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

rename gpt to common

Browse source
This commit is contained in:
slaren 2024-10-09 14:07:04 +02:00
parent 672438dce1
commit 3c0b8628cd
36 changed files with 677 additions and 677 deletions
Download patch file Download diff file Expand all files Collapse all files

442
common/arg.cpp
View file

File diff suppressed because it is too large Load diff

26
common/arg.h
View file

@ -18,29 +18,29 @@ struct common_arg {
const char * env = nullptr; const char * env = nullptr;
std::string help; std::string help;
bool is_sparam = false; // is current arg a sampling param? bool is_sparam = false; // is current arg a sampling param?
void (*handler_void) (gpt_params & params) = nullptr; void (*handler_void) (common_params & params) = nullptr;
void (*handler_string) (gpt_params & params, const std::string &) = nullptr; void (*handler_string) (common_params & params, const std::string &) = nullptr;
void (*handler_str_str)(gpt_params & params, const std::string &, const std::string &) = nullptr; void (*handler_str_str)(common_params & params, const std::string &, const std::string &) = nullptr;
void (*handler_int) (gpt_params & params, int) = nullptr; void (*handler_int) (common_params & params, int) = nullptr;
common_arg( common_arg(
const std::initializer_list<const char *> & args, const std::initializer_list<const char *> & args,
const char * value_hint, const char * value_hint,
const std::string & help, const std::string & help,
void (*handler)(gpt_params & params, const std::string &) void (*handler)(common_params & params, const std::string &)
) : args(args), value_hint(value_hint), help(help), handler_string(handler) {} ) : args(args), value_hint(value_hint), help(help), handler_string(handler) {}
common_arg( common_arg(
const std::initializer_list<const char *> & args, const std::initializer_list<const char *> & args,
const char * value_hint, const char * value_hint,
const std::string & help, const std::string & help,
void (*handler)(gpt_params & params, int) void (*handler)(common_params & params, int)
) : args(args), value_hint(value_hint), help(help), handler_int(handler) {} ) : args(args), value_hint(value_hint), help(help), handler_int(handler) {}
common_arg( common_arg(
const std::initializer_list<const char *> & args, const std::initializer_list<const char *> & args,
const std::string & help, const std::string & help,
void (*handler)(gpt_params & params) void (*handler)(common_params & params)
) : args(args), help(help), handler_void(handler) {} ) : args(args), help(help), handler_void(handler) {}
// support 2 values for arg // support 2 values for arg
@ -49,7 +49,7 @@ struct common_arg {
const char * value_hint, const char * value_hint,
const char * value_hint_2, const char * value_hint_2,
const std::string & help, const std::string & help,
void (*handler)(gpt_params & params, const std::string &, const std::string &) void (*handler)(common_params & params, const std::string &, const std::string &)
) : args(args), value_hint(value_hint), value_hint_2(value_hint_2), help(help), handler_str_str(handler) {} ) : args(args), value_hint(value_hint), value_hint_2(value_hint_2), help(help), handler_str_str(handler) {}
common_arg & set_examples(std::initializer_list<enum llama_example> examples); common_arg & set_examples(std::initializer_list<enum llama_example> examples);
@ -61,17 +61,17 @@ struct common_arg {
std::string to_string(); std::string to_string();
}; };
struct gpt_params_context { struct common_params_context {
enum llama_example ex = LLAMA_EXAMPLE_COMMON; enum llama_example ex = LLAMA_EXAMPLE_COMMON;
gpt_params & params; common_params & params;
std::vector<common_arg> options; std::vector<common_arg> options;
void(*print_usage)(int, char **) = nullptr; void(*print_usage)(int, char **) = nullptr;
gpt_params_context(gpt_params & params) : params(params) {} common_params_context(common_params & params) : params(params) {}
}; };
// parse input arguments from CLI // parse input arguments from CLI
// if one argument has invalid value, it will automatically display usage of the specific argument (and not the full usage message) // if one argument has invalid value, it will automatically display usage of the specific argument (and not the full usage message)
bool gpt_params_parse(int argc, char ** argv, gpt_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr); bool common_params_parse(int argc, char ** argv, common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
// function to be used by test-arg-parser // function to be used by test-arg-parser
gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr); common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);

20
common/common.cpp
View file

@ -362,10 +362,10 @@ bool parse_cpu_mask(const std::string & mask, bool (&boolmask)[GGML_MAX_N_THREAD
return true; return true;
} }
void gpt_init() { void common_init() {
llama_log_set([](ggml_log_level level, const char * text, void * /*user_data*/) { llama_log_set([](ggml_log_level level, const char * text, void * /*user_data*/) {
if (LOG_DEFAULT_LLAMA <= gpt_log_verbosity_thold) { if (LOG_DEFAULT_LLAMA <= common_log_verbosity_thold) {
gpt_log_add(gpt_log_main(), level, "%s", text); common_log_add(common_log_main(), level, "%s", text);
} }
}, NULL); }, NULL);
@ -378,7 +378,7 @@ void gpt_init() {
LOG_INF("build: %d (%s) with %s for %s%s\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT, LLAMA_COMPILER, LLAMA_BUILD_TARGET, build_type); LOG_INF("build: %d (%s) with %s for %s%s\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT, LLAMA_COMPILER, LLAMA_BUILD_TARGET, build_type);
} }
std::string gpt_params_get_system_info(const gpt_params & params) { std::string common_params_get_system_info(const common_params & params) {
std::ostringstream os; std::ostringstream os;
os << "system_info: n_threads = " << params.cpuparams.n_threads; os << "system_info: n_threads = " << params.cpuparams.n_threads;
@ -819,9 +819,9 @@ std::string fs_get_cache_file(const std::string & filename) {
// //
// Model utils // Model utils
// //
struct common_init_result llama_init_from_gpt_params(gpt_params & params) { struct common_init_result common_init_from_common_params(common_params & params) {
common_init_result iparams; common_init_result iparams;
auto mparams = common_model_params_from_gpt_params(params); auto mparams = common_model_params_from_common_params(params);
llama_model * model = nullptr; llama_model * model = nullptr;
@ -863,7 +863,7 @@ struct common_init_result llama_init_from_gpt_params(gpt_params & params) {
} }
} }
auto cparams = common_context_params_from_gpt_params(params); auto cparams = common_context_params_from_common_params(params);
llama_context * lctx = llama_new_context_with_model(model, cparams); llama_context * lctx = llama_new_context_with_model(model, cparams);
if (lctx == NULL) { if (lctx == NULL) {
@ -970,7 +970,7 @@ void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_l
} }
} }
struct llama_model_params common_model_params_from_gpt_params(const gpt_params & params) { struct llama_model_params common_model_params_from_common_params(const common_params & params) {
auto mparams = llama_model_default_params(); auto mparams = llama_model_default_params();
if (params.n_gpu_layers != -1) { if (params.n_gpu_layers != -1) {
@ -1022,7 +1022,7 @@ static ggml_type kv_cache_type_from_str(const std::string & s) {
throw std::runtime_error("Invalid cache type: " + s); throw std::runtime_error("Invalid cache type: " + s);
} }
struct llama_context_params common_context_params_from_gpt_params(const gpt_params & params) { struct llama_context_params common_context_params_from_common_params(const common_params & params) {
auto cparams = llama_context_default_params(); auto cparams = llama_context_default_params();
cparams.n_ctx = params.n_ctx; cparams.n_ctx = params.n_ctx;
@ -1946,7 +1946,7 @@ void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const cha
} }
} }
void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const llama_context * lctx, void yaml_dump_non_result_info(FILE * stream, const common_params & params, const llama_context * lctx,
const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc) { const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc) {
const auto & sparams = params.sparams; const auto & sparams = params.sparams;

48
common/common.h
View file

@ -82,14 +82,14 @@ enum llama_example {
LLAMA_EXAMPLE_COUNT, LLAMA_EXAMPLE_COUNT,
}; };
enum gpt_sampler_type { enum common_sampler_type {
GPT_SAMPLER_TYPE_NONE = 0, COMMON_SAMPLER_TYPE_NONE = 0,
GPT_SAMPLER_TYPE_TOP_K = 1, COMMON_SAMPLER_TYPE_TOP_K = 1,
GPT_SAMPLER_TYPE_TOP_P = 2, COMMON_SAMPLER_TYPE_TOP_P = 2,
GPT_SAMPLER_TYPE_MIN_P = 3, COMMON_SAMPLER_TYPE_MIN_P = 3,
GPT_SAMPLER_TYPE_TFS_Z = 4, COMMON_SAMPLER_TYPE_TFS_Z = 4,
GPT_SAMPLER_TYPE_TYPICAL_P = 5, COMMON_SAMPLER_TYPE_TYPICAL_P = 5,
GPT_SAMPLER_TYPE_TEMPERATURE = 6, COMMON_SAMPLER_TYPE_TEMPERATURE = 6,
}; };
// dimensionality reduction methods, used by cvector-generator // dimensionality reduction methods, used by cvector-generator
@ -99,7 +99,7 @@ enum dimre_method {
}; };
// sampler parameters // sampler parameters
struct gpt_sampler_params { struct common_sampler_params {
uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
int32_t n_prev = 64; // number of previous tokens to remember int32_t n_prev = 64; // number of previous tokens to remember
@ -124,13 +124,13 @@ struct gpt_sampler_params {
bool ignore_eos = false; bool ignore_eos = false;
bool no_perf = false; // disable performance metrics bool no_perf = false; // disable performance metrics
std::vector<enum gpt_sampler_type> samplers = { std::vector<enum common_sampler_type> samplers = {
GPT_SAMPLER_TYPE_TOP_K, COMMON_SAMPLER_TYPE_TOP_K,
GPT_SAMPLER_TYPE_TFS_Z, COMMON_SAMPLER_TYPE_TFS_Z,
GPT_SAMPLER_TYPE_TYPICAL_P, COMMON_SAMPLER_TYPE_TYPICAL_P,
GPT_SAMPLER_TYPE_TOP_P, COMMON_SAMPLER_TYPE_TOP_P,
GPT_SAMPLER_TYPE_MIN_P, COMMON_SAMPLER_TYPE_MIN_P,
GPT_SAMPLER_TYPE_TEMPERATURE COMMON_SAMPLER_TYPE_TEMPERATURE
}; };
std::string grammar; // optional BNF-like grammar to constrain sampling std::string grammar; // optional BNF-like grammar to constrain sampling
@ -141,7 +141,7 @@ struct gpt_sampler_params {
std::string print() const; std::string print() const;
}; };
struct gpt_params { struct common_params {
int32_t n_predict = -1; // new tokens to predict int32_t n_predict = -1; // new tokens to predict
int32_t n_ctx = 0; // context size int32_t n_ctx = 0; // context size
int32_t n_batch = 2048; // logical batch size for prompt processing (must be >=32 to use BLAS) int32_t n_batch = 2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
@ -183,7 +183,7 @@ struct gpt_params {
enum llama_pooling_type pooling_type = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings enum llama_pooling_type pooling_type = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
enum llama_attention_type attention_type = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings enum llama_attention_type attention_type = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings
struct gpt_sampler_params sparams; struct common_sampler_params sparams;
std::string model = ""; // model path // NOLINT std::string model = ""; // model path // NOLINT
std::string model_draft = ""; // draft model for speculative decoding // NOLINT std::string model_draft = ""; // draft model for speculative decoding // NOLINT
@ -348,9 +348,9 @@ struct gpt_params {
// call once at the start of a program if it uses libcommon // call once at the start of a program if it uses libcommon
// initializes the logging system and prints info about the build // initializes the logging system and prints info about the build
void gpt_init(); void common_init();
std::string gpt_params_get_system_info(const gpt_params & params); std::string common_params_get_system_info(const common_params & params);
bool parse_cpu_range(const std::string& range, bool(&boolmask)[GGML_MAX_N_THREADS]); bool parse_cpu_range(const std::string& range, bool(&boolmask)[GGML_MAX_N_THREADS]);
bool parse_cpu_mask(const std::string& mask, bool(&boolmask)[GGML_MAX_N_THREADS]); bool parse_cpu_mask(const std::string& mask, bool(&boolmask)[GGML_MAX_N_THREADS]);
@ -410,10 +410,10 @@ struct common_init_result {
std::vector<common_lora_adapter_container> lora_adapters; std::vector<common_lora_adapter_container> lora_adapters;
}; };
struct common_init_result llama_init_from_gpt_params(gpt_params & params); struct common_init_result common_init_from_common_params(common_params & params);
struct llama_model_params common_model_params_from_gpt_params (const gpt_params & params); struct llama_model_params common_model_params_from_common_params (const common_params & params);
struct llama_context_params common_context_params_from_gpt_params (const gpt_params & params); struct llama_context_params common_context_params_from_common_params(const common_params & params);
struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params); struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params);
struct llama_model * common_load_model_from_url(const char * model_url, const char * path_model, const char * hf_token, const struct llama_model_params & params); struct llama_model * common_load_model_from_url(const char * model_url, const char * path_model, const char * hf_token, const struct llama_model_params & params);
@ -554,5 +554,5 @@ void yaml_dump_vector_int (FILE * stream, const char * prop_name, const std
void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const char * data); void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const char * data);
void yaml_dump_non_result_info( void yaml_dump_non_result_info(
FILE * stream, const gpt_params & params, const llama_context * lctx, FILE * stream, const common_params & params, const llama_context * lctx,
const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc); const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);

