));
concat_output += llama_token_to_str(ctx, id);
}
else
{
// some user input remains from prompt or interaction, forward it to processing
while ((int)embd_inp.size() > input_consumed)
{
embd.push_back(embd_inp[input_consumed]);
last_n_tokens.erase(last_n_tokens.begin());
last_n_tokens.push_back(embd_inp[input_consumed]);
current_context_tokens.push_back(embd_inp[input_consumed]);
++input_consumed;
if ((int)embd.size() >= params.n_batch)
{
break;
}
}
}
}
time2 = timer_check();
printf("\nTime Taken - Processing:%.1fs, Generation:%.1fs, Total:%.1fs", time1, time2, (time1 + time2));
output.status = 1;
snprintf(output.text, sizeof(output.text), "%s", concat_output.c_str());
return output;
}

6
llama_adapter.h Normal file
View file

@ -0,0 +1,6 @@
#pragma once
#include "expose.h"
#include "extra.h"
bool llama_load_model(const load_model_inputs inputs, FileFormat file_format);
generation_outputs llama_generate(const generation_inputs inputs, generation_outputs &output);

Binary file not shown.

Binary file not shown.

8622
otherarch/ggml_old.c Normal file

File diff suppressed because it is too large Load diff

748
otherarch/ggml_old.h Normal file
View file

@ -0,0 +1,748 @@
#pragma once
//
// GGML Tensor Library
//
// This documentation is still a work in progress.
// If you wish some specific topics to be covered, feel free to drop a comment:
//
// https://github.com/ggerganov/whisper.cpp/issues/40
//
// ## Overview
//
// This library implements:
//
// - a set of tensor operations
// - automatic differentiation
// - basic optimization algorithms
//
// The aim of this library is to provide a minimalistic approach for various machine learning tasks. This includes,
// but is not limited to, the following:
//
// - linear regression
// - support vector machines
// - neural networks
//
// The library allows the user to define a certain function using the available tensor operations. This function
// definition is represented internally via a computation graph. Each tensor operation in the function definition
// corresponds to a node in the graph. Having the computation graph defined, the user can choose to compute the
// function's value and/or its gradient with respect to the input variables. Optionally, the function can be optimized
// using one of the available optimization algorithms.
//
// For example, here we define the function: f(x) = a*x^2 + b
//
// {
// struct ggml_init_params params = {
// .mem_size = 16*1024*1024,
// .mem_buffer = NULL,
// };
//
// // memory allocation happens here
// struct ggml_context * ctx = ggml_init(params);
//
// struct ggml_tensor * x = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
//
// ggml_set_param(ctx, x); // x is an input variable
//
// struct ggml_tensor * a = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
// struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
// struct ggml_tensor * x2 = ggml_mul(ctx, x, x);
// struct ggml_tensor * f = ggml_add(ctx, ggml_mul(ctx, a, x2), b);
//
// ...
// }
//
// Notice that the function definition above does not involve any actual computation. The computation is performed only
// when the user explicitly requests it. For example, to compute the function's value at x = 2.0:
//
// {
// ...
//
// struct ggml_cgraph gf = ggml_build_forward(f);
//
// // set the input variable and parameter values
// ggml_set_f32(x, 2.0f);
// ggml_set_f32(a, 3.0f);
// ggml_set_f32(b, 4.0f);
//
// ggml_graph_compute(ctx0, &gf);
//
// printf("f = %f\n", ggml_get_f32_1d(f, 0));
//
// ...
// }
//
// The actual computation is performed in the ggml_graph_compute() function.
//
// The ggml_new_tensor_...() functions create new tensors. They are allocated in the memory buffer provided to the
// ggml_init() function. You have to be careful not to exceed the memory buffer size. Therefore, you have to know
// in advance how much memory you need for your computation. Alternatively, you can allocate a large enough memory
// and after defining the computation graph, call the ggml_used_mem() function to find out how much memory was
// actually needed.
//
// The ggml_set_param() function marks a tensor as an input variable. This is used by the automatic
// differentiation and optimization algorithms.
//
// The described approach allows to define the function graph once and then compute its forward or backward graphs
// multiple times. All computations will use the same memory buffer allocated in the ggml_init() function. This way
// the user can avoid the memory allocation overhead at runtime.
//
// The library supports multi-dimensional tensors - up to 4 dimensions. The FP16 and FP32 data types are first class
// citizens, but in theory the library can be extended to support FP8 and integer data types.
//
// Each tensor operation produces a new tensor. Initially the library was envisioned to support only the use of unary
// and binary operations. Most of the available operations fall into one of these two categories. With time, it became
// clear that the library needs to support more complex operations. The way to support these operations is not clear
// yet, but a few examples are demonstrated in the following operations:
//
// - ggml_permute()
// - ggml_conv_1d_1s()
// - ggml_conv_1d_2s()
//
// For each tensor operator, the library implements a forward and backward computation function. The forward function
// computes the output tensor value given the input tensor values. The backward function computes the adjoint of the
// input tensors given the adjoint of the output tensor. For a detailed explanation of what this means, take a
// calculus class, or watch the following video:
//
// What is Automatic Differentiation?
// https://www.youtube.com/watch?v=wG_nF1awSSY
//
//
// ## Tensor data (struct ggml_tensor)
//
// The tensors are stored in memory via the ggml_tensor struct. The structure provides information about the size of
// the tensor, the data type, and the memory buffer where the tensor data is stored. Additionally, it contains
// pointers to the "source" tensors - i.e. the tensors that were used to compute the current tensor. For example:
//
// {
// struct ggml_tensor * c = ggml_add(ctx, a, b);
//
// assert(c->src[0] == a);
// assert(c->src[1] == b);
// }
//
// The multi-dimensional tensors are stored in row-major order. The ggml_tensor struct contains fields for the
// number of elements in each dimension ("ne") as well as the number of bytes ("nb", a.k.a. stride). This allows
