vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Various enhancement and integration pygmalion.cpp

Browse source
This commit is contained in:
Concedo 2023-04-03 00:04:43 +08:00
parent 3f4967b827
commit 8dd8ab1659
20 changed files with 2362 additions and 526 deletions
Download patch file Download diff file Expand all files Collapse all files

8
Makefile
View file

@ -31,8 +31,8 @@ endif
# #
# keep standard at C11 and C++11 # keep standard at C11 and C++11
CFLAGS = -I. -O3 -DNDEBUG -std=c11 -fPIC CFLAGS = -I. -Ofast -DNDEBUG -std=c11 -fPIC
CXXFLAGS = -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC CXXFLAGS = -I. -I./examples -Ofast -DNDEBUG -std=c++11 -fPIC
LDFLAGS = LDFLAGS =
#lets try enabling everything #lets try enabling everything
@ -152,7 +152,7 @@ ggml_blas.o: ggml.c ggml.h
$(CC) $(CFLAGS) -DGGML_USE_OPENBLAS -c ggml.c -o ggml_blas.o $(CC) $(CFLAGS) -DGGML_USE_OPENBLAS -c ggml.c -o ggml_blas.o
ggml_v1.o: otherarch/ggml_v1.c otherarch/ggml_v1.h ggml_v1.o: otherarch/ggml_v1.c otherarch/ggml_v1.h
$(CC) $(CFLAGS) -c otherarch/ggml_v1.c -o ggml_v1.o $(CC) $(CFLAGS) -c otherarch/ggml_v1.c -o ggml_v1.o
llama.o: llama.cpp llama.h llama.o: llama.cpp llama.h
$(CXX) $(CXXFLAGS) -c llama.cpp -o llama.o $(CXX) $(CXXFLAGS) -c llama.cpp -o llama.o
@ -193,6 +193,8 @@ perplexity: examples/perplexity/perplexity.cpp ggml.o llama.o common.o
embedding: examples/embedding/embedding.cpp ggml.o llama.o common.o embedding: examples/embedding/embedding.cpp ggml.o llama.o common.o
$(CXX) $(CXXFLAGS) examples/embedding/embedding.cpp ggml.o llama.o common.o -o embedding $(LDFLAGS) $(CXX) $(CXXFLAGS) examples/embedding/embedding.cpp ggml.o llama.o common.o -o embedding $(LDFLAGS)
gptj: ggml_v1.o
$(CXX) $(CXXFLAGS) otherarch/gptj_v1_main.cpp otherarch/utils.cpp ggml_v1.o -o gptj $(LDFLAGS)
# #
# Tests # Tests
# #

31
expose.cpp
View file

@ -24,17 +24,38 @@ extern "C"
{ {
//return val: 0=fail, 1=(original ggml, alpaca), 2=(ggmf), 3=(ggjt) //return val: 0=fail, 1=(original ggml, alpaca), 2=(ggmf), 3=(ggjt)
static FileFormat file_format = FAIL; static FileFormat file_format = FileFormat::BADFORMAT;
bool load_model(const load_model_inputs inputs) bool load_model(const load_model_inputs inputs)
{ {
std::string model = inputs.model_filename; std::string model = inputs.model_filename;
file_format = check_file_format(model.c_str()); file_format = check_file_format(model.c_str());
if(file_format==GPTJ1 || file_format==GPTJ2) if(file_format==FileFormat::GPTJ1 || file_format==FileFormat::GPTJ2 || file_format==FileFormat::GPTJ3)
{ {
printf("\n---\nIdentified as GPT-J model: (ver %d)\nAttempting to Load...\n---\n", file_format); printf("\n---\nIdentified as GPT-J model: (ver %d)\nAttempting to Load...\n---\n", file_format);
return gptj_load_model(inputs, file_format); ModelLoadResult lr = gptj_load_model(inputs, file_format);
if (lr == ModelLoadResult::RETRY_LOAD)
{
file_format = FileFormat::GPTJ2;
printf("\n---\nRetrying as GPT-J model: (ver %d)\nAttempting to Load...\n---\n", file_format);
lr = gptj_load_model(inputs, file_format);
}
if (lr == ModelLoadResult::RETRY_LOAD)
{
file_format = FileFormat::GPTJ3;
printf("\n---\nRetrying as GPT-J model: (ver %d)\nAttempting to Load...\n---\n", file_format);
lr = gptj_load_model(inputs, file_format);
}
if (lr == ModelLoadResult::FAIL || lr == ModelLoadResult::RETRY_LOAD)
{
return false;
}
else
{
return true;
}
} }
else else
{ {
@ -45,7 +66,7 @@ extern "C"
generation_outputs generate(const generation_inputs inputs, generation_outputs &output) generation_outputs generate(const generation_inputs inputs, generation_outputs &output)
{ {
if (file_format == GPTJ1 || file_format == GPTJ2) if (file_format == FileFormat::GPTJ1 || file_format == FileFormat::GPTJ2 || file_format==FileFormat::GPTJ3)
{ {
return gptj_generate(inputs, output); return gptj_generate(inputs, output);
} }