100
common/log.cpp
View file

@ -8,10 +8,10 @@
#include <thread> #include <thread>
#include <vector> #include <vector>
int gpt_log_verbosity_thold = LOG_DEFAULT_LLAMA; int common_log_verbosity_thold = LOG_DEFAULT_LLAMA;
void gpt_log_set_verbosity_thold(int verbosity) { void common_log_set_verbosity_thold(int verbosity) {
gpt_log_verbosity_thold = verbosity; common_log_verbosity_thold = verbosity;
} }
#define LOG_COL_DEFAULT "\033[0m" #define LOG_COL_DEFAULT "\033[0m"
@ -29,16 +29,16 @@ static int64_t t_us() {
} }
// colors // colors
enum gpt_log_col : int { enum common_log_col : int {
GPT_LOG_COL_DEFAULT = 0, COMMON_LOG_COL_DEFAULT = 0,
GPT_LOG_COL_BOLD, COMMON_LOG_COL_BOLD,
GPT_LOG_COL_RED, COMMON_LOG_COL_RED,
GPT_LOG_COL_GREEN, COMMON_LOG_COL_GREEN,
GPT_LOG_COL_YELLOW, COMMON_LOG_COL_YELLOW,
GPT_LOG_COL_BLUE, COMMON_LOG_COL_BLUE,
GPT_LOG_COL_MAGENTA, COMMON_LOG_COL_MAGENTA,
GPT_LOG_COL_CYAN, COMMON_LOG_COL_CYAN,
GPT_LOG_COL_WHITE, COMMON_LOG_COL_WHITE,
}; };
// disable colors by default // disable colors by default
@ -54,7 +54,7 @@ static std::vector<const char *> g_col = {
"", "",
}; };
struct gpt_log_entry { struct common_log_entry {
enum ggml_log_level level; enum ggml_log_level level;
bool prefix; bool prefix;
@ -71,7 +71,7 @@ struct gpt_log_entry {
if (!fcur) { if (!fcur) {
// stderr displays DBG messages only when their verbosity level is not higher than the threshold // stderr displays DBG messages only when their verbosity level is not higher than the threshold
// these messages will still be logged to a file // these messages will still be logged to a file
if (level == GGML_LOG_LEVEL_DEBUG && gpt_log_verbosity_thold < LOG_DEFAULT_DEBUG) { if (level == GGML_LOG_LEVEL_DEBUG && common_log_verbosity_thold < LOG_DEFAULT_DEBUG) {
return; return;
} }
@ -86,19 +86,19 @@ struct gpt_log_entry {
if (timestamp) { if (timestamp) {
// [M.s.ms.us] // [M.s.ms.us]
fprintf(fcur, "%s%d.%02d.%03d.%03d%s ", fprintf(fcur, "%s%d.%02d.%03d.%03d%s ",
g_col[GPT_LOG_COL_BLUE], g_col[COMMON_LOG_COL_BLUE],
(int) (timestamp / 1000000 / 60), (int) (timestamp / 1000000 / 60),
(int) (timestamp / 1000000 % 60), (int) (timestamp / 1000000 % 60),
(int) (timestamp / 1000 % 1000), (int) (timestamp / 1000 % 1000),
(int) (timestamp % 1000), (int) (timestamp % 1000),
g_col[GPT_LOG_COL_DEFAULT]); g_col[COMMON_LOG_COL_DEFAULT]);
} }
switch (level) { switch (level) {
case GGML_LOG_LEVEL_INFO: fprintf(fcur, "%sI %s", g_col[GPT_LOG_COL_GREEN], g_col[GPT_LOG_COL_DEFAULT]); break; case GGML_LOG_LEVEL_INFO: fprintf(fcur, "%sI %s", g_col[COMMON_LOG_COL_GREEN], g_col[COMMON_LOG_COL_DEFAULT]); break;
case GGML_LOG_LEVEL_WARN: fprintf(fcur, "%sW %s", g_col[GPT_LOG_COL_MAGENTA], "" ); break; case GGML_LOG_LEVEL_WARN: fprintf(fcur, "%sW %s", g_col[COMMON_LOG_COL_MAGENTA], "" ); break;
case GGML_LOG_LEVEL_ERROR: fprintf(fcur, "%sE %s", g_col[GPT_LOG_COL_RED], "" ); break; case GGML_LOG_LEVEL_ERROR: fprintf(fcur, "%sE %s", g_col[COMMON_LOG_COL_RED], "" ); break;
case GGML_LOG_LEVEL_DEBUG: fprintf(fcur, "%sD %s", g_col[GPT_LOG_COL_YELLOW], "" ); break; case GGML_LOG_LEVEL_DEBUG: fprintf(fcur, "%sD %s", g_col[COMMON_LOG_COL_YELLOW], "" ); break;
default: default:
break; break;
} }
@ -107,18 +107,18 @@ struct gpt_log_entry {
fprintf(fcur, "%s", msg.data()); fprintf(fcur, "%s", msg.data());
if (level == GGML_LOG_LEVEL_WARN || level == GGML_LOG_LEVEL_ERROR || level == GGML_LOG_LEVEL_DEBUG) { if (level == GGML_LOG_LEVEL_WARN || level == GGML_LOG_LEVEL_ERROR || level == GGML_LOG_LEVEL_DEBUG) {
fprintf(fcur, "%s", g_col[GPT_LOG_COL_DEFAULT]); fprintf(fcur, "%s", g_col[COMMON_LOG_COL_DEFAULT]);
} }
fflush(fcur); fflush(fcur);
} }
}; };
struct gpt_log { struct common_log {
// default capacity - will be expanded if needed // default capacity - will be expanded if needed
gpt_log() : gpt_log(256) {} common_log() : common_log(256) {}
gpt_log(size_t capacity) { common_log(size_t capacity) {
file = nullptr; file = nullptr;
prefix = false; prefix = false;
timestamps = false; timestamps = false;
@ -137,7 +137,7 @@ struct gpt_log {
resume(); resume();
} }
~gpt_log() { ~common_log() {
pause(); pause();
if (file) { if (file) {
fclose(file); fclose(file);
@ -158,12 +158,12 @@ private:
int64_t t_start; int64_t t_start;
// ring buffer of entries // ring buffer of entries
std::vector<gpt_log_entry> entries; std::vector<common_log_entry> entries;
size_t head; size_t head;
size_t tail; size_t tail;
// worker thread copies into this // worker thread copies into this
gpt_log_entry cur; common_log_entry cur;
public: public:
void add(enum ggml_log_level level, const char * fmt, va_list args) { void add(enum ggml_log_level level, const char * fmt, va_list args) {
@ -219,7 +219,7 @@ public:
tail = (tail + 1) % entries.size(); tail = (tail + 1) % entries.size();
if (tail == head) { if (tail == head) {
// expand the buffer // expand the buffer
std::vector<gpt_log_entry> new_entries(2*entries.size()); std::vector<common_log_entry> new_entries(2*entries.size());
size_t new_tail = 0; size_t new_tail = 0;
@ -320,15 +320,15 @@ public:
pause(); pause();
if (colors) { if (colors) {
g_col[GPT_LOG_COL_DEFAULT] = LOG_COL_DEFAULT; g_col[COMMON_LOG_COL_DEFAULT] = LOG_COL_DEFAULT;
g_col[GPT_LOG_COL_BOLD] = LOG_COL_BOLD; g_col[COMMON_LOG_COL_BOLD] = LOG_COL_BOLD;
g_col[GPT_LOG_COL_RED] = LOG_COL_RED; g_col[COMMON_LOG_COL_RED] = LOG_COL_RED;
g_col[GPT_LOG_COL_GREEN] = LOG_COL_GREEN; g_col[COMMON_LOG_COL_GREEN] = LOG_COL_GREEN;
g_col[GPT_LOG_COL_YELLOW] = LOG_COL_YELLOW; g_col[COMMON_LOG_COL_YELLOW] = LOG_COL_YELLOW;
g_col[GPT_LOG_COL_BLUE] = LOG_COL_BLUE; g_col[COMMON_LOG_COL_BLUE] = LOG_COL_BLUE;
g_col[GPT_LOG_COL_MAGENTA] = LOG_COL_MAGENTA; g_col[COMMON_LOG_COL_MAGENTA] = LOG_COL_MAGENTA;
g_col[GPT_LOG_COL_CYAN] = LOG_COL_CYAN; g_col[COMMON_LOG_COL_CYAN] = LOG_COL_CYAN;
g_col[GPT_LOG_COL_WHITE] = LOG_COL_WHITE; g_col[COMMON_LOG_COL_WHITE] = LOG_COL_WHITE;
} else { } else {
for (size_t i = 0; i < g_col.size(); i++) { for (size_t i = 0; i < g_col.size(); i++) {
g_col[i] = ""; g_col[i] = "";
@ -355,47 +355,47 @@ public:
// public API // public API
// //
struct gpt_log * gpt_log_init() { struct common_log * common_log_init() {
return new gpt_log; return new common_log;
} }
struct gpt_log * gpt_log_main() { struct common_log * common_log_main() {
static struct gpt_log log; static struct common_log log;
return &log; return &log;
} }
void gpt_log_pause(struct gpt_log * log) { void common_log_pause(struct common_log * log) {
log->pause(); log->pause();
} }
void gpt_log_resume(struct gpt_log * log) { void common_log_resume(struct common_log * log) {
log->resume(); log->resume();
} }
void gpt_log_free(struct gpt_log * log) { void common_log_free(struct common_log * log) {
delete log; delete log;
} }
void gpt_log_add(struct gpt_log * log, enum ggml_log_level level, const char * fmt, ...) { void common_log_add(struct common_log * log, enum ggml_log_level level, const char * fmt, ...) {
va_list args; va_list args;
va_start(args, fmt); va_start(args, fmt);
log->add(level, fmt, args); log->add(level, fmt, args);
va_end(args); va_end(args);
} }
void gpt_log_set_file(struct gpt_log * log, const char * file) { void common_log_set_file(struct common_log * log, const char * file) {
log->set_file(file); log->set_file(file);
} }
void gpt_log_set_colors(struct gpt_log * log, bool colors) { void common_log_set_colors(struct common_log * log, bool colors) {
log->set_colors(colors); log->set_colors(colors);
} }
void gpt_log_set_prefix(struct gpt_log * log, bool prefix) { void common_log_set_prefix(struct common_log * log, bool prefix) {
log->set_prefix(prefix); log->set_prefix(prefix);
} }
void gpt_log_set_timestamps(struct gpt_log * log, bool timestamps) { void common_log_set_timestamps(struct common_log * log, bool timestamps) {
log->set_timestamps(timestamps); log->set_timestamps(timestamps);
} }