// to store tensors that are not contiguous in memory, which is useful for operations such as transposition and
// permutation. All tensor operations have to take the stride into account and not assume that the tensor is
// contiguous in memory.
//
// The data of the tensor is accessed via the "data" pointer. For example:
//
// {
// struct ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 2, 3);
//
// // a[1, 2] = 1.0f;
// *(float *) ((char *) a->data + 2*a->nb[1] + 1*a->nb[0]) = 1.0f;
//
// // a[2, 0] = 2.0f;
// *(float *) ((char *) a->data + 0*a->nb[1] + 2*a->nb[0]) = 2.0f;
//
// ...
// }
//
// Alternatively, there are helper functions, such as ggml_get_f32_1d() and ggml_set_f32_1d() that can be used.
//
// ## The matrix multiplication operator (ggml_mul_mat)
//
// TODO
//
//
// ## Multi-threading
//
// TODO
//
//
// ## Overview of ggml.c
//
// TODO
//
//
// ## SIMD optimizations
//
// TODO
//
//
// ## Debugging ggml
//
// TODO
//
//
#ifdef __cplusplus
extern "C" {
#endif
#include <stdint.h>
#include <stddef.h>
#include <stdbool.h>
#define GGML_MAX_DIMS 4
#define GGML_MAX_NODES 4096
#define GGML_MAX_PARAMS 16
#define GGML_MAX_CONTEXTS 64
#define GGML_MAX_OPT 4
#ifdef __ARM_NEON
// we use the built-in 16-bit float type
typedef __fp16 ggml_fp16_t;
#else
typedef uint16_t ggml_fp16_t;
#endif
// convert FP16 <-> FP32
float ggml_fp16_to_fp32(ggml_fp16_t x);
ggml_fp16_t ggml_fp32_to_fp16(float x);
struct ggml_object;
struct ggml_context;
enum ggml_type {
GGML_TYPE_I8,
GGML_TYPE_I16,
GGML_TYPE_I32,
GGML_TYPE_F16,
GGML_TYPE_F32,
GGML_TYPE_COUNT,
};
// available tensor operations:
enum ggml_op {
GGML_OP_NONE = 0,
GGML_OP_DUP,
GGML_OP_ADD,
GGML_OP_SUB,
GGML_OP_MUL,
GGML_OP_DIV,
GGML_OP_SQR,
GGML_OP_SQRT,
GGML_OP_SUM,
GGML_OP_MEAN,
GGML_OP_REPEAT,
GGML_OP_ABS,
GGML_OP_SGN,
GGML_OP_NEG,
GGML_OP_STEP,
GGML_OP_RELU,
GGML_OP_GELU,
GGML_OP_NORM, // normalize
GGML_OP_MUL_MAT,
GGML_OP_SCALE,
GGML_OP_CPY,
GGML_OP_RESHAPE,
GGML_OP_VIEW,
GGML_OP_PERMUTE,
GGML_OP_TRANSPOSE,
GGML_OP_GET_ROWS,
GGML_OP_DIAG_MASK_INF,
GGML_OP_SOFT_MAX,
GGML_OP_ROPE,
GGML_OP_CONV_1D_1S,
GGML_OP_CONV_1D_2S,
GGML_OP_FLASH_ATTN,
GGML_OP_FLASH_FF,
GGML_OP_COUNT,
};
// n-dimensional tensor
struct ggml_tensor {
enum ggml_type type;
int n_dims;
int ne[GGML_MAX_DIMS]; // number of elements
size_t nb[GGML_MAX_DIMS]; // stride in bytes:
// nb[0] = sizeof(type)
// nb[1] = nb[0] * ne[0] + padding
// nb[i] = nb[i-1] * ne[i-1]
// compute data
enum ggml_op op;
bool is_param;
struct ggml_tensor * grad;
struct ggml_tensor * src0;
struct ggml_tensor * src1;
struct ggml_tensor * opt[GGML_MAX_OPT];
// thread scheduling
int n_tasks;
// performance
int perf_runs;
int64_t perf_cycles;
int64_t perf_time_us;
void * data;
char padding[8];
};
// computation graph
struct ggml_cgraph {
int n_nodes;
int n_leafs;
int n_threads;
size_t work_size;
struct ggml_tensor * work;
struct ggml_tensor * nodes[GGML_MAX_NODES];
struct ggml_tensor * grads[GGML_MAX_NODES];
struct ggml_tensor * leafs[GGML_MAX_NODES];
// performance
int perf_runs;
int64_t perf_cycles;
int64_t perf_time_us;
};
// scratch buffer
struct ggml_scratch {
size_t offs;
size_t size;
void * data;
};
struct ggml_init_params {
// memory pool
size_t mem_size; // bytes
void * mem_buffer; // if NULL, memory will be allocated internally
};
void ggml_time_init(void); // call this once at the beginning of the program
int64_t ggml_time_ms(void);
int64_t ggml_time_us(void);
int64_t ggml_cycles(void);
int64_t ggml_cycles_per_ms(void);
void ggml_print_object (const struct ggml_object * obj);
void ggml_print_objects(const struct ggml_context * ctx);
int ggml_nelements(const struct ggml_tensor * tensor);
size_t ggml_nbytes (const struct ggml_tensor * tensor);
size_t ggml_type_size (enum ggml_type type);
size_t ggml_element_size(const struct ggml_tensor * tensor);
struct ggml_context * ggml_init(struct ggml_init_params params);
void ggml_free(struct ggml_context * ctx);
size_t ggml_used_mem(const struct ggml_context * ctx);
size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch);
struct ggml_tensor * ggml_new_tensor(
struct ggml_context * ctx,
enum ggml_type type,
int n_dims,
const int *ne);
struct ggml_tensor * ggml_new_tensor_1d(
struct ggml_context * ctx,
enum ggml_type type,
int ne0);
struct ggml_tensor * ggml_new_tensor_2d(
struct ggml_context * ctx,
enum ggml_type type,
int ne0,
int ne1);
struct ggml_tensor * ggml_new_tensor_3d(
struct ggml_context * ctx,
enum ggml_type type,
int ne0,
int ne1,
int ne2);
struct ggml_tensor * ggml_new_tensor_4d(
struct ggml_context * ctx,
enum ggml_type type,
int ne0,
int ne1,
int ne2,
int ne3);
struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
struct ggml_tensor * ggml_dup_tensor (struct ggml_context * ctx, const struct ggml_tensor * src);
struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, const struct ggml_tensor * src);
struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i);
void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value);
float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);
void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);
void * ggml_get_data (const struct ggml_tensor * tensor);
float * ggml_get_data_f32(const struct ggml_tensor * tensor);
//
// operations on tensors with backpropagation
//
struct ggml_tensor * ggml_dup(
struct ggml_context * ctx,
struct ggml_tensor * a);
struct ggml_tensor * ggml_add(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
struct ggml_tensor * ggml_sub(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
struct ggml_tensor * ggml_mul(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
struct ggml_tensor * ggml_div(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
struct ggml_tensor * ggml_sqr(
struct ggml_context * ctx,
struct ggml_tensor * a);
struct ggml_tensor * ggml_sqrt(
struct ggml_context * ctx,
struct ggml_tensor * a);
// return scalar
// TODO: compute sum along rows
struct ggml_tensor * ggml_sum(
struct ggml_context * ctx,
struct ggml_tensor * a);
// mean along rows
struct ggml_tensor * ggml_mean(
struct ggml_context * ctx,
struct ggml_tensor * a);
// if a is the same shape as b, and a is not parameter, return a
// otherwise, return a new tensor: repeat(a) to fit in b
struct ggml_tensor * ggml_repeat(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
struct ggml_tensor * ggml_abs(
struct ggml_context * ctx,
struct ggml_tensor * a);
struct ggml_tensor * ggml_sgn(
struct ggml_context * ctx,
struct ggml_tensor * a);