68
gptj_adapter.cpp
View file

@ -17,7 +17,7 @@
#include "otherarch/gptj_v2.cpp" #include "otherarch/gptj_v2.cpp"
//return val: 0=fail, 1=(original ggml, alpaca), 2=(ggmf), 3=(ggjt) //return val: 0=fail, 1=(original ggml, alpaca), 2=(ggmf), 3=(ggjt)
static FileFormat file_format = FileFormat::FAIL; static FileFormat file_format = FileFormat::BADFORMAT;
static gpt_vocab vocab; static gpt_vocab vocab;
static gptj_model_v1 model_v1; static gptj_model_v1 model_v1;
static gptj_model model_v2; static gptj_model model_v2;
@ -30,9 +30,8 @@ static std::vector<gpt_vocab::id> current_context_tokens;
static size_t mem_per_token = 0; static size_t mem_per_token = 0;
static std::vector<float> logits; static std::vector<float> logits;
bool gptj_load_model(const load_model_inputs inputs, FileFormat in_file_format) ModelLoadResult gptj_load_model(const load_model_inputs inputs, FileFormat in_file_format)
{ {
ggml_time_init(); ggml_time_init();
file_format = in_file_format; file_format = in_file_format;
@ -40,20 +39,42 @@ bool gptj_load_model(const load_model_inputs inputs, FileFormat in_file_format)
n_batch = params.n_batch = inputs.batch_size; n_batch = params.n_batch = inputs.batch_size;
modelname = params.model = inputs.model_filename; modelname = params.model = inputs.model_filename;
if (!legacy_gptj_model_load(params.model, model_v1, vocab)) { if (file_format == FileFormat::GPTJ1 || file_format == FileFormat::GPTJ2)
fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str()); {
return false; ModelLoadResult res = legacy_gptj_model_load(params.model, model_v1, vocab, file_format);
if(res==ModelLoadResult::FAIL)
{
fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
return res;
}
else if(res==ModelLoadResult::RETRY_LOAD)
{
printf("\nTensor Transposition Detected! Retrying GPT-J model loading...");
return res;
}
// determine the required inference memory per token:
legacy_gptj_eval(model_v1, params.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token, file_format);
return ModelLoadResult::SUCCESS;
} }
else
if (file_format != FileFormat::GPTJ2)
{ {
printf("\n---\nWarning: Your model has an INVALID or OUTDATED format (ver %d). Please reconvert it for better results!\n---\n", file_format); ModelLoadResult loadresult = gptj_model_load(params.model, model_v2, vocab);
if (loadresult == ModelLoadResult::FAIL)
{
fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
return loadresult;
}
else if (loadresult == ModelLoadResult::RETRY_LOAD)
{
printf("\nTensor Transposition Detected! Retrying GPT-J model loading...");
return loadresult;
}
// determine the required inference memory per token:
gptj_eval(model_v2, params.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token);
return ModelLoadResult::SUCCESS;
} }
// determine the required inference memory per token:
legacy_gptj_eval(model_v1, params.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token);
return true;
} }
@ -82,9 +103,10 @@ generation_outputs gptj_generate(const generation_inputs inputs, generation_outp
std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(vocab, params.prompt); std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(vocab, params.prompt);
//truncate to front of the prompt if its too long //truncate to front of the prompt if its too long
if (embd_inp.size() + params.n_predict > model_v1.hparams.n_ctx) auto nctx = ( (file_format == FileFormat::GPTJ1||file_format == FileFormat::GPTJ2)? model_v1.hparams.n_ctx:model_v2.hparams.n_ctx);
if (embd_inp.size() + params.n_predict > nctx)
{ {
int offset = embd_inp.size() - model_v1.hparams.n_ctx + params.n_predict; int offset = embd_inp.size() - nctx + params.n_predict;
embd_inp = std::vector<llama_token>(embd_inp.begin() + offset, embd_inp.end()); embd_inp = std::vector<llama_token>(embd_inp.begin() + offset, embd_inp.end());
} }
@ -114,7 +136,7 @@ generation_outputs gptj_generate(const generation_inputs inputs, generation_outp
embd_inp.erase(embd_inp.begin(), embd_inp.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 //if using BLAS and prompt is big enough, switch to single thread and use a huge batch
bool blasmode = false;// (embd_inp.size() >= 32 && ggml_cpu_has_blas()); bool blasmode = false; //(embd_inp.size() >= 32 && ggml_cpu_has_blas());
int original_batch = params.n_batch; int original_batch = params.n_batch;
int original_threads = params.n_threads; int original_threads = params.n_threads;
if (blasmode) if (blasmode)
@ -135,7 +157,7 @@ generation_outputs gptj_generate(const generation_inputs inputs, generation_outp
timer_start(); timer_start();
double time1 = 0, time2 = 0; double time1 = 0, time2 = 0;
unsigned int embd_inp_size = embd_inp.size(); unsigned int embd_inp_size = embd_inp.size();
const int n_vocab = model_v1.hparams.n_vocab; const int n_vocab = ((file_format == FileFormat::GPTJ1||file_format == FileFormat::GPTJ2)? model_v1.hparams.n_vocab:model_v2.hparams.n_vocab);
printf("\n"); printf("\n");
@ -156,7 +178,15 @@ generation_outputs gptj_generate(const generation_inputs inputs, generation_outp
printf("\rGenerating (%d / %d tokens)", (1 + params.n_predict - remaining_tokens), params.n_predict); printf("\rGenerating (%d / %d tokens)", (1 + params.n_predict - remaining_tokens), params.n_predict);
} }
if (!legacy_gptj_eval(model_v1, params.n_threads, n_past, embd, logits, mem_per_token)) bool evalres = false;
if(file_format==FileFormat::GPTJ1 || file_format==FileFormat::GPTJ2)
{
evalres = legacy_gptj_eval(model_v1, params.n_threads, n_past, embd, logits, mem_per_token, file_format);
}else
{
evalres = gptj_eval(model_v2, params.n_threads, n_past, embd, logits, mem_per_token);
}
if (!evalres)
{ {
fprintf(stderr, "Failed to predict\n"); fprintf(stderr, "Failed to predict\n");
snprintf(output.text, sizeof(output.text), "%s", ""); snprintf(output.text, sizeof(output.text), "%s", "");

2
llama_adapter.cpp
View file

@ -16,7 +16,7 @@
#include "llamaextra.cpp" #include "llamaextra.cpp"
//return val: 0=fail, 1=(original ggml, alpaca), 2=(ggmf), 3=(ggjt) //return val: 0=fail, 1=(original ggml, alpaca), 2=(ggmf), 3=(ggjt)
static FileFormat file_format = FileFormat::FAIL; static FileFormat file_format = FileFormat::BADFORMAT;
static llama_context_params ctx_params; static llama_context_params ctx_params;
static gpt_params params; static gpt_params params;
static int n_past = 0; static int n_past = 0;

BIN
llamacpp.dll
View file

Binary file not shown.