36
common/log.h
View file

@ -14,23 +14,23 @@
#define LOG_DEFAULT_LLAMA 0 #define LOG_DEFAULT_LLAMA 0
// needed by the LOG_TMPL macro to avoid computing log arguments if the verbosity lower // needed by the LOG_TMPL macro to avoid computing log arguments if the verbosity lower
// set via gpt_log_set_verbosity() // set via common_log_set_verbosity()
extern int gpt_log_verbosity_thold; extern int common_log_verbosity_thold;
void gpt_log_set_verbosity_thold(int verbosity); // not thread-safe void common_log_set_verbosity_thold(int verbosity); // not thread-safe
// the gpt_log uses an internal worker thread to print/write log messages // the common_log uses an internal worker thread to print/write log messages
// when the worker thread is paused, incoming log messages are discarded // when the worker thread is paused, incoming log messages are discarded
struct gpt_log; struct common_log;
struct gpt_log * gpt_log_init(); struct common_log * common_log_init();
struct gpt_log * gpt_log_main(); // singleton, automatically destroys itself on exit struct common_log * common_log_main(); // singleton, automatically destroys itself on exit
void gpt_log_pause (struct gpt_log * log); // pause the worker thread, not thread-safe void common_log_pause (struct common_log * log); // pause the worker thread, not thread-safe
void gpt_log_resume(struct gpt_log * log); // resume the worker thread, not thread-safe void common_log_resume(struct common_log * log); // resume the worker thread, not thread-safe
void gpt_log_free (struct gpt_log * log); void common_log_free (struct common_log * log);
LOG_ATTRIBUTE_FORMAT(3, 4) LOG_ATTRIBUTE_FORMAT(3, 4)
void gpt_log_add(struct gpt_log * log, enum ggml_log_level level, const char * fmt, ...); void common_log_add(struct common_log * log, enum ggml_log_level level, const char * fmt, ...);
// defaults: file = NULL, colors = false, prefix = false, timestamps = false // defaults: file = NULL, colors = false, prefix = false, timestamps = false
// //
@ -54,10 +54,10 @@ void gpt_log_add(struct gpt_log * log, enum ggml_log_level level, const char * f
// D - debug (stderr, V = LOG_DEFAULT_DEBUG) // D - debug (stderr, V = LOG_DEFAULT_DEBUG)
// //
void gpt_log_set_file (struct gpt_log * log, const char * file); // not thread-safe void common_log_set_file (struct common_log * log, const char * file); // not thread-safe
void gpt_log_set_colors (struct gpt_log * log, bool colors); // not thread-safe void common_log_set_colors (struct common_log * log, bool colors); // not thread-safe
void gpt_log_set_prefix (struct gpt_log * log, bool prefix); // whether to output prefix to each log void common_log_set_prefix (struct common_log * log, bool prefix); // whether to output prefix to each log
void gpt_log_set_timestamps(struct gpt_log * log, bool timestamps); // whether to output timestamps in the prefix void common_log_set_timestamps(struct common_log * log, bool timestamps); // whether to output timestamps in the prefix
// helper macros for logging // helper macros for logging
// use these to avoid computing log arguments if the verbosity of the log is higher than the threshold // use these to avoid computing log arguments if the verbosity of the log is higher than the threshold
@ -66,13 +66,13 @@ void gpt_log_set_timestamps(struct gpt_log * log, bool timestamps); // w
// //
// LOG_DBG("this is a debug message: %d\n", expensive_function()); // LOG_DBG("this is a debug message: %d\n", expensive_function());
// //
// this will avoid calling expensive_function() if LOG_DEFAULT_DEBUG > gpt_log_verbosity_thold // this will avoid calling expensive_function() if LOG_DEFAULT_DEBUG > common_log_verbosity_thold
// //
#define LOG_TMPL(level, verbosity, ...) \ #define LOG_TMPL(level, verbosity, ...) \
do { \ do { \
if ((verbosity) <= gpt_log_verbosity_thold) { \ if ((verbosity) <= common_log_verbosity_thold) { \
gpt_log_add(gpt_log_main(), (level), __VA_ARGS__); \ common_log_add(common_log_main(), (level), __VA_ARGS__); \
} \ } \
} while (0) } while (0)

134
common/sampling.cpp
View file

@ -98,8 +98,8 @@ struct ring_buffer {
std::vector<T> data; std::vector<T> data;
}; };
struct gpt_sampler { struct common_sampler {
gpt_sampler_params params; common_sampler_params params;
struct llama_sampler * grmr; struct llama_sampler * grmr;
struct llama_sampler * chain; struct llama_sampler * chain;
@ -125,7 +125,7 @@ struct gpt_sampler {
} }
}; };
std::string gpt_sampler_params::print() const { std::string common_sampler_params::print() const {
char result[1024]; char result[1024];
snprintf(result, sizeof(result), snprintf(result, sizeof(result),
@ -139,12 +139,12 @@ std::string gpt_sampler_params::print() const {
return std::string(result); return std::string(result);
} }
struct gpt_sampler * gpt_sampler_init(const struct llama_model * model, const struct gpt_sampler_params & params) { struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_sampler_params & params) {
llama_sampler_chain_params lparams = llama_sampler_chain_default_params(); llama_sampler_chain_params lparams = llama_sampler_chain_default_params();
lparams.no_perf = params.no_perf; lparams.no_perf = params.no_perf;
auto * result = new gpt_sampler { auto * result = new common_sampler {
/* .params = */ params, /* .params = */ params,
/* .grmr = */ llama_sampler_init_grammar(model, params.grammar.c_str(), "root"), /* .grmr = */ llama_sampler_init_grammar(model, params.grammar.c_str(), "root"),
/* .chain = */ llama_sampler_chain_init(lparams), /* .chain = */ llama_sampler_chain_init(lparams),
@ -175,22 +175,22 @@ struct gpt_sampler * gpt_sampler_init(const struct llama_model * model, const st
if (params.mirostat == 0) { if (params.mirostat == 0) {
for (const auto & cnstr : params.samplers) { for (const auto & cnstr : params.samplers) {
switch (cnstr) { switch (cnstr) {
case GPT_SAMPLER_TYPE_TOP_K: case COMMON_SAMPLER_TYPE_TOP_K:
llama_sampler_chain_add(result->chain, llama_sampler_init_top_k (params.top_k)); llama_sampler_chain_add(result->chain, llama_sampler_init_top_k (params.top_k));
break; break;
case GPT_SAMPLER_TYPE_TOP_P: case COMMON_SAMPLER_TYPE_TOP_P:
llama_sampler_chain_add(result->chain, llama_sampler_init_top_p (params.top_p, params.min_keep)); llama_sampler_chain_add(result->chain, llama_sampler_init_top_p (params.top_p, params.min_keep));
break; break;
case GPT_SAMPLER_TYPE_MIN_P: case COMMON_SAMPLER_TYPE_MIN_P:
llama_sampler_chain_add(result->chain, llama_sampler_init_min_p (params.min_p, params.min_keep)); llama_sampler_chain_add(result->chain, llama_sampler_init_min_p (params.min_p, params.min_keep));
break; break;
case GPT_SAMPLER_TYPE_TFS_Z: case COMMON_SAMPLER_TYPE_TFS_Z:
llama_sampler_chain_add(result->chain, llama_sampler_init_tail_free(params.tfs_z, params.min_keep)); llama_sampler_chain_add(result->chain, llama_sampler_init_tail_free(params.tfs_z, params.min_keep));
break; break;
case GPT_SAMPLER_TYPE_TYPICAL_P: case COMMON_SAMPLER_TYPE_TYPICAL_P:
llama_sampler_chain_add(result->chain, llama_sampler_init_typical (params.typ_p, params.min_keep)); llama_sampler_chain_add(result->chain, llama_sampler_init_typical (params.typ_p, params.min_keep));
break; break;
case GPT_SAMPLER_TYPE_TEMPERATURE: case COMMON_SAMPLER_TYPE_TEMPERATURE:
llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent)); llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
break; break;
default: default:
@ -224,7 +224,7 @@ struct gpt_sampler * gpt_sampler_init(const struct llama_model * model, const st
return result; return result;
} }
void gpt_sampler_free(struct gpt_sampler * gsmpl) { void common_sampler_free(struct common_sampler * gsmpl) {
if (gsmpl) { if (gsmpl) {
llama_sampler_free(gsmpl->grmr); llama_sampler_free(gsmpl->grmr);
@ -234,7 +234,7 @@ void gpt_sampler_free(struct gpt_sampler * gsmpl) {
} }
} }
void gpt_sampler_accept(struct gpt_sampler * gsmpl, llama_token token, bool accept_grammar) { void common_sampler_accept(struct common_sampler * gsmpl, llama_token token, bool accept_grammar) {
if (accept_grammar) { if (accept_grammar) {
llama_sampler_accept(gsmpl->grmr, token); llama_sampler_accept(gsmpl->grmr, token);
} }
@ -244,14 +244,14 @@ void gpt_sampler_accept(struct gpt_sampler * gsmpl, llama_token token, bool acce
gsmpl->prev.push_back(token); gsmpl->prev.push_back(token);
} }
void gpt_sampler_reset(struct gpt_sampler * gsmpl) { void common_sampler_reset(struct common_sampler * gsmpl) {
llama_sampler_reset(gsmpl->grmr); llama_sampler_reset(gsmpl->grmr);
llama_sampler_reset(gsmpl->chain); llama_sampler_reset(gsmpl->chain);
} }
struct gpt_sampler * gpt_sampler_clone(gpt_sampler * gsmpl) { struct common_sampler * common_sampler_clone(common_sampler * gsmpl) {
return new gpt_sampler { return new common_sampler {
/* .params = */ gsmpl->params, /* .params = */ gsmpl->params,
/* .grmr = */ llama_sampler_clone(gsmpl->grmr), /* .grmr = */ llama_sampler_clone(gsmpl->grmr),
/* .chain = */ llama_sampler_clone(gsmpl->chain), /* .chain = */ llama_sampler_clone(gsmpl->chain),
@ -261,7 +261,7 @@ struct gpt_sampler * gpt_sampler_clone(gpt_sampler * gsmpl) {
}; };
} }
void gpt_perf_print(const struct llama_context * ctx, const struct gpt_sampler * gsmpl) { void common_perf_print(const struct llama_context * ctx, const struct common_sampler * gsmpl) {
// TODO: measure grammar performance // TODO: measure grammar performance
if (gsmpl) { if (gsmpl) {
@ -272,7 +272,7 @@ void gpt_perf_print(const struct llama_context * ctx, const struct gpt_sampler *
} }
} }
llama_token gpt_sampler_sample(struct gpt_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first) { llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first) {
gsmpl->set_logits(ctx, idx); gsmpl->set_logits(ctx, idx);
auto & grmr = gsmpl->grmr; auto & grmr = gsmpl->grmr;
@ -318,21 +318,21 @@ llama_token gpt_sampler_sample(struct gpt_sampler * gsmpl, struct llama_context
return cur_p.data[cur_p.selected].id; return cur_p.data[cur_p.selected].id;
} }
uint32_t gpt_sampler_get_seed(const struct gpt_sampler * gsmpl) { uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl) {
return llama_sampler_get_seed(gsmpl->chain); return llama_sampler_get_seed(gsmpl->chain);
} }
// helpers // helpers
llama_token_data_array * gpt_sampler_get_candidates(struct gpt_sampler * gsmpl) { llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl) {
return &gsmpl->cur_p; return &gsmpl->cur_p;
} }
llama_token gpt_sampler_last(const struct gpt_sampler * gsmpl) { llama_token common_sampler_last(const struct common_sampler * gsmpl) {
return gsmpl->prev.rat(0); return gsmpl->prev.rat(0);
} }
std::string gpt_sampler_print(const struct gpt_sampler * gsmpl) { std::string common_sampler_print(const struct common_sampler * gsmpl) {
std::string result = "logits "; std::string result = "logits ";
for (int i = 0; i < llama_sampler_chain_n(gsmpl->chain); i++) { for (int i = 0; i < llama_sampler_chain_n(gsmpl->chain); i++) {
@ -343,7 +343,7 @@ std::string gpt_sampler_print(const struct gpt_sampler * gsmpl) {
return result; return result;
} }
std::string gpt_sampler_prev_str(gpt_sampler * gsmpl, llama_context * ctx_main, int n) { std::string common_sampler_prev_str(common_sampler * gsmpl, llama_context * ctx_main, int n) {
n = std::min(n, (int) gsmpl->prev.size()); n = std::min(n, (int) gsmpl->prev.size());
if (n <= 0) { if (n <= 0) {
@ -364,57 +364,57 @@ std::string gpt_sampler_prev_str(gpt_sampler * gsmpl, llama_context * ctx_main,
return result; return result;
} }
char gpt_sampler_type_to_chr(enum gpt_sampler_type cnstr) { char common_sampler_type_to_chr(enum common_sampler_type cnstr) {
switch (cnstr) { switch (cnstr) {
case GPT_SAMPLER_TYPE_TOP_K: return 'k'; case COMMON_SAMPLER_TYPE_TOP_K: return 'k';
case GPT_SAMPLER_TYPE_TFS_Z: return 'f'; case COMMON_SAMPLER_TYPE_TFS_Z: return 'f';
case GPT_SAMPLER_TYPE_TYPICAL_P: return 'y'; case COMMON_SAMPLER_TYPE_TYPICAL_P: return 'y';
case GPT_SAMPLER_TYPE_TOP_P: return 'p'; case COMMON_SAMPLER_TYPE_TOP_P: return 'p';
case GPT_SAMPLER_TYPE_MIN_P: return 'm'; case COMMON_SAMPLER_TYPE_MIN_P: return 'm';
case GPT_SAMPLER_TYPE_TEMPERATURE: return 't'; case COMMON_SAMPLER_TYPE_TEMPERATURE: return 't';
default : return '?'; default : return '?';
} }
} }
std::string gpt_sampler_type_to_str(enum gpt_sampler_type cnstr) { std::string common_sampler_type_to_str(enum common_sampler_type cnstr) {
switch (cnstr) { switch (cnstr) {
case GPT_SAMPLER_TYPE_TOP_K: return "top_k"; case COMMON_SAMPLER_TYPE_TOP_K: return "top_k";
case GPT_SAMPLER_TYPE_TFS_Z: return "tfs_z"; case COMMON_SAMPLER_TYPE_TFS_Z: return "tfs_z";
case GPT_SAMPLER_TYPE_TYPICAL_P: return "typ_p"; case COMMON_SAMPLER_TYPE_TYPICAL_P: return "typ_p";
case GPT_SAMPLER_TYPE_TOP_P: return "top_p"; case COMMON_SAMPLER_TYPE_TOP_P: return "top_p";
case GPT_SAMPLER_TYPE_MIN_P: return "min_p"; case COMMON_SAMPLER_TYPE_MIN_P: return "min_p";
case GPT_SAMPLER_TYPE_TEMPERATURE: return "temperature"; case COMMON_SAMPLER_TYPE_TEMPERATURE: return "temperature";
default : return ""; default : return "";
} }
} }
std::vector<gpt_sampler_type> gpt_sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) { std::vector<common_sampler_type> common_sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) {
std::unordered_map<std::string, gpt_sampler_type> sampler_canonical_name_map { std::unordered_map<std::string, common_sampler_type> sampler_canonical_name_map {
{ "top_k", GPT_SAMPLER_TYPE_TOP_K }, { "top_k", COMMON_SAMPLER_TYPE_TOP_K },
{ "top_p", GPT_SAMPLER_TYPE_TOP_P }, { "top_p", COMMON_SAMPLER_TYPE_TOP_P },
{ "typ_p", GPT_SAMPLER_TYPE_TYPICAL_P }, { "typ_p", COMMON_SAMPLER_TYPE_TYPICAL_P },
{ "min_p", GPT_SAMPLER_TYPE_MIN_P }, { "min_p", COMMON_SAMPLER_TYPE_MIN_P },
{ "tfs_z", GPT_SAMPLER_TYPE_TFS_Z }, { "tfs_z", COMMON_SAMPLER_TYPE_TFS_Z },
{ "temperature", GPT_SAMPLER_TYPE_TEMPERATURE }, { "temperature", COMMON_SAMPLER_TYPE_TEMPERATURE },
}; };
// since samplers names are written multiple ways // since samplers names are written multiple ways
// make it ready for both system names and input names // make it ready for both system names and input names
std::unordered_map<std::string, gpt_sampler_type> sampler_alt_name_map { std::unordered_map<std::string, common_sampler_type> sampler_alt_name_map {
{ "top-k", GPT_SAMPLER_TYPE_TOP_K }, { "top-k", COMMON_SAMPLER_TYPE_TOP_K },
{ "top-p", GPT_SAMPLER_TYPE_TOP_P }, { "top-p", COMMON_SAMPLER_TYPE_TOP_P },
{ "nucleus", GPT_SAMPLER_TYPE_TOP_P }, { "nucleus", COMMON_SAMPLER_TYPE_TOP_P },
{ "typical-p", GPT_SAMPLER_TYPE_TYPICAL_P }, { "typical-p", COMMON_SAMPLER_TYPE_TYPICAL_P },
{ "typical", GPT_SAMPLER_TYPE_TYPICAL_P }, { "typical", COMMON_SAMPLER_TYPE_TYPICAL_P },
{ "typ-p", GPT_SAMPLER_TYPE_TYPICAL_P }, { "typ-p", COMMON_SAMPLER_TYPE_TYPICAL_P },
{ "typ", GPT_SAMPLER_TYPE_TYPICAL_P }, { "typ", COMMON_SAMPLER_TYPE_TYPICAL_P },
{ "min-p", GPT_SAMPLER_TYPE_MIN_P }, { "min-p", COMMON_SAMPLER_TYPE_MIN_P },
{ "tfs-z", GPT_SAMPLER_TYPE_TFS_Z }, { "tfs-z", COMMON_SAMPLER_TYPE_TFS_Z },
{ "tfs", GPT_SAMPLER_TYPE_TFS_Z }, { "tfs", COMMON_SAMPLER_TYPE_TFS_Z },
{ "temp", GPT_SAMPLER_TYPE_TEMPERATURE }, { "temp", COMMON_SAMPLER_TYPE_TEMPERATURE },
}; };
std::vector<gpt_sampler_type> samplers; std::vector<common_sampler_type> samplers;
samplers.reserve(names.size()); samplers.reserve(names.size());
for (const auto & name : names) { for (const auto & name : names) {
@ -434,17 +434,17 @@ std::vector<gpt_sampler_type> gpt_sampler_types_from_names(const std::vector<std
return samplers; return samplers;
} }
std::vector<gpt_sampler_type> gpt_sampler_types_from_chars(const std::string & chars) { std::vector<common_sampler_type> common_sampler_types_from_chars(const std::string & chars) {
std::unordered_map<char, gpt_sampler_type> sampler_name_map = { std::unordered_map<char, common_sampler_type> sampler_name_map = {
{ gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TOP_K), GPT_SAMPLER_TYPE_TOP_K }, { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_K), COMMON_SAMPLER_TYPE_TOP_K },
{ gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TFS_Z), GPT_SAMPLER_TYPE_TFS_Z }, { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TFS_Z), COMMON_SAMPLER_TYPE_TFS_Z },
{ gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TYPICAL_P), GPT_SAMPLER_TYPE_TYPICAL_P }, { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TYPICAL_P), COMMON_SAMPLER_TYPE_TYPICAL_P },
{ gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TOP_P), GPT_SAMPLER_TYPE_TOP_P }, { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_P), COMMON_SAMPLER_TYPE_TOP_P },
{ gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_MIN_P), GPT_SAMPLER_TYPE_MIN_P }, { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_MIN_P), COMMON_SAMPLER_TYPE_MIN_P },
{ gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TEMPERATURE), GPT_SAMPLER_TYPE_TEMPERATURE } { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TEMPERATURE), COMMON_SAMPLER_TYPE_TEMPERATURE }
}; };
std::vector<gpt_sampler_type> samplers; std::vector<common_sampler_type> samplers;
samplers.reserve(chars.size()); samplers.reserve(chars.size());
for (const auto & c : chars) { for (const auto & c : chars) {