struct ggml_tensor * ggml_neg(
struct ggml_context * ctx,
struct ggml_tensor * a);
struct ggml_tensor * ggml_step(
struct ggml_context * ctx,
struct ggml_tensor * a);
struct ggml_tensor * ggml_relu(
struct ggml_context * ctx,
struct ggml_tensor * a);
// TODO: double-check this computation is correct
struct ggml_tensor * ggml_gelu(
struct ggml_context * ctx,
struct ggml_tensor * a);
// normalize along rows
// TODO: eps is hardcoded to 1e-5 for now
struct ggml_tensor * ggml_norm(
struct ggml_context * ctx,
struct ggml_tensor * a);
// A: m rows, n columns
// B: p rows, n columns (i.e. we transpose it internally)
// result is m columns, p rows
struct ggml_tensor * ggml_mul_mat(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
//
// operations on tensors without backpropagation
//
// in-place, returns view(a)
struct ggml_tensor * ggml_scale(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
// a -> b, return view(b)
struct ggml_tensor * ggml_cpy(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
// return view(a), b specifies the new shape
// TODO: when we start computing gradient, make a copy instead of view
struct ggml_tensor * ggml_reshape(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
// return view(a)
// TODO: when we start computing gradient, make a copy instead of view
struct ggml_tensor * ggml_reshape_2d(
struct ggml_context * ctx,
struct ggml_tensor * a,
int ne0,
int ne1);
// return view(a)
// TODO: when we start computing gradient, make a copy instead of view
struct ggml_tensor * ggml_reshape_3d(
struct ggml_context * ctx,
struct ggml_tensor * a,
int ne0,
int ne1,
int ne2);
// offset in bytes
struct ggml_tensor * ggml_view_1d(
struct ggml_context * ctx,
struct ggml_tensor * a,
int ne0,
size_t offset);
struct ggml_tensor * ggml_view_2d(
struct ggml_context * ctx,
struct ggml_tensor * a,
int ne0,
int ne1,
size_t nb1, // row stride in bytes
size_t offset);
struct ggml_tensor * ggml_permute(
struct ggml_context * ctx,
struct ggml_tensor * a,
int axis0,
int axis1,
int axis2,
int axis3);
// alias for ggml_permute(ctx, a, 1, 0, 2, 3)
struct ggml_tensor * ggml_transpose(
struct ggml_context * ctx,
struct ggml_tensor * a);
struct ggml_tensor * ggml_get_rows(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
// set elements above the diagonal to -INF
// in-place, returns view(a)
struct ggml_tensor * ggml_diag_mask_inf(
struct ggml_context * ctx,
struct ggml_tensor * a,
int n_past);
// in-place, returns view(a)
struct ggml_tensor * ggml_soft_max(
struct ggml_context * ctx,
struct ggml_tensor * a);
// rotary position embedding
// in-place, returns view(a)
// if mode == 1, skip n_past elements
// TODO: avoid creating a new tensor every time
struct ggml_tensor * ggml_rope(
struct ggml_context * ctx,
struct ggml_tensor * a,
int n_past,
int n_dims,
int mode);
// padding = 1
// TODO: we don't support extra parameters for now
// that's why we are hard-coding the stride, padding, and dilation
// not great ..
struct ggml_tensor * ggml_conv_1d_1s(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
struct ggml_tensor * ggml_conv_1d_2s(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
struct ggml_tensor * ggml_flash_attn(
struct ggml_context * ctx,
struct ggml_tensor * q,
struct ggml_tensor * k,
struct ggml_tensor * v,
bool masked);
struct ggml_tensor * ggml_flash_ff(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b0,
struct ggml_tensor * b1,
struct ggml_tensor * c0,
struct ggml_tensor * c1);
//
// automatic differentiation
//
void ggml_set_param(
struct ggml_context * ctx,
struct ggml_tensor * tensor);
void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
struct ggml_cgraph ggml_build_forward (struct ggml_tensor * tensor);
struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cgraph * gf, bool keep);
void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph);
void ggml_graph_reset (struct ggml_cgraph * cgraph);
// print info and performance information for the graph
void ggml_graph_print(const struct ggml_cgraph * cgraph);
// dump the graph into a file using the dot format
void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename);
//
// optimization
//
// optimization methods
enum ggml_opt_type {
GGML_OPT_ADAM,
GGML_OPT_LBFGS,
};
// linesearch methods
enum ggml_linesearch {
GGML_LINESEARCH_DEFAULT = 1,
GGML_LINESEARCH_BACKTRACKING_ARMIJO = 0,
GGML_LINESEARCH_BACKTRACKING_WOLFE = 1,
GGML_LINESEARCH_BACKTRACKING_STRONG_WOLFE = 2,
};
// optimization return values
enum ggml_opt_result {
GGML_OPT_OK = 0,
GGML_OPT_DID_NOT_CONVERGE,
GGML_OPT_NO_CONTEXT,
GGML_OPT_INVALID_WOLFE,
GGML_OPT_FAIL,
GGML_LINESEARCH_FAIL = -128,
GGML_LINESEARCH_MINIMUM_STEP,
GGML_LINESEARCH_MAXIMUM_STEP,
GGML_LINESEARCH_MAXIMUM_ITERATIONS,
GGML_LINESEARCH_INVALID_PARAMETERS,
};
// optimization parameters
//
// see ggml.c (ggml_opt_default_params) for default values
//
struct ggml_opt_params {
enum ggml_opt_type type;
int n_threads;
// delta-based convergence test
//
// if past == 0 - disabled
// if past > 0:
// stop if |f(x) - f(x_past)| < delta * max(1, |f(x)|)
//
int past;
float delta;
// maximum number of iterations without improvement
//
// if 0 - disabled
// if > 0:
// assume convergence if no cost improvement in this number of iterations
//
int max_no_improvement;
bool print_forward_graph;
bool print_backward_graph;
// ADAM parameters
struct {
int n_iter;
float alpha; // learning rate
float beta1;
float beta2;
float eps; // epsilon for numerical stability
float eps_f; // epsilon for convergence test
float eps_g; // epsilon for convergence test
} adam;
// LBFGS parameters
struct {
int m; // number of corrections to approximate the inv. Hessian
int n_iter;
int max_linesearch;
float eps; // convergence tolerance
float ftol; // line search tolerance
float wolfe;
float min_step;
float max_step;
enum ggml_linesearch linesearch;
} lbfgs;
};
struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type);
// optimize the function defined by the tensor f
enum ggml_opt_result ggml_opt(
struct ggml_context * ctx,
struct ggml_opt_params params,
struct ggml_tensor * f);
//
// system info
//
int ggml_cpu_has_avx(void);
int ggml_cpu_has_avx2(void);
int ggml_cpu_has_avx512(void);
int ggml_cpu_has_fma(void);
int ggml_cpu_has_neon(void);
int ggml_cpu_has_arm_fma(void);
int ggml_cpu_has_f16c(void);
int ggml_cpu_has_fp16_va(void);
int ggml_cpu_has_wasm_simd(void);
int ggml_cpu_has_blas(void);
int ggml_cpu_has_sse3(void);
int ggml_cpu_has_vsx(void);
#ifdef __cplusplus
}
#endif