BIN
llamacpp_blas.dll
View file

Binary file not shown.

13
llamacpp_for_kobold.py
View file

@ -37,12 +37,17 @@ use_blas = False # if true, uses OpenBLAS for acceleration. libopenblas.dll must
def init_library(): def init_library():
global handle, use_blas global handle, use_blas
dir_path = os.path.dirname(os.path.realpath(__file__)) libname = ""
if use_blas: if use_blas:
#OpenBLAS should provide about a 2x speedup on prompt ingestion if compatible. libname = "llamacpp_blas.dll"
handle = ctypes.CDLL(os.path.join(dir_path, "llamacpp_blas.dll"))
else: else:
handle = ctypes.CDLL(os.path.join(dir_path, "llamacpp.dll")) libname = "llamacpp.dll"
print("Initializing dynamic library: " + libname)
dir_path = os.path.dirname(os.path.realpath(__file__))
#OpenBLAS should provide about a 2x speedup on prompt ingestion if compatible.
handle = ctypes.CDLL(os.path.join(dir_path, libname ))
handle.load_model.argtypes = [load_model_inputs] handle.load_model.argtypes = [load_model_inputs]
handle.load_model.restype = ctypes.c_bool handle.load_model.restype = ctypes.c_bool

BIN
main.exe
View file

Binary file not shown.

4
model_adapter.cpp
View file

@ -49,10 +49,10 @@ void print_tok_vec(std::vector<int> &embd)
fin.rdbuf()->pubsetbuf(f_buf.data(), f_buf.size()); fin.rdbuf()->pubsetbuf(f_buf.data(), f_buf.size());
if (!fin) { if (!fin) {
fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str()); fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
return FileFormat::FAIL; return FileFormat::BADFORMAT;
} }
FileFormat fileformat = FileFormat::FAIL; FileFormat fileformat = FileFormat::BADFORMAT;
uint32_t magic; uint32_t magic;
fin.read((char *) &magic, sizeof(magic)); fin.read((char *) &magic, sizeof(magic));
if (magic == 0x67676d6c) { //v1 format ggml, alpaca, old gptj and gpt2 models if (magic == 0x67676d6c) { //v1 format ggml, alpaca, old gptj and gpt2 models

23
model_adapter.h
View file

@ -13,23 +13,30 @@
#include "expose.h" #include "expose.h"
//return val: 0=fail, 1=(original ggml, alpaca), 2=(ggmf), 3=(ggjt)
enum FileFormat enum FileFormat
{ {
FAIL=0, BADFORMAT=0, //unknown, uninit, or failed to load
GGML=1, GGML=1, // 1=(original llama ggml, alpaca, GPT4ALL, GPTJ header)
GGHF=2, GGHF=2, // 2=(llama ggmf)
GGJT=3, GGJT=3, // 3=(llama ggjt)
GPTJ1=100, GPTJ1=100, //the very first super old GPTJ format
GPTJ2=101, GPTJ2=101, //pygmalion, uses old ggml lib
GPTJ3=102, //uses new ggml lib
GPT2=200, GPT2=200,
}; };
enum ModelLoadResult
{
FAIL = 0,
SUCCESS = 1,
RETRY_LOAD = 2, //used if it's suspected that the model is an older format
};
bool llama_load_model(const load_model_inputs inputs, FileFormat file_format); bool llama_load_model(const load_model_inputs inputs, FileFormat file_format);
generation_outputs llama_generate(const generation_inputs inputs, generation_outputs &output); generation_outputs llama_generate(const generation_inputs inputs, generation_outputs &output);
bool gptj_load_model(const load_model_inputs inputs, FileFormat in_file_format); ModelLoadResult gptj_load_model(const load_model_inputs inputs, FileFormat in_file_format);
generation_outputs gptj_generate(const generation_inputs inputs, generation_outputs &output); generation_outputs gptj_generate(const generation_inputs inputs, generation_outputs &output);

1796
otherarch/ggml_v1.c
View file

File diff suppressed because it is too large Load diff

7
otherarch/ggml_v1.h
View file

@ -198,6 +198,8 @@ struct ggml_v1_object;
struct ggml_v1_context; struct ggml_v1_context;
enum ggml_v1_type { enum ggml_v1_type {
GGML_V1_TYPE_Q4_0,
GGML_V1_TYPE_Q4_1,
GGML_V1_TYPE_I8, GGML_V1_TYPE_I8,
GGML_V1_TYPE_I16, GGML_V1_TYPE_I16,
GGML_V1_TYPE_I32, GGML_V1_TYPE_I32,
@ -326,7 +328,10 @@ void ggml_v1_print_objects(const struct ggml_v1_context * ctx);
int ggml_v1_nelements(const struct ggml_v1_tensor * tensor); int ggml_v1_nelements(const struct ggml_v1_tensor * tensor);
size_t ggml_v1_nbytes (const struct ggml_v1_tensor * tensor); size_t ggml_v1_nbytes (const struct ggml_v1_tensor * tensor);
size_t ggml_v1_type_size (enum ggml_v1_type type); int ggml_v1_blck_size (enum ggml_v1_type type);
size_t ggml_v1_type_size (enum ggml_v1_type type); // size in bytes for all elements in a block
float ggml_v1_type_sizef(enum ggml_v1_type type); // ggml_v1_type_size()/ggml_v1_blck_size() as float
size_t ggml_v1_element_size(const struct ggml_v1_tensor * tensor); size_t ggml_v1_element_size(const struct ggml_v1_tensor * tensor);
struct ggml_v1_context * ggml_v1_init(struct ggml_v1_init_params params); struct ggml_v1_context * ggml_v1_init(struct ggml_v1_init_params params);