40
common/sampling.h
View file

@ -7,7 +7,7 @@
#include <string> #include <string>
#include <vector> #include <vector>
// gpt_sampler extends llama_sampler with additional functionality: // common_sampler extends llama_sampler with additional functionality:
// //
// - grammar support // - grammar support
// - custom sampler logic based on the parameters // - custom sampler logic based on the parameters
@ -23,30 +23,30 @@
// token in order to verify if it fits the grammar. And only if the token doesn't fit the grammar, the // token in order to verify if it fits the grammar. And only if the token doesn't fit the grammar, the
// grammar constraints are applied to the full vocabulary and the token is resampled. // grammar constraints are applied to the full vocabulary and the token is resampled.
// //
// The gpt_sampler also maintains a container with the last accepted tokens. In the future, this can // The common_sampler also maintains a container with the last accepted tokens. In the future, this can
// be moved into the core llama library. // be moved into the core llama library.
// //
// For convenience, the gpt_sampler also maintains a container with the current candidate tokens. // For convenience, the common_sampler also maintains a container with the current candidate tokens.
// This can be used to access the probabilities of the rest of the non-sampled tokens. // This can be used to access the probabilities of the rest of the non-sampled tokens.
// //
// TODO: measure grammar performance // TODO: measure grammar performance
// //
struct gpt_sampler; struct common_sampler;
// llama_sampler API overloads // llama_sampler API overloads
struct gpt_sampler * gpt_sampler_init(const struct llama_model * model, const struct gpt_sampler_params & params); struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_sampler_params & params);
void gpt_sampler_free(struct gpt_sampler * gsmpl); void common_sampler_free(struct common_sampler * gsmpl);
// if accept_grammar is true, the token is accepted both by the sampling chain and the grammar // if accept_grammar is true, the token is accepted both by the sampling chain and the grammar
void gpt_sampler_accept(struct gpt_sampler * gsmpl, llama_token token, bool accept_grammar); void common_sampler_accept(struct common_sampler * gsmpl, llama_token token, bool accept_grammar);
void gpt_sampler_reset (struct gpt_sampler * gsmpl); void common_sampler_reset (struct common_sampler * gsmpl);
struct gpt_sampler * gpt_sampler_clone (struct gpt_sampler * gsmpl); struct common_sampler * common_sampler_clone (struct common_sampler * gsmpl);
// arguments can be nullptr to skip printing // arguments can be nullptr to skip printing
void gpt_perf_print(const struct llama_context * ctx, const struct gpt_sampler * gsmpl); void common_perf_print(const struct llama_context * ctx, const struct common_sampler * gsmpl);
// extended sampling implementation: // extended sampling implementation:
// //
@ -58,26 +58,26 @@ void gpt_perf_print(const struct llama_context * ctx, const struct gpt_sampler *
// if grammar_first is true, the grammar is applied before the samplers (slower) // if grammar_first is true, the grammar is applied before the samplers (slower)
// useful in cases where all the resulting candidates (not just the sampled one) must fit the grammar // useful in cases where all the resulting candidates (not just the sampled one) must fit the grammar
// //
llama_token gpt_sampler_sample(struct gpt_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first = false); llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first = false);
uint32_t gpt_sampler_get_seed(const struct gpt_sampler * gsmpl); uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl);
// helpers // helpers
// access the internal list of current candidate tokens // access the internal list of current candidate tokens
llama_token_data_array * gpt_sampler_get_candidates(struct gpt_sampler * gsmpl); llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl);
// get the last accepted token // get the last accepted token
llama_token gpt_sampler_last(const struct gpt_sampler * gsmpl); llama_token common_sampler_last(const struct common_sampler * gsmpl);
// print the sampler chain into a string // print the sampler chain into a string
std::string gpt_sampler_print(const struct gpt_sampler * gsmpl); std::string common_sampler_print(const struct common_sampler * gsmpl);
// get a string representation of the last accepted tokens // get a string representation of the last accepted tokens
std::string gpt_sampler_prev_str(gpt_sampler * gsmpl, llama_context * ctx, int n); std::string common_sampler_prev_str(common_sampler * gsmpl, llama_context * ctx, int n);
char gpt_sampler_type_to_chr(enum gpt_sampler_type cnstr); char common_sampler_type_to_chr(enum common_sampler_type cnstr);
std::string gpt_sampler_type_to_str(enum gpt_sampler_type cnstr); std::string common_sampler_type_to_str(enum common_sampler_type cnstr);
std::vector<enum gpt_sampler_type> gpt_sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names); std::vector<enum common_sampler_type> common_sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
std::vector<enum gpt_sampler_type> gpt_sampler_types_from_chars(const std::string & chars); std::vector<enum common_sampler_type> common_sampler_types_from_chars(const std::string & chars);

10
examples/batched-bench/batched-bench.cpp
View file

@ -15,13 +15,13 @@ static void print_usage(int, char ** argv) {
} }
int main(int argc, char ** argv) { int main(int argc, char ** argv) {
gpt_params params; common_params params;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_BENCH, print_usage)) { if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_BENCH, print_usage)) {
return 1; return 1;
} }
gpt_init(); common_init();
int is_pp_shared = params.is_pp_shared; int is_pp_shared = params.is_pp_shared;
@ -36,7 +36,7 @@ int main(int argc, char ** argv) {
// initialize the model // initialize the model
llama_model_params model_params = common_model_params_from_gpt_params(params); llama_model_params model_params = common_model_params_from_common_params(params);
llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params); llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
@ -45,7 +45,7 @@ int main(int argc, char ** argv) {
return 1; return 1;
} }
llama_context_params ctx_params = common_context_params_from_gpt_params(params); llama_context_params ctx_params = common_context_params_from_common_params(params);
// ensure enough sequences are available // ensure enough sequences are available
ctx_params.n_seq_max = n_pl.empty() ? 1 : *std::max_element(n_pl.begin(), n_pl.end()); ctx_params.n_seq_max = n_pl.empty() ? 1 : *std::max_element(n_pl.begin(), n_pl.end());