854
otherarch/gpt2.cpp Normal file
View file

@ -0,0 +1,854 @@
#include "ggml.h"
#include "utils.h"
#include <cassert>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <fstream>
#include <map>
#include <string>
#include <vector>
#include <iostream>
#include <unistd.h>
// 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;
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
struct ggml_tensor * lm_head; // language model head
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 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*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype); // wte
ctx_size += n_ctx*n_embd*ggml_type_sizef(GGML_TYPE_F32); // wpe
ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype); // lm_head
ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_g
ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_b
ctx_size += n_layer*(3*n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_attn_w
ctx_size += n_layer*( 3*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_attn_b
ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_proj_w
ctx_size += n_layer*( n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_proj_b
ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype)); // c_mlp_fc_w
ctx_size += n_layer*( 4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype)); // c_mlp_proj_w
ctx_size += n_layer*( n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
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 += (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);
model.lm_head = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
// 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;
model.tensors["model/lm_head"] = model.lm_head;
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, n_embd, 3*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, n_embd, 4*n_embd);
layer.c_mlp_fc_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
layer.c_mlp_proj_w = 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;
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;
bool has_lm_head = false;
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;
}
if (0) {
static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", };
printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ftype_str[ftype], ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
}
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 ((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;
}
fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
// GPT-2 models share the WTE tensor as the LM head
if (name == "model/wte" && has_lm_head == false) {
memcpy(model.lm_head->data, tensor->data, ggml_nbytes(tensor));
}
if (name == "model/lm_head") {
has_lm_head = true;
}
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 logits for 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 = 256u*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,
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_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, model.memory_v->type, n_past + N, n_embd/n_head, n_head));
// 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,
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,
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,
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.lm_head
// [ 768, N] - inpL
inpL = ggml_mul_mat(ctx0, model.lm_head, 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 just for 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;
}
int main(int argc, char ** argv) {
ggml_time_init();
const int64_t t_main_start_us = ggml_time_us();
gpt_params params;
params.model = "models/gpt-2-117M/ggml-model.bin";
if (gpt_params_parse(argc, argv, params) == false) {
return 1;
}
if (params.seed < 0) {
params.seed = time(NULL);
}
printf("%s: seed = %d\n", __func__, params.seed);
std::mt19937 rng(params.seed);
if (params.prompt.empty()) {
if( !isatty(STDIN_FILENO) ){
std::string line;
while( std::getline(std::cin, line) ){
params.prompt = params.prompt + "\n" + line;
}
} else {
params.prompt = gpt_random_prompt(rng);
}
}
int64_t t_load_us = 0;
gpt_vocab vocab;
gpt2_model model;
// load the model
{
const int64_t t_start_us = ggml_time_us();
if (!gpt2_model_load(params.model, model, vocab)) {
fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
return 1;
}
t_load_us = ggml_time_us() - t_start_us;
}
int n_past = 0;
int64_t t_sample_us = 0;
int64_t t_predict_us = 0;
std::vector<float> logits;
// tokenize the prompt
std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(vocab, params.prompt);
params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int) embd_inp.size());
printf("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
printf("%s: number of tokens in prompt = %zu, first 8 tokens: ", __func__, embd_inp.size());
for (int i = 0; i < std::min(8, (int) embd_inp.size()); i++) {
printf("%d ", embd_inp[i]);
}
printf("\n\n");
// submit the input prompt token-by-token
// this reduces the memory usage during inference, at the cost of a bit of speed at the beginning
std::vector<gpt_vocab::id> embd;
// determine the required inference memory per token:
size_t mem_per_token = 0;
gpt2_eval(model, params.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token);
for (int i = embd.size(); i < embd_inp.size() + params.n_predict; i++) {
// predict
if (embd.size() > 0) {
const int64_t t_start_us = ggml_time_us();
if (!gpt2_eval(model, params.n_threads, n_past, embd, logits, mem_per_token)) {
printf("Failed to predict\n");
return 1;
}
t_predict_us += ggml_time_us() - t_start_us;
}
n_past += embd.size();
embd.clear();
if (i >= embd_inp.size()) {
// sample next token
const int top_k = params.top_k;
const float top_p = params.top_p;
const float temp = params.temp;
const int n_vocab = model.hparams.n_vocab;
gpt_vocab::id id = 0;
{
const int64_t t_start_sample_us = ggml_time_us();
id = gpt_sample_top_k_top_p(vocab, logits.data() + (logits.size() - n_vocab), top_k, top_p, temp, rng);
t_sample_us += ggml_time_us() - t_start_sample_us;
}
// add it to the context
embd.push_back(id);
} else {
// if here, it means we are still processing the input prompt
for (int k = i; k < embd_inp.size(); k++) {
embd.push_back(embd_inp[k]);
if (embd.size() >= params.n_batch) {
break;
}
}
i += embd.size() - 1;
}
// display text
for (auto id : embd) {
printf("%s", vocab.id_to_token[id].c_str());
}
fflush(stdout);
// end of text token
if (embd.back() == 50256) {
break;
}
}
// report timing
{
const int64_t t_main_end_us = ggml_time_us();
printf("\n\n");
printf("%s: mem per token = %8zu bytes\n", __func__, mem_per_token);
printf("%s: load time = %8.2f ms\n", __func__, t_load_us/1000.0f);
printf("%s: sample time = %8.2f ms\n", __func__, t_sample_us/1000.0f);
printf("%s: predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past);
printf("%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
}
ggml_free(model.ctx);
return 0;
}

323
otherarch/gpt2_quantize.cpp Normal file
View file

@ -0,0 +1,323 @@
#include "ggml.h"
#include "utils.h"
#include <cassert>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <fstream>
#include <map>
#include <string>
#include <vector>
#include <regex>
// 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;
};
// quantize a model
bool gpt2_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));
}
gpt2_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_head, sizeof(hparams.n_head));
finp.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
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_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_head, sizeof(hparams.n_head));
fout.write((char *) &hparams.n_layer, sizeof(hparams.n_layer));
fout.write((char *) &itype, sizeof(hparams.f16));
}
// 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;
}
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("%24s - [%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 = {
"model/wte",
"model/lm_head",
"model/h.*/attn/c_attn/w",
"model/h.*/attn/c_proj/w",
"model/h.*/mlp/c_fc/w",
"model/h.*/mlp/c_proj/w",
};
bool quantize = false;
for (const auto & s : k_names) {
if (std::regex_match(name, std::regex(s))) {
quantize = true;
break;
}
}
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], hist_cur.data());
} break;
case GGML_TYPE_Q4_1:
{
cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], 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;
}
// usage:
// ./gpt-2-quantize models/gpt-2-117M/ggml-model.bin models/gpt-2-117M/ggml-model-quant.bin type
//
int main(int argc, char ** argv) {
if (argc != 4) {
fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
fprintf(stderr, " type = 2 - q4_0\n");
fprintf(stderr, " type = 3 - q4_1\n");
return 1;
}
const std::string fname_inp = argv[1];
const std::string fname_out = argv[2];
const int itype = atoi(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 (!gpt2_model_quantize(fname_inp, fname_out, itype)) {
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;
}