316
otherarch/gptj_v1.cpp
View file

@ -17,13 +17,15 @@
// load the model's weights from a file // load the model's weights from a file
bool legacy_gptj_model_load(const std::string & fname, gptj_model_v1 & model, gpt_vocab & vocab) { ModelLoadResult legacy_gptj_model_load(const std::string & fname, gptj_model_v1 & model, gpt_vocab & vocab, FileFormat file_format) {
printf("%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str()); printf("%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str());
bool super_old_format = (file_format==FileFormat::GPTJ1);
auto fin = std::ifstream(fname, std::ios::binary); auto fin = std::ifstream(fname, std::ios::binary);
if (!fin) { if (!fin) {
fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str()); fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
return false; return ModelLoadResult::FAIL;
} }
// verify magic // verify magic
@ -32,7 +34,7 @@ bool legacy_gptj_model_load(const std::string & fname, gptj_model_v1 & model, gp
fin.read((char *) &magic, sizeof(magic)); fin.read((char *) &magic, sizeof(magic));
if (magic != 0x67676d6c) { if (magic != 0x67676d6c) {
fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str()); fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
return false; return ModelLoadResult::FAIL;
} }
} }
@ -65,7 +67,7 @@ bool legacy_gptj_model_load(const std::string & fname, gptj_model_v1 & model, gp
if (n_vocab != model.hparams.n_vocab) { if (n_vocab != model.hparams.n_vocab) {
fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n", fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
__func__, fname.c_str(), n_vocab, model.hparams.n_vocab); __func__, fname.c_str(), n_vocab, model.hparams.n_vocab);
return false; return ModelLoadResult::FAIL;
} }
std::string word; std::string word;
@ -81,9 +83,23 @@ bool legacy_gptj_model_load(const std::string & fname, gptj_model_v1 & model, gp
} }
} }
// for the big tensors, we have the option to store the data in 16-bit floats // 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 // in order to save memory and also to speed up the computation
const ggml_v1_type wtype = model.hparams.f16 ? GGML_V1_TYPE_F16 : GGML_V1_TYPE_F32; ggml_v1_type wtype = GGML_V1_TYPE_COUNT;
switch (model.hparams.f16) {
case 0: wtype = GGML_V1_TYPE_F32; break;
case 1: wtype = GGML_V1_TYPE_F16; break;
case 2: wtype = GGML_V1_TYPE_Q4_0; break;
case 3: wtype = GGML_V1_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 ModelLoadResult::FAIL;
}
}
const ggml_v1_type wtype2 = GGML_V1_TYPE_F32;
auto & ctx = model.ctx; auto & ctx = model.ctx;
@ -97,31 +113,31 @@ bool legacy_gptj_model_load(const std::string & fname, gptj_model_v1 & model, gp
const int n_ctx = hparams.n_ctx; const int n_ctx = hparams.n_ctx;
const int n_vocab = hparams.n_vocab; const int n_vocab = hparams.n_vocab;
ctx_size += n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32); // ln_f_g ctx_size += n_embd*ggml_v1_type_sizef(GGML_V1_TYPE_F32); // ln_f_g
ctx_size += n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32); // ln_f_b ctx_size += n_embd*ggml_v1_type_sizef(GGML_V1_TYPE_F32); // ln_f_b
ctx_size += n_embd*n_vocab*ggml_v1_type_size(wtype); // wte ctx_size += n_embd*n_vocab*ggml_v1_type_sizef(wtype); // wte
ctx_size += n_embd*n_vocab*ggml_v1_type_size(wtype); // lmh_g ctx_size += n_embd*n_vocab*ggml_v1_type_sizef(wtype); // lmh_g
ctx_size += n_vocab*ggml_v1_type_size(GGML_V1_TYPE_F32); // lmh_b ctx_size += n_vocab*ggml_v1_type_sizef(GGML_V1_TYPE_F32); // lmh_b
ctx_size += n_layer*(n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32)); // ln_1_g ctx_size += n_layer*(n_embd*ggml_v1_type_sizef(GGML_V1_TYPE_F32)); // ln_1_g
ctx_size += n_layer*(n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32)); // ln_1_b ctx_size += n_layer*(n_embd*ggml_v1_type_sizef(GGML_V1_TYPE_F32)); // ln_1_b
ctx_size += n_layer*(n_embd*n_embd*ggml_v1_type_size(wtype)); // c_attn_q_proj_w ctx_size += n_layer*(n_embd*n_embd*ggml_v1_type_sizef(wtype)); // c_attn_q_proj_w
ctx_size += n_layer*(n_embd*n_embd*ggml_v1_type_size(wtype)); // c_attn_k_proj_w ctx_size += n_layer*(n_embd*n_embd*ggml_v1_type_sizef(wtype)); // c_attn_k_proj_w
ctx_size += n_layer*(n_embd*n_embd*ggml_v1_type_size(wtype)); // c_attn_v_proj_w ctx_size += n_layer*(n_embd*n_embd*ggml_v1_type_sizef(wtype)); // c_attn_v_proj_w
ctx_size += n_layer*(n_embd*n_embd*ggml_v1_type_size(wtype)); // c_attn_proj_w ctx_size += n_layer*(n_embd*n_embd*ggml_v1_type_sizef(wtype)); // c_attn_proj_w
ctx_size += n_layer*(4*n_embd*n_embd*ggml_v1_type_size(wtype)); // c_mlp_fc_w ctx_size += n_layer*(4*n_embd*n_embd*ggml_v1_type_sizef(wtype)); // c_mlp_fc_w