10
examples/batched/batched.cpp
View file

@ -15,16 +15,16 @@ static void print_usage(int, char ** argv) {
} }
int main(int argc, char ** argv) { int main(int argc, char ** argv) {
gpt_params params; common_params params;
params.prompt = "Hello my name is"; params.prompt = "Hello my name is";
params.n_predict = 32; params.n_predict = 32;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, print_usage)) { if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, print_usage)) {
return 1; return 1;
} }
gpt_init(); common_init();
// number of parallel batches // number of parallel batches
int n_parallel = params.n_parallel; int n_parallel = params.n_parallel;
@ -39,7 +39,7 @@ int main(int argc, char ** argv) {
// initialize the model // initialize the model
llama_model_params model_params = common_model_params_from_gpt_params(params); llama_model_params model_params = common_model_params_from_common_params(params);
llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params); llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
@ -57,7 +57,7 @@ int main(int argc, char ** argv) {
// initialize the context // initialize the context
llama_context_params ctx_params = common_context_params_from_gpt_params(params); llama_context_params ctx_params = common_context_params_from_common_params(params);
ctx_params.n_ctx = n_kv_req; ctx_params.n_ctx = n_kv_req;
ctx_params.n_batch = std::max(n_predict, n_parallel); ctx_params.n_batch = std::max(n_predict, n_parallel);

2
examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
View file

@ -872,7 +872,7 @@ static std::string basename(const std::string &path) {
} }
int main(int argc, char ** argv) { int main(int argc, char ** argv) {
gpt_init(); common_init();
struct train_params params = get_default_train_params(); struct train_params params = get_default_train_params();
if (!params_parse(argc, argv, &params)) { if (!params_parse(argc, argv, &params)) {

8
examples/cvector-generator/cvector-generator.cpp
View file

@ -370,7 +370,7 @@ static void export_gguf(const std::vector<struct ggml_tensor *> & v_ctrl, const
* Load prompt files and completion file. * Load prompt files and completion file.
* Then format each pair of prompt + completion to make an entry. * Then format each pair of prompt + completion to make an entry.
*/ */
static int prepare_entries(gpt_params & params, train_context & ctx_train) { static int prepare_entries(common_params & params, train_context & ctx_train) {
// load prompts // load prompts
std::vector<std::string> positive_prompts = ctrlvec_load_prompt_file(params.cvector_positive_file, true); std::vector<std::string> positive_prompts = ctrlvec_load_prompt_file(params.cvector_positive_file, true);
std::vector<std::string> negative_prompts = ctrlvec_load_prompt_file(params.cvector_negative_file, true); std::vector<std::string> negative_prompts = ctrlvec_load_prompt_file(params.cvector_negative_file, true);
@ -388,9 +388,9 @@ static int prepare_entries(gpt_params & params, train_context & ctx_train) {
} }
int main(int argc, char ** argv) { int main(int argc, char ** argv) {
gpt_params params; common_params params;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_CVECTOR_GENERATOR, print_usage)) { if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_CVECTOR_GENERATOR, print_usage)) {
return 1; return 1;
} }
@ -413,7 +413,7 @@ int main(int argc, char ** argv) {
llama_numa_init(params.numa); llama_numa_init(params.numa);
// load the model to get hparams // load the model to get hparams
common_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_common_params(params);
llama_model * model = llama_init.model; llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context; llama_context * ctx = llama_init.context;

10
examples/embedding/embedding.cpp
View file

@ -79,13 +79,13 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
} }
int main(int argc, char ** argv) { int main(int argc, char ** argv) {
gpt_params params; common_params params;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_EMBEDDING)) { if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_EMBEDDING)) {
return 1; return 1;
} }
gpt_init(); common_init();
params.embedding = true; params.embedding = true;
// For non-causal models, batch size must be equal to ubatch size // For non-causal models, batch size must be equal to ubatch size
@ -95,7 +95,7 @@ int main(int argc, char ** argv) {
llama_numa_init(params.numa); llama_numa_init(params.numa);
// load the model // load the model
common_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_common_params(params);
llama_model * model = llama_init.model; llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context; llama_context * ctx = llama_init.context;
@ -122,7 +122,7 @@ int main(int argc, char ** argv) {
// print system information // print system information
{ {
LOG_INF("\n"); LOG_INF("\n");
LOG_INF("%s\n", gpt_params_get_system_info(params).c_str()); LOG_INF("%s\n", common_params_get_system_info(params).c_str());
} }
// split the prompt into lines // split the prompt into lines

12
examples/eval-callback/eval-callback.cpp
View file

@ -126,7 +126,7 @@ static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
return true; return true;
} }
static bool run(llama_context * ctx, const gpt_params & params) { static bool run(llama_context * ctx, const common_params & params) {
const bool add_bos = llama_add_bos_token(llama_get_model(ctx)); const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, add_bos); std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, add_bos);
@ -142,13 +142,13 @@ static bool run(llama_context * ctx, const gpt_params & params) {
int main(int argc, char ** argv) { int main(int argc, char ** argv) {
callback_data cb_data; callback_data cb_data;
gpt_params params; common_params params;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) { if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
return 1; return 1;
} }
gpt_init(); common_init();
llama_backend_init(); llama_backend_init();
llama_numa_init(params.numa); llama_numa_init(params.numa);
@ -160,7 +160,7 @@ int main(int argc, char ** argv) {
params.warmup = false; params.warmup = false;
// init // init
common_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_common_params(params);
llama_model * model = llama_init.model; llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context; llama_context * ctx = llama_init.context;
@ -172,7 +172,7 @@ int main(int argc, char ** argv) {
// print system information // print system information
{ {
LOG_INF("\n"); LOG_INF("\n");
LOG_INF("%s\n", gpt_params_get_system_info(params).c_str()); LOG_INF("%s\n", common_params_get_system_info(params).c_str());
LOG_INF("\n"); LOG_INF("\n");
} }

4
examples/export-lora/export-lora.cpp
View file

@ -400,9 +400,9 @@ static void print_usage(int, char ** argv) {
} }
int main(int argc, char ** argv) { int main(int argc, char ** argv) {
gpt_params params; common_params params;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_EXPORT_LORA, print_usage)) { if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_EXPORT_LORA, print_usage)) {
return 1; return 1;
} }

4
examples/gen-docs/gen-docs.cpp
View file

@ -50,8 +50,8 @@ static void write_table(std::ofstream & file, std::vector<common_arg *> & opts)
static void export_md(std::string fname, llama_example ex) { static void export_md(std::string fname, llama_example ex) {
std::ofstream file(fname, std::ofstream::out | std::ofstream::trunc); std::ofstream file(fname, std::ofstream::out | std::ofstream::trunc);
gpt_params params; common_params params;
auto ctx_arg = gpt_params_parser_init(params, ex); auto ctx_arg = common_params_parser_init(params, ex);
std::vector<common_arg *> common_options; std::vector<common_arg *> common_options;
std::vector<common_arg *> sparam_options; std::vector<common_arg *> sparam_options;

10
examples/gritlm/gritlm.cpp
View file

@ -152,16 +152,16 @@ static std::string gritlm_instruction(const std::string & instruction) {
} }
int main(int argc, char * argv[]) { int main(int argc, char * argv[]) {
gpt_params params; common_params params;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) { if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
return 1; return 1;
} }
gpt_init(); common_init();
llama_model_params mparams = common_model_params_from_gpt_params(params); llama_model_params mparams = common_model_params_from_common_params(params);
llama_context_params cparams = common_context_params_from_gpt_params(params); llama_context_params cparams = common_context_params_from_common_params(params);
llama_backend_init(); llama_backend_init();

16
examples/imatrix/imatrix.cpp
View file

@ -37,13 +37,13 @@ struct Stats {
class IMatrixCollector { class IMatrixCollector {
public: public:
IMatrixCollector() = default; IMatrixCollector() = default;
void set_params(gpt_params params) { m_params = std::move(params); } void set_params(common_params params) { m_params = std::move(params); }
bool collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data); bool collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data);
void save_imatrix(int ncall = -1) const; void save_imatrix(int ncall = -1) const;
bool load_imatrix(const char * file_name); bool load_imatrix(const char * file_name);
private: private:
std::unordered_map<std::string, Stats> m_stats; std::unordered_map<std::string, Stats> m_stats;
gpt_params m_params; common_params m_params;
std::mutex m_mutex; std::mutex m_mutex;
int m_last_call = 0; int m_last_call = 0;
std::vector<float> m_src1_data; std::vector<float> m_src1_data;
@ -428,7 +428,7 @@ static void process_logits(
} }
} }
static bool compute_imatrix(llama_context * ctx, const gpt_params & params) { static bool compute_imatrix(llama_context * ctx, const common_params & params) {
const bool add_bos = llama_add_bos_token(llama_get_model(ctx)); const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx))); GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
const int n_ctx = llama_n_ctx(ctx); const int n_ctx = llama_n_ctx(ctx);
@ -568,17 +568,17 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
} }
int main(int argc, char ** argv) { int main(int argc, char ** argv) {
gpt_params params; common_params params;
params.n_ctx = 512; params.n_ctx = 512;
params.logits_all = true; params.logits_all = true;
params.escape = false; params.escape = false;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_IMATRIX, print_usage)) { if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_IMATRIX, print_usage)) {
return 1; return 1;
} }
gpt_init(); common_init();
params.n_batch = std::min(params.n_batch, params.n_ctx); params.n_batch = std::min(params.n_batch, params.n_ctx);
@ -607,7 +607,7 @@ int main(int argc, char ** argv) {
params.warmup = false; params.warmup = false;
// init // init
common_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_common_params(params);
llama_model * model = llama_init.model; llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context; llama_context * ctx = llama_init.context;
@ -625,7 +625,7 @@ int main(int argc, char ** argv) {
// print system information // print system information
{ {
LOG_INF("\n"); LOG_INF("\n");
LOG_INF("%s\n", gpt_params_get_system_info(params).c_str()); LOG_INF("%s\n", common_params_get_system_info(params).c_str());
} }
if (!compute_imatrix(ctx, params)) { if (!compute_imatrix(ctx, params)) {