715
otherarch/gptj.cpp Normal file
View file

@ -0,0 +1,715 @@
#include "ggml.h"
#include "otherarch.h"
#include "utils.h"
#include <cassert>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <fstream>
#include <map>
#include <string>
#include <vector>
#include <iostream>
#include <unistd.h>
// load the model's weights from a file
bool gptj_model_load(const std::string & fname, gptj_model & model, gpt_vocab & vocab) {
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;
}
}
// 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.n_rot, sizeof(hparams.n_rot));
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: n_rot = %d\n", __func__, hparams.n_rot);
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 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*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // wte
ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // lmh_g
ctx_size += n_vocab*ggml_type_sizef(GGML_TYPE_F32); // lmh_b
ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_q_proj_w
ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_k_proj_w
ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_v_proj_w
ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_proj_w
ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype)); // c_mlp_fc_w
ctx_size += n_layer*( 4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype)); // c_mlp_proj_w
ctx_size += n_layer*( n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
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.wte = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
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.lmh_g = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
model.lmh_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_vocab);
// map by name
model.tensors["transformer.wte.weight"] = model.wte;
model.tensors["transformer.ln_f.weight"] = model.ln_f_g;
model.tensors["transformer.ln_f.bias"] = model.ln_f_b;
model.tensors["lm_head.weight"] = model.lmh_g;
model.tensors["lm_head.bias"] = model.lmh_b;
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.c_attn_q_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
layer.c_attn_k_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
layer.c_attn_v_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
layer.c_attn_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
layer.c_mlp_fc_w = ggml_new_tensor_2d(ctx, wtype, n_embd, 4*n_embd);
layer.c_mlp_fc_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
layer.c_mlp_proj_w = 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["transformer.h." + std::to_string(i) + ".ln_1.weight"] = layer.ln_1_g;
model.tensors["transformer.h." + std::to_string(i) + ".ln_1.bias"] = layer.ln_1_b;
model.tensors["transformer.h." + std::to_string(i) + ".attn.q_proj.weight"] = layer.c_attn_q_proj_w;
model.tensors["transformer.h." + std::to_string(i) + ".attn.k_proj.weight"] = layer.c_attn_k_proj_w;
model.tensors["transformer.h." + std::to_string(i) + ".attn.v_proj.weight"] = layer.c_attn_v_proj_w;
model.tensors["transformer.h." + std::to_string(i) + ".attn.out_proj.weight"] = layer.c_attn_proj_w;
model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_in.weight"] = layer.c_mlp_fc_w;
model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_in.bias"] = layer.c_mlp_fc_b;
model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_out.weight"] = layer.c_mlp_proj_w;
model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_out.bias"] = 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
{
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;
}
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;
}
if (0) {
static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", };
printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ftype_str[ftype], ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
}
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 ((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;
}
fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
//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);
total_size += ggml_nbytes(tensor);
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 gptj_eval(
const gptj_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_rot;
const int d_key = n_embd/n_head;
static size_t buf_size = 256u*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));
// wte
struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.wte, embd);
for (int il = 0; il < n_layer; ++il) {
struct ggml_tensor * cur;
// norm
{
cur = ggml_norm(ctx0, inpL);
// cur = ln_1_g*cur + ln_1_b
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));
}
struct ggml_tensor * inpSA = cur;
// self-attention
{
struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].c_attn_q_proj_w, cur);
struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].c_attn_k_proj_w, cur);
struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].c_attn_v_proj_w, 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_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, model.memory_v->type, n_past + N, n_embd/n_head, n_head));
// 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].c_attn_proj_w,
cur);
}
struct ggml_tensor * inpFF = cur;
// feed-forward network
// this is independent of the self-attention result, so it could be done in parallel to the self-attention
{
// note here we pass inpSA instead of cur
cur = ggml_mul_mat(ctx0,
model.layers[il].c_mlp_fc_w,
inpSA);
cur = ggml_add(ctx0,
ggml_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur),
cur);
// GELU activation
cur = ggml_gelu(ctx0, cur);
// projection
// cur = proj_w*cur + proj_b
cur = ggml_mul_mat(ctx0,
model.layers[il].c_mlp_proj_w,
cur);
cur = ggml_add(ctx0,
ggml_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur),
cur);
}
// self-attention + FF
cur = ggml_add(ctx0, cur, inpFF);
// input for next layer
inpL = ggml_add(ctx0, cur, inpL);
}
// norm
{
inpL = ggml_norm(ctx0, inpL);
// inpL = ln_f_g*inpL + ln_f_b
inpL = ggml_add(ctx0,
ggml_mul(ctx0,
ggml_repeat(ctx0, model.ln_f_g, inpL),
inpL),
ggml_repeat(ctx0, model.ln_f_b, inpL));
}
// lm_head
{
inpL = ggml_mul_mat(ctx0, model.lmh_g, inpL);
inpL = ggml_add(ctx0,
ggml_repeat(ctx0, model.lmh_b, inpL),
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;
}
int main(int argc, char ** argv) {
ggml_time_init();
const int64_t t_main_start_us = ggml_time_us();
gpt_params params;
params.model = "models/gpt-j-6B/ggml-model.bin";
if (gpt_params_parse(argc, argv, params) == false) {
return 1;
}
if (params.seed < 0) {
params.seed = time(NULL);
}
printf("%s: seed = %d\n", __func__, params.seed);
std::mt19937 rng(params.seed);
if (params.prompt.empty()) {
if( !isatty(STDIN_FILENO) ){
std::string line;
while( std::getline(std::cin, line) ){
params.prompt = params.prompt + "\n" + line;
}
} else {
params.prompt = gpt_random_prompt(rng);
}
}
int64_t t_load_us = 0;
gpt_vocab vocab;
gptj_model model;
// load the model
{
const int64_t t_start_us = ggml_time_us();
if (!gptj_model_load(params.model, model, vocab)) {
fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
return 1;
}
t_load_us = ggml_time_us() - t_start_us;
}
int n_past = 0;
int64_t t_sample_us = 0;
int64_t t_predict_us = 0;
std::vector<float> logits;
// tokenize the prompt
std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(vocab, params.prompt);
params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int) embd_inp.size());
printf("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
printf("\n");
std::vector<gpt_vocab::id> embd;
// determine the required inference memory per token:
size_t mem_per_token = 0;
gptj_eval(model, params.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token);
for (int i = embd.size(); i < embd_inp.size() + params.n_predict; i++) {
// predict
if (embd.size() > 0) {
const int64_t t_start_us = ggml_time_us();
if (!gptj_eval(model, params.n_threads, n_past, embd, logits, mem_per_token)) {
printf("Failed to predict\n");
return 1;
}
t_predict_us += ggml_time_us() - t_start_us;
}
n_past += embd.size();
embd.clear();
if (i >= embd_inp.size()) {
// sample next token
const int top_k = params.top_k;
const float top_p = params.top_p;
const float temp = params.temp;
const int n_vocab = model.hparams.n_vocab;
gpt_vocab::id id = 0;
{
const int64_t t_start_sample_us = ggml_time_us();
id = gpt_sample_top_k_top_p(vocab, logits.data() + (logits.size() - n_vocab), top_k, top_p, temp, rng);
t_sample_us += ggml_time_us() - t_start_sample_us;
}
// add it to the context
embd.push_back(id);
} else {
// if here, it means we are still processing the input prompt
for (int k = i; k < embd_inp.size(); k++) {
embd.push_back(embd_inp[k]);
if (embd.size() > params.n_batch) {
break;
}
}
i += embd.size() - 1;
}
// display text
for (auto id : embd) {
printf("%s", vocab.id_to_token[id].c_str());
}
fflush(stdout);
// end of text token
if (embd.back() == 50256) {
break;
}
}
// report timing
{
const int64_t t_main_end_us = ggml_time_us();
printf("\n\n");
printf("%s: mem per token = %8zu bytes\n", __func__, mem_per_token);
printf("%s: load time = %8.2f ms\n", __func__, t_load_us/1000.0f);
printf("%s: sample time = %8.2f ms\n", __func__, t_sample_us/1000.0f);
printf("%s: predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past);
printf("%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
}
ggml_free(model.ctx);
return 0;
}