ctx_size += n_layer*( 4*n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32)); // c_mlp_fc_b ctx_size += n_layer*( 4*n_embd*ggml_v1_type_sizef(GGML_V1_TYPE_F32)); // c_mlp_fc_b
ctx_size += n_layer*(4*n_embd*n_embd*ggml_v1_type_size(wtype)); // c_mlp_proj_w_trans ctx_size += n_layer*(4*n_embd*n_embd*ggml_v1_type_sizef(wtype)); // c_mlp_proj_w_trans
ctx_size += n_layer*( n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32)); // c_mlp_proj_b ctx_size += n_layer*( n_embd*ggml_v1_type_sizef(GGML_V1_TYPE_F32)); // c_mlp_proj_b
ctx_size += n_ctx*n_layer*n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32); // memory_k ctx_size += n_ctx*n_layer*n_embd*ggml_v1_type_sizef(GGML_V1_TYPE_F32); // memory_k
ctx_size += n_ctx*n_layer*n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32); // memory_v ctx_size += n_ctx*n_layer*n_embd*ggml_v1_type_sizef(GGML_V1_TYPE_F32); // memory_v
ctx_size += (5 + 10*n_layer)*256; // object overhead ctx_size += (5 + 10*n_layer)*256; // object overhead
@ -138,7 +154,7 @@ bool legacy_gptj_model_load(const std::string & fname, gptj_model_v1 & model, gp
model.ctx = ggml_v1_init(params); model.ctx = ggml_v1_init(params);
if (!model.ctx) { if (!model.ctx) {
fprintf(stderr, "%s: ggml_v1_init() failed\n", __func__); fprintf(stderr, "%s: ggml_v1_init() failed\n", __func__);
return false; return ModelLoadResult::FAIL;
} }
} }
@ -181,8 +197,15 @@ bool legacy_gptj_model_load(const std::string & fname, gptj_model_v1 & model, gp
layer.c_attn_v_proj_w = ggml_v1_new_tensor_2d(ctx, wtype, n_embd, n_embd); layer.c_attn_v_proj_w = ggml_v1_new_tensor_2d(ctx, wtype, n_embd, n_embd);
layer.c_attn_proj_w = ggml_v1_new_tensor_2d(ctx, wtype, n_embd, n_embd); layer.c_attn_proj_w = ggml_v1_new_tensor_2d(ctx, wtype, n_embd, n_embd);
layer.c_mlp_fc_w = ggml_v1_new_tensor_2d(ctx, wtype, 4*n_embd, n_embd); if(super_old_format)
{
layer.c_mlp_fc_w = ggml_v1_new_tensor_2d(ctx, wtype, 4*n_embd, n_embd);
}
else
{
layer.c_mlp_fc_w = ggml_v1_new_tensor_2d(ctx, wtype, n_embd, 4*n_embd);
}
layer.c_mlp_fc_b = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32, 4*n_embd); layer.c_mlp_fc_b = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32, 4*n_embd);
layer.c_mlp_proj_w_trans = ggml_v1_new_tensor_2d(ctx, wtype, 4*n_embd, n_embd); layer.c_mlp_proj_w_trans = ggml_v1_new_tensor_2d(ctx, wtype, 4*n_embd, n_embd);
@ -257,27 +280,55 @@ bool legacy_gptj_model_load(const std::string & fname, gptj_model_v1 & model, gp
if (model.tensors.find(name.data()) == model.tensors.end()) { if (model.tensors.find(name.data()) == model.tensors.end()) {
fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data()); fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
return false; return ModelLoadResult::FAIL;
} }
auto tensor = model.tensors[name.data()]; auto tensor = model.tensors[name.data()];
if (ggml_v1_nelements(tensor) != nelements) { if (ggml_v1_nelements(tensor) != nelements) {
fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data()); fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
return false; return ModelLoadResult::FAIL;
} }
if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) { if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1])
fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n", {
__func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]); //test for transposition and retry older loader
return false; if(tensor->ne[0]==ne[1] && tensor->ne[1]==ne[0] && should_transpose_layer(name))
{
printf("\nFound a transposed tensor. This could be an older model. Retrying load...");
ggml_v1_free(ctx);
return ModelLoadResult::RETRY_LOAD;
}
else
{
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 ModelLoadResult::FAIL;
}
} }
const size_t bpe = tensor->type == GGML_V1_TYPE_I8 ? 1 : (ftype == 0) ? sizeof(float) : sizeof(ggml_v1_fp16_t); 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_v1_nbytes(tensor)/1024.0/1024.0, ggml_v1_nbytes(tensor));
}
if (nelements*bpe != ggml_v1_nbytes(tensor)) { size_t bpe = 0;
switch (ftype) {
case 0: bpe = ggml_v1_type_size(GGML_V1_TYPE_F32); break;
case 1: bpe = ggml_v1_type_size(GGML_V1_TYPE_F16); break;
case 2: bpe = ggml_v1_type_size(GGML_V1_TYPE_Q4_0); assert(ne[0] % 64 == 0); break;
case 3: bpe = ggml_v1_type_size(GGML_V1_TYPE_Q4_1); assert(ne[0] % 64 == 0); break;
default:
{
fprintf(stderr, "%s: unknown ftype %d in model file\n", __func__, ftype);
return ModelLoadResult::FAIL;
}
};
if ((nelements*bpe)/ggml_v1_blck_size(tensor->type) != ggml_v1_nbytes(tensor)) {
fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n", fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
__func__, name.data(), ggml_v1_nbytes(tensor), nelements*bpe); __func__, name.data(), ggml_v1_nbytes(tensor), nelements*bpe);
return false; return ModelLoadResult::FAIL;
} }
fin.read(reinterpret_cast<char *>(tensor->data), ggml_v1_nbytes(tensor)); fin.read(reinterpret_cast<char *>(tensor->data), ggml_v1_nbytes(tensor));
@ -297,7 +348,7 @@ bool legacy_gptj_model_load(const std::string & fname, gptj_model_v1 & model, gp
fin.close(); fin.close();
return true; return ModelLoadResult::SUCCESS;
} }
// evaluate the transformer // evaluate the transformer
@ -316,7 +367,10 @@ bool legacy_gptj_eval(
const int n_past, const int n_past,
const std::vector<gpt_vocab::id> & embd_inp, const std::vector<gpt_vocab::id> & embd_inp,
std::vector<float> & embd_w, std::vector<float> & embd_w,
size_t & mem_per_token) { size_t & mem_per_token,
FileFormat file_format) {
bool super_old_format = (file_format==FileFormat::GPTJ1);
const int N = embd_inp.size(); const int N = embd_inp.size();
const auto & hparams = model.hparams; const auto & hparams = model.hparams;
@ -379,9 +433,21 @@ bool legacy_gptj_eval(
// self-attention // self-attention
{ {
struct ggml_v1_tensor * Qcur = ggml_v1_mul_mat(ctx0, ggml_v1_transpose(ctx0, model.layers[il].c_attn_q_proj_w), cur); struct ggml_v1_tensor * Qcur;
struct ggml_v1_tensor * Kcur = ggml_v1_mul_mat(ctx0, ggml_v1_transpose(ctx0, model.layers[il].c_attn_k_proj_w), cur); struct ggml_v1_tensor * Kcur;
struct ggml_v1_tensor * Vcur = ggml_v1_mul_mat(ctx0, ggml_v1_transpose(ctx0, model.layers[il].c_attn_v_proj_w), cur); struct ggml_v1_tensor * Vcur;
if(super_old_format)
{
Qcur = ggml_v1_mul_mat(ctx0, ggml_v1_transpose(ctx0, model.layers[il].c_attn_q_proj_w), cur);
Kcur = ggml_v1_mul_mat(ctx0, ggml_v1_transpose(ctx0, model.layers[il].c_attn_k_proj_w), cur);
Vcur = ggml_v1_mul_mat(ctx0, ggml_v1_transpose(ctx0, model.layers[il].c_attn_v_proj_w), cur);
}
else
{
Qcur = ggml_v1_mul_mat(ctx0, model.layers[il].c_attn_q_proj_w, cur);
Kcur = ggml_v1_mul_mat(ctx0, model.layers[il].c_attn_k_proj_w, cur);
Vcur = ggml_v1_mul_mat(ctx0, model.layers[il].c_attn_v_proj_w, cur);
}
// store key and value to memory // store key and value to memory
if (N >= 1) { if (N >= 1) {
@ -448,9 +514,18 @@ bool legacy_gptj_eval(
ggml_v1_new_tensor_2d(ctx0, GGML_V1_TYPE_F32, n_embd, N)); ggml_v1_new_tensor_2d(ctx0, GGML_V1_TYPE_F32, n_embd, N));
// projection (no bias) // projection (no bias)
cur = ggml_v1_mul_mat(ctx0, if(super_old_format)
ggml_v1_transpose(ctx0, model.layers[il].c_attn_proj_w), {
cur); cur = ggml_v1_mul_mat(ctx0,
ggml_v1_transpose(ctx0, model.layers[il].c_attn_proj_w),
cur);
}
else
{
cur = ggml_v1_mul_mat(ctx0,
model.layers[il].c_attn_proj_w,
cur);
}
} }
struct ggml_v1_tensor * inpFF = cur; struct ggml_v1_tensor * inpFF = cur;
@ -459,9 +534,16 @@ bool legacy_gptj_eval(
// this is independent of the self-attention result, so it could be done in parallel to the self-attention // 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 // note here we pass inpSA instead of cur
cur = ggml_v1_mul_mat(ctx0, if(super_old_format)
ggml_v1_transpose(ctx0, model.layers[il].c_mlp_fc_w), {
inpSA); cur = ggml_v1_mul_mat(ctx0,
ggml_v1_transpose(ctx0, model.layers[il].c_mlp_fc_w),
inpSA);
}else{
cur = ggml_v1_mul_mat(ctx0,
model.layers[il].c_mlp_fc_w,
inpSA);
}
cur = ggml_v1_add(ctx0, cur = ggml_v1_add(ctx0,
ggml_v1_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur), ggml_v1_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur),
@ -538,145 +620,3 @@ bool legacy_gptj_eval(
return true; return true;
} }
// int main(int argc, char ** argv) {
// ggml_v1_time_init();
// const int64_t t_main_start_us = ggml_v1_time_us();
// gpt_params params;
// params.model = "models/gpt-j-6B/ggml-model.bin";
// if (utils_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 = utils_gpt_random_prompt(rng);
// }
// }
// int64_t t_load_us = 0;
// gpt_vocab vocab;
// gptj_model_v1 model;
// // load the model
// {
// const int64_t t_start_us = ggml_v1_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_v1_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_v1_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_v1_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_v1_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_v1_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_v1_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_v1_free(model.ctx);
// return 0;
// }