44
examples/infill/infill.cpp
View file

@ -35,8 +35,8 @@
static llama_context ** g_ctx; static llama_context ** g_ctx;
static llama_model ** g_model; static llama_model ** g_model;
static gpt_sampler ** g_smpl; static common_sampler ** g_smpl;
static gpt_params * g_params; static common_params * g_params;
static std::vector<llama_token> * g_input_tokens; static std::vector<llama_token> * g_input_tokens;
static std::ostringstream * g_output_ss; static std::ostringstream * g_output_ss;
static std::vector<llama_token> * g_output_tokens; static std::vector<llama_token> * g_output_tokens;
@ -44,7 +44,7 @@ static std::vector<llama_token> * g_output_tokens;
static bool is_interacting = false; static bool is_interacting = false;
static void write_logfile( static void write_logfile(
const llama_context * ctx, const gpt_params & params, const llama_model * model, const llama_context * ctx, const common_params & params, const llama_model * model,
const std::vector<llama_token> & input_tokens, const std::string & output, const std::vector<llama_token> & input_tokens, const std::string & output,
const std::vector<llama_token> & output_tokens const std::vector<llama_token> & output_tokens
) { ) {
@ -95,12 +95,12 @@ static void sigint_handler(int signo) {
} else { } else {
console::cleanup(); console::cleanup();
LOG("\n"); LOG("\n");
gpt_perf_print(*g_ctx, *g_smpl); common_perf_print(*g_ctx, *g_smpl);
write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens); write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
// make sure all logs are flushed // make sure all logs are flushed
LOG("Interrupted by user\n"); LOG("Interrupted by user\n");
gpt_log_pause(gpt_log_main()); common_log_pause(common_log_main());
_exit(130); _exit(130);
} }
@ -109,14 +109,14 @@ static void sigint_handler(int signo) {
#endif #endif
int main(int argc, char ** argv) { int main(int argc, char ** argv) {
gpt_params params; common_params params;
g_params = &params; g_params = &params;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_INFILL)) { if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_INFILL)) {
return 1; return 1;
} }
gpt_init(); common_init();
auto & sparams = params.sparams; auto & sparams = params.sparams;
@ -166,7 +166,7 @@ int main(int argc, char ** argv) {
llama_model * model = nullptr; llama_model * model = nullptr;
llama_context * ctx = nullptr; llama_context * ctx = nullptr;
gpt_sampler * smpl = nullptr; common_sampler * smpl = nullptr;
g_model = &model; g_model = &model;
g_ctx = &ctx; g_ctx = &ctx;
@ -174,7 +174,7 @@ int main(int argc, char ** argv) {
// load the model and apply lora adapter, if any // load the model and apply lora adapter, if any
LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__); LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__);
common_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_common_params(params);
model = llama_init.model; model = llama_init.model;
ctx = llama_init.context; ctx = llama_init.context;
@ -195,7 +195,7 @@ int main(int argc, char ** argv) {
// print system information // print system information
{ {
LOG_INF("\n"); LOG_INF("\n");
LOG_INF("%s\n", gpt_params_get_system_info(params).c_str()); LOG_INF("%s\n", common_params_get_system_info(params).c_str());
} }
const bool add_bos = llama_add_bos_token(model); const bool add_bos = llama_add_bos_token(model);
GGML_ASSERT(!llama_add_eos_token(model)); GGML_ASSERT(!llama_add_eos_token(model));
@ -298,11 +298,11 @@ int main(int argc, char ** argv) {
LOG_INF("Input suffix: '%s'\n", params.input_suffix.c_str()); LOG_INF("Input suffix: '%s'\n", params.input_suffix.c_str());
} }
} }
smpl = gpt_sampler_init(model, sparams); smpl = common_sampler_init(model, sparams);
LOG_INF("sampler seed: %u\n", gpt_sampler_get_seed(smpl)); LOG_INF("sampler seed: %u\n", common_sampler_get_seed(smpl));
LOG_INF("sampler params: \n%s\n", sparams.print().c_str()); LOG_INF("sampler params: \n%s\n", sparams.print().c_str());
LOG_INF("sampler chain: %s\n", gpt_sampler_print(smpl).c_str()); LOG_INF("sampler chain: %s\n", common_sampler_print(smpl).c_str());
LOG_INF("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep); LOG_INF("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
@ -411,9 +411,9 @@ int main(int argc, char ** argv) {
embd.clear(); embd.clear();
if ((int) embd_inp.size() <= n_consumed && !is_interacting) { if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
const llama_token id = gpt_sampler_sample(smpl, ctx, -1); const llama_token id = common_sampler_sample(smpl, ctx, -1);
gpt_sampler_accept(smpl, id, true); common_sampler_accept(smpl, id, true);
// LOG_DBG("last: %s\n", string_from(ctx, smpl->prev.to_vector()).c_str()); // LOG_DBG("last: %s\n", string_from(ctx, smpl->prev.to_vector()).c_str());
@ -434,7 +434,7 @@ int main(int argc, char ** argv) {
// push the prompt in the sampling context in order to apply repetition penalties later // push the prompt in the sampling context in order to apply repetition penalties later
// for the prompt, we don't apply grammar rules // for the prompt, we don't apply grammar rules
gpt_sampler_accept(smpl, embd_inp[n_consumed], false); common_sampler_accept(smpl, embd_inp[n_consumed], false);
++n_consumed; ++n_consumed;
if ((int) embd.size() >= params.n_batch) { if ((int) embd.size() >= params.n_batch) {
@ -465,7 +465,7 @@ int main(int argc, char ** argv) {
// if not currently processing queued inputs; // if not currently processing queued inputs;
if ((int) embd_inp.size() <= n_consumed) { if ((int) embd_inp.size() <= n_consumed) {
// deal with eot token in infill mode // deal with eot token in infill mode
if ((gpt_sampler_last(smpl) == llama_token_eot(model) || is_interacting) && params.interactive){ if ((common_sampler_last(smpl) == llama_token_eot(model) || is_interacting) && params.interactive){
if (is_interacting && !params.interactive_first) { if (is_interacting && !params.interactive_first) {
// print an eot token // print an eot token
LOG("%s", common_token_to_piece(ctx, llama_token_eot(model)).c_str()); LOG("%s", common_token_to_piece(ctx, llama_token_eot(model)).c_str());
@ -529,7 +529,7 @@ int main(int argc, char ** argv) {
is_interacting = false; is_interacting = false;
} }
// deal with end of generation tokens in interactive mode // deal with end of generation tokens in interactive mode
else if (llama_token_is_eog(model, gpt_sampler_last(smpl))) { else if (llama_token_is_eog(model, common_sampler_last(smpl))) {
LOG_DBG("found EOS token\n"); LOG_DBG("found EOS token\n");
if (params.interactive) { if (params.interactive) {
@ -601,7 +601,7 @@ int main(int argc, char ** argv) {
if (n_past > 0) { if (n_past > 0) {
if (is_interacting) { if (is_interacting) {
gpt_sampler_reset(smpl); common_sampler_reset(smpl);
} }
is_interacting = false; is_interacting = false;
} }
@ -624,13 +624,13 @@ int main(int argc, char ** argv) {
} }
LOG("\n"); LOG("\n");
gpt_perf_print(ctx, smpl); common_perf_print(ctx, smpl);
write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens); write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
llama_free(ctx); llama_free(ctx);
llama_free_model(model); llama_free_model(model);
gpt_sampler_free(smpl); common_sampler_free(smpl);
llama_backend_free(); llama_backend_free();
return 0; return 0;

28
examples/llava/llava-cli.cpp
View file

@ -42,11 +42,11 @@ static bool eval_string(struct llama_context * ctx_llama, const char* str, int n
return true; return true;
} }
static const char * sample(struct gpt_sampler * smpl, static const char * sample(struct common_sampler * smpl,
struct llama_context * ctx_llama, struct llama_context * ctx_llama,
int * n_past) { int * n_past) {
const llama_token id = gpt_sampler_sample(smpl, ctx_llama, -1); const llama_token id = common_sampler_sample(smpl, ctx_llama, -1);
gpt_sampler_accept(smpl, id, true); common_sampler_accept(smpl, id, true);
static std::string ret; static std::string ret;
if (llama_token_is_eog(llama_get_model(ctx_llama), id)) { if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
ret = "</s>"; ret = "</s>";
@ -120,7 +120,7 @@ static void print_usage(int, char ** argv) {
LOG("\n note: a lower temperature value like 0.1 is recommended for better quality.\n"); LOG("\n note: a lower temperature value like 0.1 is recommended for better quality.\n");
} }
static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_params * params, const std::string & fname) { static struct llava_image_embed * load_image(llava_context * ctx_llava, common_params * params, const std::string & fname) {
// load and preprocess the image // load and preprocess the image
llava_image_embed * embed = NULL; llava_image_embed * embed = NULL;
@ -146,7 +146,7 @@ static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_para
return embed; return embed;
} }
static void process_prompt(struct llava_context * ctx_llava, struct llava_image_embed * image_embed, gpt_params * params, const std::string & prompt) { static void process_prompt(struct llava_context * ctx_llava, struct llava_image_embed * image_embed, common_params * params, const std::string & prompt) {
int n_past = 0; int n_past = 0;
const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict; const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict;
@ -191,7 +191,7 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
LOG("\n"); LOG("\n");
struct gpt_sampler * smpl = gpt_sampler_init(ctx_llava->model, params->sparams); struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sparams);
if (!smpl) { if (!smpl) {
LOG_ERR("%s: failed to initialize sampling subsystem\n", __func__); LOG_ERR("%s: failed to initialize sampling subsystem\n", __func__);
exit(1); exit(1);
@ -211,15 +211,15 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
fflush(stdout); fflush(stdout);
} }
gpt_sampler_free(smpl); common_sampler_free(smpl);
LOG("\n"); LOG("\n");
} }
static struct llama_model * llava_init(gpt_params * params) { static struct llama_model * llava_init(common_params * params) {
llama_backend_init(); llama_backend_init();
llama_numa_init(params->numa); llama_numa_init(params->numa);
llama_model_params model_params = common_model_params_from_gpt_params(*params); llama_model_params model_params = common_model_params_from_common_params(*params);
llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params); llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
if (model == NULL) { if (model == NULL) {
@ -229,7 +229,7 @@ static struct llama_model * llava_init(gpt_params * params) {
return model; return model;
} }
static struct llava_context * llava_init_context(gpt_params * params, llama_model * model) { static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
const char * clip_path = params->mmproj.c_str(); const char * clip_path = params->mmproj.c_str();
auto prompt = params->prompt; auto prompt = params->prompt;
@ -240,7 +240,7 @@ static struct llava_context * llava_init_context(gpt_params * params, llama_mode
auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1); auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
llama_context_params ctx_params = common_context_params_from_gpt_params(*params); llama_context_params ctx_params = common_context_params_from_common_params(*params);
ctx_params.n_ctx = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings ctx_params.n_ctx = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings
llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params); llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
@ -272,13 +272,13 @@ static void llava_free(struct llava_context * ctx_llava) {
int main(int argc, char ** argv) { int main(int argc, char ** argv) {
ggml_time_init(); ggml_time_init();
gpt_params params; common_params params;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, print_usage)) { if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, print_usage)) {
return 1; return 1;
} }
gpt_init(); common_init();
if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) { if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
print_usage(argc, argv); print_usage(argc, argv);

36
examples/llava/minicpmv-cli.cpp
View file

@ -25,11 +25,11 @@ static void show_additional_info(int /*argc*/, char ** argv) {
LOG("\nnote: a lower temperature value like 0.1 is recommended for better quality.\n"); LOG("\nnote: a lower temperature value like 0.1 is recommended for better quality.\n");
} }
static struct llama_model * llava_init(gpt_params * params) { static struct llama_model * llava_init(common_params * params) {
llama_backend_init(); llama_backend_init();
llama_numa_init(params->numa); llama_numa_init(params->numa);
llama_model_params model_params = common_model_params_from_gpt_params(*params); llama_model_params model_params = common_model_params_from_common_params(*params);
llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params); llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
if (model == NULL) { if (model == NULL) {
@ -39,13 +39,13 @@ static struct llama_model * llava_init(gpt_params * params) {
return model; return model;
} }
static struct llava_context * llava_init_context(gpt_params * params, llama_model * model) { static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
auto prompt = params->prompt; auto prompt = params->prompt;
if (prompt.empty()) { if (prompt.empty()) {
prompt = "describe the image in detail."; prompt = "describe the image in detail.";
} }
llama_context_params ctx_params = common_context_params_from_gpt_params(*params); llama_context_params ctx_params = common_context_params_from_common_params(*params);
if (params->n_ctx < 2048) { if (params->n_ctx < 2048) {
// warn user here, "Image processing requires at least 2048 context, setting context to 2048" // warn user here, "Image processing requires at least 2048 context, setting context to 2048"
LOG_WRN("%s: Image processing requires at least 2048 context, setting context to 2048\n" , __func__); LOG_WRN("%s: Image processing requires at least 2048 context, setting context to 2048\n" , __func__);
@ -79,7 +79,7 @@ static void llava_free(struct llava_context * ctx_llava) {
llama_backend_free(); llama_backend_free();
} }
static struct clip_ctx * clip_init_context(gpt_params * params) { static struct clip_ctx * clip_init_context(common_params * params) {
const char * clip_path = params->mmproj.c_str(); const char * clip_path = params->mmproj.c_str();
auto prompt = params->prompt; auto prompt = params->prompt;
@ -129,7 +129,7 @@ static void process_eval_image_embed(struct llava_context * ctx_llava, const str
llava_image_embed_free(slice_embed); llava_image_embed_free(slice_embed);
} }
static void process_image(struct llava_context * ctx_llava, struct llava_image_embed * embeds, gpt_params * params, int &n_past) { static void process_image(struct llava_context * ctx_llava, struct llava_image_embed * embeds, common_params * params, int &n_past) {
std::string system_prompt; std::string system_prompt;
int idx = 0; int idx = 0;
int num_image_embeds = embeds->n_image_pos / clip_n_patches(ctx_llava->ctx_clip); int num_image_embeds = embeds->n_image_pos / clip_n_patches(ctx_llava->ctx_clip);
@ -162,11 +162,11 @@ static void process_image(struct llava_context * ctx_llava, struct llava_image_e
LOG_INF("%s: image token past: %d\n", __func__, n_past); LOG_INF("%s: image token past: %d\n", __func__, n_past);
} }
static const char * sample(struct gpt_sampler * smpl, static const char * sample(struct common_sampler * smpl,
struct llama_context * ctx_llama, struct llama_context * ctx_llama,
int * n_past) { int * n_past) {
const llama_token id = gpt_sampler_sample(smpl, ctx_llama, -1); const llama_token id = common_sampler_sample(smpl, ctx_llama, -1);
gpt_sampler_accept(smpl, id, true); common_sampler_accept(smpl, id, true);
static std::string ret; static std::string ret;
if (llama_token_is_eog(llama_get_model(ctx_llama), id)) { if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
ret = "</s>"; ret = "</s>";
@ -177,7 +177,7 @@ static const char * sample(struct gpt_sampler * smpl,
return ret.c_str(); return ret.c_str();
} }
static struct llava_context * minicpmv_init(gpt_params * params, const std::string & fname, int &n_past){ static struct llava_context * minicpmv_init(common_params * params, const std::string & fname, int &n_past){
auto * ctx_clip = clip_init_context(params); auto * ctx_clip = clip_init_context(params);
auto * embeds = llava_image_embed_make_with_filename(ctx_clip, params->cpuparams.n_threads, fname.c_str()); auto * embeds = llava_image_embed_make_with_filename(ctx_clip, params->cpuparams.n_threads, fname.c_str());
if (!embeds) { if (!embeds) {
@ -213,7 +213,7 @@ static struct llava_context * minicpmv_init(gpt_params * params, const std::stri
return ctx_llava; return ctx_llava;
} }
static struct gpt_sampler * llama_init(struct llava_context * ctx_llava, gpt_params * params, const std::string & prompt, int & n_past, bool is_first = false){ static struct common_sampler * llama_init(struct llava_context * ctx_llava, common_params * params, const std::string & prompt, int & n_past, bool is_first = false){
std::string user_prompt = prompt; std::string user_prompt = prompt;
int has_minicpmv_projector = clip_is_minicpmv(ctx_llava->ctx_clip); int has_minicpmv_projector = clip_is_minicpmv(ctx_llava->ctx_clip);
if (!is_first) { if (!is_first) {
@ -237,11 +237,11 @@ static struct gpt_sampler * llama_init(struct llava_context * ctx_llava, gpt_par
LOG_INF("\n"); LOG_INF("\n");
struct gpt_sampler * smpl = gpt_sampler_init(ctx_llava->model, params->sparams); struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sparams);
return smpl; return smpl;
} }
static const char * llama_loop(struct llava_context * ctx_llava,struct gpt_sampler * smpl, int &n_past){ static const char * llama_loop(struct llava_context * ctx_llava,struct common_sampler * smpl, int &n_past){
const char * tmp = sample(smpl, ctx_llava->ctx_llama, &n_past); const char * tmp = sample(smpl, ctx_llava->ctx_llama, &n_past);
return tmp; return tmp;
@ -250,13 +250,13 @@ static const char * llama_loop(struct llava_context * ctx_llava,struct gpt_sampl
int main(int argc, char ** argv) { int main(int argc, char ** argv) {
ggml_time_init(); ggml_time_init();
gpt_params params; common_params params;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, show_additional_info)) { if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, show_additional_info)) {
return 1; return 1;
} }
gpt_init(); common_init();
if (params.mmproj.empty() || (params.image.empty())) { if (params.mmproj.empty() || (params.image.empty())) {
show_additional_info(argc, argv); show_additional_info(argc, argv);
@ -290,7 +290,7 @@ int main(int argc, char ** argv) {
fflush(stdout); fflush(stdout);
} }
gpt_sampler_free(smpl); common_sampler_free(smpl);
}else { }else {
while (true) { while (true) {
LOG("<user>"); LOG("<user>");
@ -309,7 +309,7 @@ int main(int argc, char ** argv) {
if (strstr(response.c_str(), "<user>")) break; // minicpm-v if (strstr(response.c_str(), "<user>")) break; // minicpm-v
fflush(stdout); fflush(stdout);
} }
gpt_sampler_free(smpl); common_sampler_free(smpl);
} }
} }
printf("\n"); printf("\n");