682
otherarch/gptj_old.cpp Normal file
View file

@ -0,0 +1,682 @@
#include "ggml_old.h"
#include "otherarch.h"
#include "utils.h"
#include <cassert>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <fstream>
#include <map>
#include <string>
#include <vector>
#include <iostream>
#include <unistd.h>
// load the model's weights from a file
bool legacy_gptj_model_load(const std::string & fname, gptj_model & model, gpt_vocab & vocab) {
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;
}
}
// 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.n_rot, sizeof(hparams.n_rot));
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: n_rot = %d\n", __func__, hparams.n_rot);
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_embd*n_vocab*ggml_type_size(wtype); // wte
ctx_size += n_embd*n_vocab*ggml_type_size(wtype); // lmh_g
ctx_size += n_vocab*ggml_type_size(GGML_TYPE_F32); // lmh_b
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*n_embd*ggml_type_size(wtype)); // c_attn_q_proj_w
ctx_size += n_layer*(n_embd*n_embd*ggml_type_size(wtype)); // c_attn_k_proj_w
ctx_size += n_layer*(n_embd*n_embd*ggml_type_size(wtype)); // c_attn_v_proj_w
ctx_size += n_layer*(n_embd*n_embd*ggml_type_size(wtype)); // c_attn_proj_w
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_trans
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 += (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.wte = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
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.lmh_g = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
model.lmh_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_vocab);
// map by name
model.tensors["transformer.wte.weight"] = model.wte;
model.tensors["transformer.ln_f.weight"] = model.ln_f_g;
model.tensors["transformer.ln_f.bias"] = model.ln_f_b;
model.tensors["lm_head.weight"] = model.lmh_g;
model.tensors["lm_head.bias"] = model.lmh_b;
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.c_attn_q_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
layer.c_attn_k_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
layer.c_attn_v_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
layer.c_attn_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, 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["transformer.h." + std::to_string(i) + ".ln_1.weight"] = layer.ln_1_g;
model.tensors["transformer.h." + std::to_string(i) + ".ln_1.bias"] = layer.ln_1_b;
model.tensors["transformer.h." + std::to_string(i) + ".attn.q_proj.weight"] = layer.c_attn_q_proj_w;
model.tensors["transformer.h." + std::to_string(i) + ".attn.k_proj.weight"] = layer.c_attn_k_proj_w;
model.tensors["transformer.h." + std::to_string(i) + ".attn.v_proj.weight"] = layer.c_attn_v_proj_w;
model.tensors["transformer.h." + std::to_string(i) + ".attn.out_proj.weight"] = layer.c_attn_proj_w;
model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_in.weight"] = layer.c_mlp_fc_w;
model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_in.bias"] = layer.c_mlp_fc_b;
model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_out.weight"] = layer.c_mlp_proj_w_trans;
model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_out.bias"] = 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
{
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;
}
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 = tensor->type == GGML_TYPE_I8 ? 1 : (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("%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);
total_size += ggml_nbytes(tensor);
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 legacy_gptj_eval(
const gptj_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_rot;
const int d_key = n_embd/n_head;
static size_t buf_size = 256u*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));
// wte
struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.wte, embd);
for (int il = 0; il < n_layer; ++il) {
struct ggml_tensor * cur;
// norm
{
cur = ggml_norm(ctx0, inpL);
// cur = ln_1_g*cur + ln_1_b
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));
}
struct ggml_tensor * inpSA = cur;
// self-attention
{
struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, ggml_transpose(ctx0, model.layers[il].c_attn_q_proj_w), cur);
struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, ggml_transpose(ctx0, model.layers[il].c_attn_k_proj_w), cur);
struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, ggml_transpose(ctx0, model.layers[il].c_attn_v_proj_w), 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,
ggml_transpose(ctx0, model.layers[il].c_attn_proj_w),
cur);
}
struct ggml_tensor * inpFF = cur;
// feed-forward network
// this is independent of the self-attention result, so it could be done in parallel to the self-attention
{
// note here we pass inpSA instead of cur
cur = ggml_mul_mat(ctx0,
ggml_transpose(ctx0, model.layers[il].c_mlp_fc_w),
inpSA);
cur = ggml_add(ctx0,
ggml_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur),
cur);
// GELU activation
cur = ggml_gelu(ctx0, cur);
// projection
// cur = proj_w*cur + proj_b
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);
}
// self-attention + FF
cur = ggml_add(ctx0, cur, inpFF);
// input for next layer
inpL = ggml_add(ctx0, cur, inpL);
}
// norm
{
inpL = ggml_norm(ctx0, inpL);
// inpL = ln_f_g*inpL + ln_f_b
inpL = ggml_add(ctx0,
ggml_mul(ctx0,
ggml_repeat(ctx0, model.ln_f_g, inpL),
inpL),
ggml_repeat(ctx0, model.ln_f_b, inpL));
}
// lm_head
{
inpL = ggml_mul_mat(ctx0, model.lmh_g, inpL);
inpL = ggml_add(ctx0,
ggml_repeat(ctx0, model.lmh_b, inpL),
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;
}
int main(int argc, char ** argv) {
ggml_time_init();
const int64_t t_main_start_us = ggml_time_us();
gpt_params params;
params.model = "models/gpt-j-6B/ggml-model.bin";
if (gpt_params_parse(argc, argv, params) == false) {
return 1;
}
if (params.seed < 0) {
params.seed = time(NULL);
}
printf("%s: seed = %d\n", __func__, params.seed);
std::mt19937 rng(params.seed);
if (params.prompt.empty()) {
if( !isatty(STDIN_FILENO) ){
std::string line;
while( std::getline(std::cin, line) ){
params.prompt = params.prompt + "\n" + line;
}
} else {
params.prompt = gpt_random_prompt(rng);
}
}
int64_t t_load_us = 0;
gpt_vocab vocab;
gptj_model model;
// load the model
{
const int64_t t_start_us = ggml_time_us();
if (!legacy_gptj_model_load(params.model, model, vocab)) {
fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
return 1;
}
t_load_us = ggml_time_us() - t_start_us;
}
int n_past = 0;
int64_t t_sample_us = 0;
int64_t t_predict_us = 0;
std::vector<float> logits;
// tokenize the prompt
std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(vocab, params.prompt);
params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int) embd_inp.size());
printf("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
printf("\n");
std::vector<gpt_vocab::id> embd;
// determine the required inference memory per token:
size_t mem_per_token = 0;
legacy_gptj_eval(model, params.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token);
for (int i = embd.size(); i < embd_inp.size() + params.n_predict; i++) {
// predict
if (embd.size() > 0) {
const int64_t t_start_us = ggml_time_us();
if (!legacy_gptj_eval(model, params.n_threads, n_past, embd, logits, mem_per_token)) {
printf("Failed to predict\n");
return 1;
}
t_predict_us += ggml_time_us() - t_start_us;
}
n_past += embd.size();
embd.clear();
if (i >= embd_inp.size()) {
// sample next token
const int top_k = params.top_k;
const float top_p = params.top_p;
const float temp = params.temp;
const int n_vocab = model.hparams.n_vocab;
gpt_vocab::id id = 0;
{
const int64_t t_start_sample_us = ggml_time_us();
id = gpt_sample_top_k_top_p(vocab, logits.data() + (logits.size() - n_vocab), top_k, top_p, temp, rng);
t_sample_us += ggml_time_us() - t_start_sample_us;
}
// add it to the context
embd.push_back(id);
} else {
// if here, it means we are still processing the input prompt
for (int k = i; k < embd_inp.size(); k++) {
embd.push_back(embd_inp[k]);
if (embd.size() > params.n_batch) {
break;
}
}
i += embd.size() - 1;
}
// display text
for (auto id : embd) {
printf("%s", vocab.id_to_token[id].c_str());
}
fflush(stdout);
// end of text token
if (embd.back() == 50256) {
break;
}
}
// report timing
{
const int64_t t_main_end_us = ggml_time_us();
printf("\n\n");
printf("%s: mem per token = %8zu bytes\n", __func__, mem_per_token);
printf("%s: load time = %8.2f ms\n", __func__, t_load_us/1000.0f);
printf("%s: sample time = %8.2f ms\n", __func__, t_sample_us/1000.0f);
printf("%s: predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past);
printf("%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
}
ggml_free(model.ctx);
return 0;
}