145
otherarch/gptj_v1_main.cpp Normal file
View file

@ -0,0 +1,145 @@
#include "gptj_v1.cpp"
int main(int argc, char ** argv) {
ggml_v1_time_init();
const int64_t t_main_start_us = ggml_v1_time_us();
gpt_params params;
params.model = "models/gpt-j-6B/ggml-model.bin";
if (utils_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 = utils_gpt_random_prompt(rng);
}
}
int64_t t_load_us = 0;
gpt_vocab vocab;
gptj_model_v1 model;
FileFormat file_format = FileFormat::GPTJ2;
// load the model
{
const int64_t t_start_us = ggml_v1_time_us();
if (legacy_gptj_model_load(params.model, model, vocab, file_format)!=ModelLoadResult::SUCCESS) {
fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
return 1;
}
t_load_us = ggml_v1_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, file_format);
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_v1_time_us();
if (!legacy_gptj_eval(model, params.n_threads, n_past, embd, logits, mem_per_token,file_format)) {
printf("Failed to predict\n");
return 1;
}
t_predict_us += ggml_v1_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_v1_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_v1_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_v1_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_v1_free(model.ctx);
return 0;
}

199
otherarch/gptj_v2.cpp
View file

@ -14,28 +14,18 @@
#include <iostream> #include <iostream>
#include <unistd.h> #include <unistd.h>
bool should_transpose_layer(std::string name) #include "model_adapter.h"
{
if(name.find(".mlp.fc_in.weight")!=std::string::npos ||
name.find(".attn.out_proj.weight")!=std::string::npos ||
name.find(".attn.q_proj.weight")!=std::string::npos ||
name.find(".attn.k_proj.weight")!=std::string::npos ||
name.find(".attn.v_proj.weight")!=std::string::npos)
{
return true;
}
return false;
}
// load the model's weights from a file // load the model's weights from a file
bool gptj_model_load(const std::string & fname, gptj_model & model, gpt_vocab & vocab) { ModelLoadResult 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()); printf("%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str());
auto fin = std::ifstream(fname, std::ios::binary); auto fin = std::ifstream(fname, std::ios::binary);
if (!fin) { if (!fin) {
fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str()); fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
return false; return ModelLoadResult::FAIL;
} }
// verify magic // verify magic
@ -44,7 +34,7 @@ bool gptj_model_load(const std::string & fname, gptj_model & model, gpt_vocab &
fin.read((char *) &magic, sizeof(magic)); fin.read((char *) &magic, sizeof(magic));
if (magic != 0x67676d6c) { if (magic != 0x67676d6c) {
fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str()); fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
return false; return ModelLoadResult::FAIL;
} }
} }
@ -77,7 +67,7 @@ bool gptj_model_load(const std::string & fname, gptj_model & model, gpt_vocab &
if (n_vocab != model.hparams.n_vocab) { if (n_vocab != model.hparams.n_vocab) {
fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n", fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
__func__, fname.c_str(), n_vocab, model.hparams.n_vocab); __func__, fname.c_str(), n_vocab, model.hparams.n_vocab);
return false; return ModelLoadResult::FAIL;
} }
std::string word; std::string word;
@ -105,7 +95,7 @@ bool gptj_model_load(const std::string & fname, gptj_model & model, gpt_vocab &
{ {
fprintf(stderr, "%s: invalid model file '%s' (bad f16 value %d)\n", fprintf(stderr, "%s: invalid model file '%s' (bad f16 value %d)\n",
__func__, fname.c_str(), model.hparams.f16); __func__, fname.c_str(), model.hparams.f16);
return false; return ModelLoadResult::FAIL;
} }
} }
@ -165,7 +155,7 @@ bool gptj_model_load(const std::string & fname, gptj_model & model, gpt_vocab &
model.ctx = ggml_init(params); model.ctx = ggml_init(params);
if (!model.ctx) { if (!model.ctx) {
fprintf(stderr, "%s: ggml_init() failed\n", __func__); fprintf(stderr, "%s: ggml_init() failed\n", __func__);
return false; return ModelLoadResult::FAIL;
} }
} }
@ -284,20 +274,32 @@ bool gptj_model_load(const std::string & fname, gptj_model & model, gpt_vocab &
if (model.tensors.find(name.data()) == model.tensors.end()) { if (model.tensors.find(name.data()) == model.tensors.end()) {
fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data()); fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
return false; return ModelLoadResult::FAIL;
} }
auto tensor = model.tensors[name.data()]; auto tensor = model.tensors[name.data()];
if (ggml_nelements(tensor) != nelements) { if (ggml_nelements(tensor) != nelements) {
fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data()); fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
return false; return ModelLoadResult::FAIL;
} }
if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) { if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
__func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]); //test for transposition and retry older loader
return false; if(tensor->ne[0]==ne[1] && tensor->ne[1]==ne[0] && should_transpose_layer(name))
{
printf("\nFound a transposed tensor. This could be an older model. Retrying load...");
ggml_free(ctx);
return ModelLoadResult::RETRY_LOAD;
}
else
{
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 ModelLoadResult::FAIL;
}
} }
if (0) { if (0) {
@ -315,14 +317,14 @@ bool gptj_model_load(const std::string & fname, gptj_model & model, gpt_vocab &
default: default:
{ {
fprintf(stderr, "%s: unknown ftype %d in model file\n", __func__, ftype); fprintf(stderr, "%s: unknown ftype %d in model file\n", __func__, ftype);
return false; return ModelLoadResult::FAIL;
} }
}; };
if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) { 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", fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
__func__, name.data(), ggml_nbytes(tensor), nelements*bpe); __func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
return false; return ModelLoadResult::FAIL;
} }
fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor)); fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
@ -342,7 +344,7 @@ bool gptj_model_load(const std::string & fname, gptj_model & model, gpt_vocab &
fin.close(); fin.close();
return true; return ModelLoadResult::SUCCESS;
} }
// evaluate the transformer // evaluate the transformer
@ -584,146 +586,3 @@ bool gptj_eval(
return true; 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 (utils_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 = utils_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;
// }