24
examples/lookahead/lookahead.cpp
View file

@ -37,13 +37,13 @@ struct ngram_container {
}; };
int main(int argc, char ** argv) { int main(int argc, char ** argv) {
gpt_params params; common_params params;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) { if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
return 1; return 1;
} }
gpt_init(); common_init();
const int W = 15; // lookahead window const int W = 15; // lookahead window
const int N = 5; // n-gram size const int N = 5; // n-gram size
@ -56,7 +56,7 @@ int main(int argc, char ** argv) {
llama_numa_init(params.numa); llama_numa_init(params.numa);
// load the target model // load the target model
common_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_common_params(params);
llama_model * model = llama_init.model; llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context; llama_context * ctx = llama_init.context;
@ -115,7 +115,7 @@ int main(int argc, char ** argv) {
llama_batch batch = llama_batch_init(params.n_ctx, 0, W + G + 1); llama_batch batch = llama_batch_init(params.n_ctx, 0, W + G + 1);
// target model sampling context // target model sampling context
struct gpt_sampler * smpl = gpt_sampler_init(model, params.sparams); struct common_sampler * smpl = common_sampler_init(model, params.sparams);
// verification n-grams // verification n-grams
std::vector<ngram_data> ngrams_cur(G); std::vector<ngram_data> ngrams_cur(G);
@ -156,9 +156,9 @@ int main(int argc, char ** argv) {
// sample first token // sample first token
{ {
id = gpt_sampler_sample(smpl, ctx, 0); id = common_sampler_sample(smpl, ctx, 0);
gpt_sampler_accept(smpl, id, true); common_sampler_accept(smpl, id, true);
{ {
const std::string token_str = common_token_to_piece(ctx, id); const std::string token_str = common_token_to_piece(ctx, id);
@ -281,9 +281,9 @@ int main(int argc, char ** argv) {
} }
// sample the next token // sample the next token
id = gpt_sampler_sample(smpl, ctx, i_batch); id = common_sampler_sample(smpl, ctx, i_batch);
gpt_sampler_accept(smpl, id, true); common_sampler_accept(smpl, id, true);
// print // print
{ {
@ -358,7 +358,7 @@ int main(int argc, char ** argv) {
if (v == 0) { if (v == 0) {
// sample from the last level // sample from the last level
for (int i = 0; i < W; i++) { for (int i = 0; i < W; i++) {
tokens_j[N - 2][i] = gpt_sampler_sample(smpl, ctx, ngrams_cur.size()*(N-1) + W*(N - 2) + i); tokens_j[N - 2][i] = common_sampler_sample(smpl, ctx, ngrams_cur.size()*(N-1) + W*(N - 2) + i);
} }
} else { } else {
for (int i = 0; i < W; i++) { for (int i = 0; i < W; i++) {
@ -466,9 +466,9 @@ int main(int argc, char ** argv) {
LOG_INF("n_accept = %d\n", n_accept); LOG_INF("n_accept = %d\n", n_accept);
LOG_INF("\n"); LOG_INF("\n");
gpt_perf_print(ctx, smpl); common_perf_print(ctx, smpl);
gpt_sampler_free(smpl); common_sampler_free(smpl);
llama_kv_cache_view_free(&kvc_view); llama_kv_cache_view_free(&kvc_view);

6
examples/lookup/lookup-create.cpp
View file

@ -12,9 +12,9 @@
#include <vector> #include <vector>
int main(int argc, char ** argv){ int main(int argc, char ** argv){
gpt_params params; common_params params;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) { if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {
return 1; return 1;
} }
@ -23,7 +23,7 @@ int main(int argc, char ** argv){
llama_numa_init(params.numa); llama_numa_init(params.numa);
// load the model // load the model
common_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_common_params(params);
llama_model * model = llama_init.model; llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context; llama_context * ctx = llama_init.context;

8
examples/lookup/lookup-stats.cpp
View file

@ -13,13 +13,13 @@
#include <vector> #include <vector>
int main(int argc, char ** argv){ int main(int argc, char ** argv){
gpt_params params; common_params params;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) { if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {
return 1; return 1;
} }
gpt_init(); common_init();
const int n_draft = params.n_draft; const int n_draft = params.n_draft;
@ -28,7 +28,7 @@ int main(int argc, char ** argv){
llama_numa_init(params.numa); llama_numa_init(params.numa);
// load the model // load the model
common_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_common_params(params);
llama_model * model = llama_init.model; llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context; llama_context * ctx = llama_init.context;

18
examples/lookup/lookup.cpp
View file

@ -13,13 +13,13 @@
#include <vector> #include <vector>
int main(int argc, char ** argv){ int main(int argc, char ** argv){
gpt_params params; common_params params;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) { if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {
return 1; return 1;
} }
gpt_init(); common_init();
// max. number of additional tokens to draft if match is found // max. number of additional tokens to draft if match is found
const int n_draft = params.n_draft; const int n_draft = params.n_draft;
@ -31,7 +31,7 @@ int main(int argc, char ** argv){
llama_numa_init(params.numa); llama_numa_init(params.numa);
// load the model // load the model
common_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_common_params(params);
llama_model * model = llama_init.model; llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context; llama_context * ctx = llama_init.context;
@ -102,7 +102,7 @@ int main(int argc, char ** argv){
bool has_eos = false; bool has_eos = false;
struct gpt_sampler * smpl = gpt_sampler_init(model, params.sparams); struct common_sampler * smpl = common_sampler_init(model, params.sparams);
std::vector<llama_token> draft; std::vector<llama_token> draft;
@ -126,9 +126,9 @@ int main(int argc, char ** argv){
int i_dft = 0; int i_dft = 0;
while (true) { while (true) {
// sample from the target model // sample from the target model
llama_token id = gpt_sampler_sample(smpl, ctx, i_dft); llama_token id = common_sampler_sample(smpl, ctx, i_dft);
gpt_sampler_accept(smpl, id, true); common_sampler_accept(smpl, id, true);
const std::string token_str = common_token_to_piece(ctx, id); const std::string token_str = common_token_to_piece(ctx, id);
@ -237,9 +237,9 @@ int main(int argc, char ** argv){
LOG_INF("accept = %.3f%%\n", 100.0f * n_accept / n_drafted); LOG_INF("accept = %.3f%%\n", 100.0f * n_accept / n_drafted);
LOG_INF("\ntarget:\n\n"); LOG_INF("\ntarget:\n\n");
gpt_perf_print(ctx, smpl); common_perf_print(ctx, smpl);
gpt_sampler_free(smpl); common_sampler_free(smpl);
llama_batch_free(batch_tgt); llama_batch_free(batch_tgt);