324
otherarch/gptj_quantize.cpp Normal file
View file

@ -0,0 +1,324 @@
#include "ggml.h"
#include "utils.h"
#include <cassert>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <fstream>
#include <map>
#include <string>
#include <vector>
#include <regex>
// default hparams (GPT-J 6B)
struct gptj_hparams {
int32_t n_vocab = 50400;
int32_t n_ctx = 2048;
int32_t n_embd = 4096;
int32_t n_head = 16;
int32_t n_layer = 28;
int32_t n_rot = 64;
int32_t f16 = 1;
};
// quantize a model
bool gptj_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));
}
gptj_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_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_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_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
{
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;
}
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], hist_cur.data());
} break;
case GGML_TYPE_Q4_1:
{
cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], 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;
}
// usage:
// ./gpt-2-quantize models/gpt-2-117M/ggml-model.bin models/gpt-2-117M/ggml-model-quant.bin type
//
int main(int argc, char ** argv) {
if (argc != 4) {
fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
fprintf(stderr, " type = 2 - q4_0\n");
fprintf(stderr, " type = 3 - q4_1\n");
return 1;
}
const std::string fname_inp = argv[1];
const std::string fname_out = argv[2];
const int itype = atoi(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 (!gptj_model_quantize(fname_inp, fname_out, itype)) {
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;
}

74
otherarch/otherarch.h Normal file
View file

@ -0,0 +1,74 @@
#pragma once
#include <cassert>
#include <cinttypes>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <fstream>
#include <iostream>
#include <map>
#include <string>
#include <vector>
#include "utils.h"
// default hparams (GPT-J 6B)
struct gptj_hparams {
int32_t n_vocab = 50400;
int32_t n_ctx = 2048;
int32_t n_embd = 4096;
int32_t n_head = 16;
int32_t n_layer = 28;
int32_t n_rot = 64;
int32_t f16 = 1;
};
struct gptj_layer {
// normalization
struct ggml_tensor * ln_1_g;
struct ggml_tensor * ln_1_b;
// attention
struct ggml_tensor * c_attn_q_proj_w;
struct ggml_tensor * c_attn_k_proj_w;
struct ggml_tensor * c_attn_v_proj_w;
struct ggml_tensor * c_attn_proj_w;
// ff
struct ggml_tensor * c_mlp_fc_w;
struct ggml_tensor * c_mlp_fc_b;
struct ggml_tensor * c_mlp_proj_w;
struct ggml_tensor * c_mlp_proj_w_trans; //for backwards compatibility
struct ggml_tensor * c_mlp_proj_b;
};
struct gptj_model {
gptj_hparams hparams;
// normalization
struct ggml_tensor * ln_f_g;
struct ggml_tensor * ln_f_b;
struct ggml_tensor * wte; // position embedding
struct ggml_tensor * lmh_g; // language model head
struct ggml_tensor * lmh_b; // language model bias
std::vector<gptj_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;
};
bool legacy_gptj_model_load(const std::string &fname, gptj_model &model, gpt_vocab &vocab);
bool legacy_gptj_eval(const gptj_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);