145
otherarch/gptj_v2_main.cpp Normal file
View file

@ -0,0 +1,145 @@
#include "gptj_v2.cpp"
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 (utils_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 = utils_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)==ModelLoadResult::FAIL) {
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;
}

8
otherarch/otherarch.h
View file

@ -12,7 +12,7 @@
#include <vector> #include <vector>
#include "utils.h" #include "utils.h"
#include "model_adapter.h"
// default hparams (GPT-J 6B) // default hparams (GPT-J 6B)
@ -113,7 +113,7 @@ struct gptj_model {
std::map<std::string, struct ggml_tensor *> tensors; std::map<std::string, struct ggml_tensor *> tensors;
}; };
bool legacy_gptj_model_load(const std::string &fname, gptj_model_v1 &model, gpt_vocab &vocab); ModelLoadResult legacy_gptj_model_load(const std::string &fname, gptj_model_v1 &model, gpt_vocab &vocab, FileFormat file_format);
bool legacy_gptj_eval(const gptj_model_v1 &model, const int n_threads, const int n_past, const std::vector<gpt_vocab::id> &embd_inp, std::vector<float> &embd_w, size_t &mem_per_token); bool legacy_gptj_eval(const gptj_model_v1 &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, FileFormat file_format);
bool gptj_model_load(const std::string &fname, gptj_model &model, gpt_vocab &vocab); ModelLoadResult gptj_model_load(const std::string &fname, gptj_model &model, gpt_vocab &vocab);
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); 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);

111
otherarch/utils.cpp
View file

@ -249,6 +249,103 @@ bool gpt_vocab_init(const std::string & fname, gpt_vocab & vocab) {
return true; return true;
} }
void gptj_sample_top_k(std::vector<std::pair<double, gpt_vocab::id>> & logits_id, int top_k) {
// 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);
}
gpt_vocab::id gptj_sample_top_p_top_k(
const gpt_vocab & vocab,
const float * logits,
std::vector<gpt_vocab::id> & last_n_tokens,
double repeat_penalty,
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) {
// repetition penalty from CTRL paper (https://arxiv.org/abs/1909.05858)
// credit https://github.com/facebookresearch/llama/compare/main...shawwn:llama:main
if (std::find(last_n_tokens.begin(), last_n_tokens.end(), i) != last_n_tokens.end()) {
// if score < 0 then repetition penalty has to multiplied to reduce the previous token probability
if (logits[i] < 0.0) {
logits_id.push_back(std::make_pair(logits[i]*scale*repeat_penalty, i));
} else {
logits_id.push_back(std::make_pair(logits[i]*scale/repeat_penalty, i));
}
} else {
logits_id.push_back(std::make_pair(logits[i]*scale, i));
}
}
}
gptj_sample_top_k(logits_id, 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 < (int) probs.size(); i++) {
cumsum += probs[i];
if (cumsum >= top_p) {
probs.resize(i + 1);
logits_id.resize(i + 1);
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) 10; i++) {
// printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), probs[i]);
//}
//printf("\n\n");
//exit(0);
std::discrete_distribution<> dist(probs.begin(), probs.end());
int idx = dist(rng);
return logits_id[idx].second;
}
gpt_vocab::id gpt_sample_top_k_top_p( gpt_vocab::id gpt_sample_top_k_top_p(
const gpt_vocab & vocab, const gpt_vocab & vocab,
const float * logits, const float * logits,
@ -327,4 +424,18 @@ gpt_vocab::id gpt_sample_top_k_top_p(
int idx = dist(rng); int idx = dist(rng);
return logits_id[idx].second; return logits_id[idx].second;
}
static bool should_transpose_layer(std::string name)
{
if(name.find(".mlp.fc_in.weight")!=std::string::npos ||
name.find(".attn.out_proj.weight")!=std::string::npos ||
name.find(".attn.q_proj.weight")!=std::string::npos ||
name.find(".attn.k_proj.weight")!=std::string::npos ||
name.find(".attn.v_proj.weight")!=std::string::npos)
{
return true;
}
return false;
} }

12
otherarch/utils.h
View file

@ -62,6 +62,18 @@ gpt_vocab::id gpt_sample_top_k_top_p(
double temp, double temp,
std::mt19937 & rng); std::mt19937 & rng);
gpt_vocab::id gptj_sample_top_p_top_k(
const gpt_vocab & vocab,
const float * logits,
std::vector<gpt_vocab::id> & last_n_tokens,
double repeat_penalty,
int top_k,
double top_p,
double temp,
std::mt19937 & rng);
bool utils_gpt_params_parse(int argc, char ** argv, gpt_params & params); bool utils_gpt_params_parse(int argc, char ** argv, gpt_params & params);
void utils_gpt_print_usage(int argc, char ** argv, const gpt_params & params); void utils_gpt_print_usage(int argc, char ** argv, const gpt_params & params);
std::string utils_gpt_random_prompt(std::mt19937 & rng); std::string utils_gpt_random_prompt(std::mt19937 & rng);
static bool should_transpose_layer(std::string name);

BIN
quantize.exe
View file

Binary file not shown.