330
otherarch/utils.cpp Normal file
View file

@ -0,0 +1,330 @@
#include "utils.h"
#include <fstream>
#include <regex>
bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
for (int i = 1; i < argc; i++) {
std::string arg = argv[i];
if (arg == "-s" || arg == "--seed") {
params.seed = std::stoi(argv[++i]);
} else if (arg == "-t" || arg == "--threads") {
params.n_threads = std::stoi(argv[++i]);
} else if (arg == "-p" || arg == "--prompt") {
params.prompt = argv[++i];
} else if (arg == "-n" || arg == "--n_predict") {
params.n_predict = std::stoi(argv[++i]);
} else if (arg == "--top_k") {
params.top_k = std::stoi(argv[++i]);
} else if (arg == "--top_p") {
params.top_p = std::stof(argv[++i]);
} else if (arg == "--temp") {
params.temp = std::stof(argv[++i]);
} else if (arg == "-b" || arg == "--batch_size") {
params.n_batch = std::stoi(argv[++i]);
} else if (arg == "-m" || arg == "--model") {
params.model = argv[++i];
} else if (arg == "-h" || arg == "--help") {
gpt_print_usage(argc, argv, params);
exit(0);
} else {
fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
gpt_print_usage(argc, argv, params);
exit(0);
}
}
return true;
}
void gpt_print_usage(int argc, char ** argv, const gpt_params & params) {
fprintf(stderr, "usage: %s [options]\n", argv[0]);
fprintf(stderr, "\n");
fprintf(stderr, "options:\n");
fprintf(stderr, " -h, --help show this help message and exit\n");
fprintf(stderr, " -s SEED, --seed SEED RNG seed (default: -1)\n");
fprintf(stderr, " -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads);
fprintf(stderr, " -p PROMPT, --prompt PROMPT\n");
fprintf(stderr, " prompt to start generation with (default: random)\n");
fprintf(stderr, " -n N, --n_predict N number of tokens to predict (default: %d)\n", params.n_predict);
fprintf(stderr, " --top_k N top-k sampling (default: %d)\n", params.top_k);
fprintf(stderr, " --top_p N top-p sampling (default: %.1f)\n", params.top_p);
fprintf(stderr, " --temp N temperature (default: %.1f)\n", params.temp);
fprintf(stderr, " -b N, --batch_size N batch size for prompt processing (default: %d)\n", params.n_batch);
fprintf(stderr, " -m FNAME, --model FNAME\n");
fprintf(stderr, " model path (default: %s)\n", params.model.c_str());
fprintf(stderr, "\n");
}
std::string gpt_random_prompt(std::mt19937 & rng) {
const int r = rng() % 10;
switch (r) {
case 0: return "So";
case 1: return "Once upon a time";
case 2: return "When";
case 3: return "The";
case 4: return "After";
case 5: return "If";
case 6: return "import";
case 7: return "He";
case 8: return "She";
case 9: return "They";
default: return "To";
}
return "The";
}
void replace(std::string & str, const std::string & needle, const std::string & replacement) {
size_t pos = 0;
while ((pos = str.find(needle, pos)) != std::string::npos) {
str.replace(pos, needle.length(), replacement);
pos += replacement.length();
}
}
std::map<std::string, int32_t> json_parse(const std::string & fname) {
std::map<std::string, int32_t> result;
// read file into string
std::string json;
{
std::ifstream ifs(fname);
if (!ifs) {
fprintf(stderr, "Failed to open %s\n", fname.c_str());
exit(1);
}
json = std::string((std::istreambuf_iterator<char>(ifs)),
(std::istreambuf_iterator<char>()));
}
if (json[0] != '{') {
return result;
}
// parse json
{
bool has_key = false;
bool in_token = false;
std::string str_key = "";
std::string str_val = "";
int n = json.size();
for (int i = 1; i < n; ++i) {
if (!in_token) {
if (json[i] == ' ') continue;
if (json[i] == '"') {
in_token = true;
continue;
}
} else {
if (json[i] == '\\' && i+1 < n) {
if (has_key == false) {
str_key += json[i];
} else {
str_val += json[i];
}
++i;
} else if (json[i] == '"') {
if (has_key == false) {
has_key = true;
++i;
while (json[i] == ' ') ++i;
++i; // :
while (json[i] == ' ') ++i;
if (json[i] != '\"') {
while (json[i] != ',' && json[i] != '}') {
str_val += json[i++];
}
has_key = false;
} else {
in_token = true;
continue;
}
} else {
has_key = false;
}
::replace(str_key, "\\u0120", " " ); // \u0120 -> space
::replace(str_key, "\\u010a", "\n"); // \u010a -> new line
::replace(str_key, "\\\"", "\""); // \\\" -> "
try {
result[str_key] = std::stoi(str_val);
} catch (...) {
//fprintf(stderr, "%s: ignoring key '%s' with value '%s'\n", fname.c_str(), str_key.c_str(), str_val.c_str());
}
str_key = "";
str_val = "";
in_token = false;
continue;
}
if (has_key == false) {
str_key += json[i];
} else {
str_val += json[i];
}
}
}
}
return result;
}
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;
}
bool gpt_vocab_init(const std::string & fname, gpt_vocab & vocab) {
printf("%s: loading vocab from '%s'\n", __func__, fname.c_str());
vocab.token_to_id = ::json_parse(fname);
for (const auto & kv : vocab.token_to_id) {
vocab.id_to_token[kv.second] = kv.first;
}
printf("%s: vocab size = %d\n", __func__, (int) vocab.token_to_id.size());
// print the vocabulary
//for (auto kv : vocab.token_to_id) {
// printf("'%s' -> %d\n", kv.first.data(), kv.second);
//}
return true;
}
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);
{
const double scale = 1.0/temp;
for (int i = 0; i < n_logits; ++i) {
logits_id.push_back(std::make_pair(logits[i]*scale, 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);
double maxl = -INFINITY;
for (const auto & kv : logits_id) {
maxl = std::max(maxl, kv.first);
}
// compute probs for the top K tokens
std::vector<double> probs;
probs.reserve(logits_id.size());
double sum = 0.0;
for (const auto & kv : logits_id) {
double p = exp(kv.first - maxl);
probs.push_back(p);
sum += p;
}
// normalize the probs
for (auto & p : probs) {
p /= sum;
}
if (top_p < 1.0f) {
double cumsum = 0.0f;
for (int i = 0; i < top_k; i++) {
cumsum += probs[i];
if (cumsum >= top_p) {
top_k = i + 1;
probs.resize(top_k);
logits_id.resize(top_k);
break;
}
}
cumsum = 1.0/cumsum;
for (int i = 0; i < (int) probs.size(); i++) {
probs[i] *= cumsum;
}
}
//printf("\n");
//for (int i = 0; i < (int) probs.size(); i++) {
// printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), probs[i]);
//}
//exit(0);
std::discrete_distribution<> dist(probs.begin(), probs.end());
int idx = dist(rng);
return logits_id[idx].second;
}

83
otherarch/utils.h Normal file
View file

@ -0,0 +1,83 @@
// Various helper functions and utilities
#pragma once
#include <string>
#include <map>
#include <vector>
#include <random>
#include <thread>
//
// CLI argument parsing
//
struct gpt_params {
int32_t seed = -1; // RNG seed
int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
int32_t n_predict = 200; // new tokens to predict
// sampling parameters
int32_t top_k = 40;
float top_p = 0.9f;
float temp = 0.9f;
int32_t n_batch = 8; // batch size for prompt processing
std::string model = "models/gpt-2-117M/ggml-model.bin"; // model path
std::string prompt;
};
bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
void gpt_print_usage(int argc, char ** argv, const gpt_params & params);
std::string gpt_random_prompt(std::mt19937 & rng);
//
// Vocab utils
//
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;
};
void replace(std::string & str, const std::string & needle, const std::string & replacement);
// poor-man's JSON parsing
std::map<std::string, int32_t> json_parse(const std::string & fname);
// split text into tokens
//
// ref: https://github.com/openai/gpt-2/blob/a74da5d99abaaba920de8131d64da2862a8f213b/src/encoder.py#L53
//
// Regex (Python):
// r"""'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+"""
//
// Regex (C++):
// R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)"
//
std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::string & text);
// load the tokens from encoder.json
bool gpt_vocab_init(const std::string & fname, gpt_vocab & vocab);
// sample next token given probabilities for each embedding
//
// - consider only the top K tokens
// - from them, consider only the top tokens with cumulative probability > P
//
// TODO: not sure if this implementation is correct
// TODO: temperature is not implemented
//
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);