48
examples/main/main.cpp
View file

@ -33,8 +33,8 @@
static llama_context ** g_ctx; static llama_context ** g_ctx;
static llama_model ** g_model; static llama_model ** g_model;
static gpt_sampler ** g_smpl; static common_sampler ** g_smpl;
static gpt_params * g_params; static common_params * g_params;
static std::vector<llama_token> * g_input_tokens; static std::vector<llama_token> * g_input_tokens;
static std::ostringstream * g_output_ss; static std::ostringstream * g_output_ss;
static std::vector<llama_token> * g_output_tokens; static std::vector<llama_token> * g_output_tokens;
@ -63,7 +63,7 @@ static bool file_is_empty(const std::string & path) {
} }
static void write_logfile( static void write_logfile(
const llama_context * ctx, const gpt_params & params, const llama_model * model, const llama_context * ctx, const common_params & params, const llama_model * model,
const std::vector<llama_token> & input_tokens, const std::string & output, const std::vector<llama_token> & input_tokens, const std::string & output,
const std::vector<llama_token> & output_tokens const std::vector<llama_token> & output_tokens
) { ) {
@ -114,12 +114,12 @@ static void sigint_handler(int signo) {
} else { } else {
console::cleanup(); console::cleanup();
LOG("\n"); LOG("\n");
gpt_perf_print(*g_ctx, *g_smpl); common_perf_print(*g_ctx, *g_smpl);
write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens); write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
// make sure all logs are flushed // make sure all logs are flushed
LOG("Interrupted by user\n"); LOG("Interrupted by user\n");
gpt_log_pause(gpt_log_main()); common_log_pause(common_log_main());
_exit(130); _exit(130);
} }
@ -136,13 +136,13 @@ static std::string chat_add_and_format(struct llama_model * model, std::vector<c
} }
int main(int argc, char ** argv) { int main(int argc, char ** argv) {
gpt_params params; common_params params;
g_params = &params; g_params = &params;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_MAIN, print_usage)) { if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_MAIN, print_usage)) {
return 1; return 1;
} }
gpt_init(); common_init();
auto & sparams = params.sparams; auto & sparams = params.sparams;
@ -187,7 +187,7 @@ int main(int argc, char ** argv) {
llama_model * model = nullptr; llama_model * model = nullptr;
llama_context * ctx = nullptr; llama_context * ctx = nullptr;
gpt_sampler * smpl = nullptr; common_sampler * smpl = nullptr;
std::vector<common_chat_msg> chat_msgs; std::vector<common_chat_msg> chat_msgs;
@ -197,7 +197,7 @@ int main(int argc, char ** argv) {
// load the model and apply lora adapter, if any // load the model and apply lora adapter, if any
LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__); LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__);
common_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_common_params(params);
model = llama_init.model; model = llama_init.model;
ctx = llama_init.context; ctx = llama_init.context;
@ -255,7 +255,7 @@ int main(int argc, char ** argv) {
// print system information // print system information
{ {
LOG_INF("\n"); LOG_INF("\n");
LOG_INF("%s\n", gpt_params_get_system_info(params).c_str()); LOG_INF("%s\n", common_params_get_system_info(params).c_str());
LOG_INF("\n"); LOG_INF("\n");
} }
@ -448,15 +448,15 @@ int main(int argc, char ** argv) {
} }
} }
smpl = gpt_sampler_init(model, sparams); smpl = common_sampler_init(model, sparams);
if (!smpl) { if (!smpl) {
LOG_ERR("%s: failed to initialize sampling subsystem\n", __func__); LOG_ERR("%s: failed to initialize sampling subsystem\n", __func__);
return 1; return 1;
} }
LOG_INF("sampler seed: %u\n", gpt_sampler_get_seed(smpl)); LOG_INF("sampler seed: %u\n", common_sampler_get_seed(smpl));
LOG_INF("sampler params: \n%s\n", sparams.print().c_str()); LOG_INF("sampler params: \n%s\n", sparams.print().c_str());
LOG_INF("sampler chain: %s\n", gpt_sampler_print(smpl).c_str()); LOG_INF("sampler chain: %s\n", common_sampler_print(smpl).c_str());
LOG_INF("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep); LOG_INF("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
@ -679,9 +679,9 @@ int main(int argc, char ** argv) {
LOG_DBG("saved session to %s\n", path_session.c_str()); LOG_DBG("saved session to %s\n", path_session.c_str());
} }
const llama_token id = gpt_sampler_sample(smpl, ctx, -1); const llama_token id = common_sampler_sample(smpl, ctx, -1);
gpt_sampler_accept(smpl, id, /* accept_grammar= */ true); common_sampler_accept(smpl, id, /* accept_grammar= */ true);
// LOG_DBG("last: %s\n", string_from(ctx, smpl->prev.to_vector()).c_str()); // LOG_DBG("last: %s\n", string_from(ctx, smpl->prev.to_vector()).c_str());
@ -702,7 +702,7 @@ int main(int argc, char ** argv) {
// push the prompt in the sampling context in order to apply repetition penalties later // push the prompt in the sampling context in order to apply repetition penalties later
// for the prompt, we don't apply grammar rules // for the prompt, we don't apply grammar rules
gpt_sampler_accept(smpl, embd_inp[n_consumed], /* accept_grammar= */ false); common_sampler_accept(smpl, embd_inp[n_consumed], /* accept_grammar= */ false);
++n_consumed; ++n_consumed;
if ((int) embd.size() >= params.n_batch) { if ((int) embd.size() >= params.n_batch) {
@ -743,7 +743,7 @@ int main(int argc, char ** argv) {
// check for reverse prompt in the last n_prev tokens // check for reverse prompt in the last n_prev tokens
if (!params.antiprompt.empty()) { if (!params.antiprompt.empty()) {
const int n_prev = 32; const int n_prev = 32;
const std::string last_output = gpt_sampler_prev_str(smpl, ctx, n_prev); const std::string last_output = common_sampler_prev_str(smpl, ctx, n_prev);
is_antiprompt = false; is_antiprompt = false;
// Check if each of the reverse prompts appears at the end of the output. // Check if each of the reverse prompts appears at the end of the output.
@ -765,7 +765,7 @@ int main(int argc, char ** argv) {
} }
// check for reverse prompt using special tokens // check for reverse prompt using special tokens
llama_token last_token = gpt_sampler_last(smpl); llama_token last_token = common_sampler_last(smpl);
for (std::vector<llama_token> ids : antiprompt_ids) { for (std::vector<llama_token> ids : antiprompt_ids) {
if (ids.size() == 1 && last_token == ids[0]) { if (ids.size() == 1 && last_token == ids[0]) {
if (params.interactive) { if (params.interactive) {
@ -782,7 +782,7 @@ int main(int argc, char ** argv) {
} }
// deal with end of generation tokens in interactive mode // deal with end of generation tokens in interactive mode
if (llama_token_is_eog(model, gpt_sampler_last(smpl))) { if (llama_token_is_eog(model, common_sampler_last(smpl))) {
LOG_DBG("found an EOG token\n"); LOG_DBG("found an EOG token\n");
if (params.interactive) { if (params.interactive) {
@ -803,7 +803,7 @@ int main(int argc, char ** argv) {
// if current token is not EOG, we add it to current assistant message // if current token is not EOG, we add it to current assistant message
if (params.conversation) { if (params.conversation) {
const auto id = gpt_sampler_last(smpl); const auto id = common_sampler_last(smpl);
assistant_ss << common_token_to_piece(ctx, id, false); assistant_ss << common_token_to_piece(ctx, id, false);
} }
@ -899,7 +899,7 @@ int main(int argc, char ** argv) {
if (n_past > 0) { if (n_past > 0) {
if (is_interacting) { if (is_interacting) {
gpt_sampler_reset(smpl); common_sampler_reset(smpl);
} }
is_interacting = false; is_interacting = false;
} }
@ -925,10 +925,10 @@ int main(int argc, char ** argv) {
} }
LOG("\n\n"); LOG("\n\n");
gpt_perf_print(ctx, smpl); common_perf_print(ctx, smpl);
write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens); write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
gpt_sampler_free(smpl); common_sampler_free(smpl);
llama_free(ctx); llama_free(ctx);
llama_free_model(model); llama_free_model(model);

20
examples/parallel/parallel.cpp
View file

@ -54,7 +54,7 @@ static std::vector<std::string> k_prompts = {
struct client { struct client {
~client() { ~client() {
if (smpl) { if (smpl) {
gpt_sampler_free(smpl); common_sampler_free(smpl);
} }
} }
@ -75,7 +75,7 @@ struct client {
std::string prompt; std::string prompt;
std::string response; std::string response;
struct gpt_sampler * smpl = nullptr; struct common_sampler * smpl = nullptr;
}; };
static void print_date_time() { static void print_date_time() {
@ -103,13 +103,13 @@ static std::vector<std::string> split_string(const std::string& input, char deli
int main(int argc, char ** argv) { int main(int argc, char ** argv) {
srand(1234); srand(1234);
gpt_params params; common_params params;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_PARALLEL)) { if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PARALLEL)) {
return 1; return 1;
} }
gpt_init(); common_init();
// number of simultaneous "clients" to simulate // number of simultaneous "clients" to simulate
const int32_t n_clients = params.n_parallel; const int32_t n_clients = params.n_parallel;
@ -130,7 +130,7 @@ int main(int argc, char ** argv) {
llama_numa_init(params.numa); llama_numa_init(params.numa);
// load the target model // load the target model
common_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_common_params(params);
llama_model * model = llama_init.model; llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context; llama_context * ctx = llama_init.context;
@ -160,7 +160,7 @@ int main(int argc, char ** argv) {
for (size_t i = 0; i < clients.size(); ++i) { for (size_t i = 0; i < clients.size(); ++i) {
auto & client = clients[i]; auto & client = clients[i];
client.id = i; client.id = i;
client.smpl = gpt_sampler_init(model, params.sparams); client.smpl = common_sampler_init(model, params.sparams);
} }
std::vector<llama_token> tokens_system; std::vector<llama_token> tokens_system;
@ -252,7 +252,7 @@ int main(int argc, char ** argv) {
client.prompt = client.input + "\nAssistant:"; client.prompt = client.input + "\nAssistant:";
client.response = ""; client.response = "";
gpt_sampler_reset(client.smpl); common_sampler_reset(client.smpl);
// do not prepend BOS because we have a system prompt! // do not prepend BOS because we have a system prompt!
std::vector<llama_token> tokens_prompt; std::vector<llama_token> tokens_prompt;
@ -340,9 +340,9 @@ int main(int argc, char ** argv) {
//printf("client %d, seq %d, token %d, pos %d, batch %d\n", //printf("client %d, seq %d, token %d, pos %d, batch %d\n",
// client.id, client.seq_id, client.sampled, client.n_decoded, client.i_batch); // client.id, client.seq_id, client.sampled, client.n_decoded, client.i_batch);
const llama_token id = gpt_sampler_sample(client.smpl, ctx, client.i_batch - i); const llama_token id = common_sampler_sample(client.smpl, ctx, client.i_batch - i);
gpt_sampler_accept(client.smpl, id, true); common_sampler_accept(client.smpl, id, true);
if (client.n_decoded == 1) { if (client.n_decoded == 1) {
// start measuring generation time after the first token to make sure all concurrent clients // start measuring generation time after the first token to make sure all concurrent clients

10
examples/passkey/passkey.cpp
View file

@ -15,17 +15,17 @@ static void print_usage(int, char ** argv) {
} }
int main(int argc, char ** argv) { int main(int argc, char ** argv) {
gpt_params params; common_params params;
params.n_junk = 250; params.n_junk = 250;
params.n_keep = 32; params.n_keep = 32;
params.i_pos = -1; params.i_pos = -1;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_PASSKEY, print_usage)) { if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PASSKEY, print_usage)) {
return 1; return 1;
} }
gpt_init(); common_init();
int n_junk = params.n_junk; int n_junk = params.n_junk;
int n_keep = params.n_keep; int n_keep = params.n_keep;
@ -61,7 +61,7 @@ int main(int argc, char ** argv) {
// initialize the model // initialize the model
llama_model_params model_params = common_model_params_from_gpt_params(params); llama_model_params model_params = common_model_params_from_common_params(params);
llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params); llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
@ -72,7 +72,7 @@ int main(int argc, char ** argv) {
// initialize the context // initialize the context
llama_context_params ctx_params = common_context_params_from_gpt_params(params); llama_context_params ctx_params = common_context_params_from_common_params(params);
ctx_params.n_ctx = llama_n_ctx_train(model)*n_grp + n_keep; ctx_params.n_ctx = llama_n_ctx_train(model)*n_grp + n_keep;

24
examples/perplexity/perplexity.cpp
View file

@ -35,7 +35,7 @@ struct results_log_softmax {
}; };
static void write_logfile( static void write_logfile(
const llama_context * ctx, const gpt_params & params, const llama_model * model, const llama_context * ctx, const common_params & params, const llama_model * model,
const struct results_perplexity & results const struct results_perplexity & results
) { ) {
if (params.logdir.empty()) { if (params.logdir.empty()) {
@ -337,7 +337,7 @@ static void process_logits(int n_vocab, const float * logits, const int * tokens
} }
} }
static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params) { static results_perplexity perplexity_v2(llama_context * ctx, const common_params & params) {
// Download: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip // Download: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
// Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw` // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
// Output: `perplexity: 13.5106 [114/114]` // Output: `perplexity: 13.5106 [114/114]`
@ -472,7 +472,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
return {tokens, std::exp(nll / count), logit_history, prob_history}; return {tokens, std::exp(nll / count), logit_history, prob_history};
} }
static results_perplexity perplexity(llama_context * ctx, const gpt_params & params, const int32_t n_ctx) { static results_perplexity perplexity(llama_context * ctx, const common_params & params, const int32_t n_ctx) {
if (params.ppl_stride > 0) { if (params.ppl_stride > 0) {
return perplexity_v2(ctx, params); return perplexity_v2(ctx, params);
} }
@ -763,7 +763,7 @@ static void compute_logprobs(const float * batch_logits, int n_vocab, std::vecto
} }
} }
static void hellaswag_score(llama_context * ctx, const gpt_params & params) { static void hellaswag_score(llama_context * ctx, const common_params & params) {
// Calculates hellaswag score (acc_norm) from prompt // Calculates hellaswag score (acc_norm) from prompt
// //
// Data extracted from the HellaSwag validation dataset (MIT license) https://github.com/rowanz/hellaswag/blob/master/data/hellaswag_val.jsonl // Data extracted from the HellaSwag validation dataset (MIT license) https://github.com/rowanz/hellaswag/blob/master/data/hellaswag_val.jsonl
@ -1102,7 +1102,7 @@ static std::vector<winogrande_entry> load_winogrande_from_csv(const std::string
* 0,Sarah was a much better surgeon than Maria so _ always got the easier cases.,Sarah,Maria,2 * 0,Sarah was a much better surgeon than Maria so _ always got the easier cases.,Sarah,Maria,2
* *
*/ */
static void winogrande_score(llama_context * ctx, const gpt_params & params) { static void winogrande_score(llama_context * ctx, const common_params & params) {
constexpr int k_min_trailing_ctx = 3; constexpr int k_min_trailing_ctx = 3;
@ -1403,7 +1403,7 @@ static bool multiple_choice_prepare_one_task(llama_context * ctx, multiple_choic
// git@hf.co:datasets/Stevross/mmlu // git@hf.co:datasets/Stevross/mmlu
// https://huggingface.co/datasets/truthful_qa // https://huggingface.co/datasets/truthful_qa
// //
static void multiple_choice_score(llama_context * ctx, const gpt_params & params) { static void multiple_choice_score(llama_context * ctx, const common_params & params) {
std::istringstream strstream(params.prompt); std::istringstream strstream(params.prompt);
uint32_t n_task; uint32_t n_task;
@ -1683,7 +1683,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
LOG_INF("\n"); LOG_INF("\n");
} }
static void kl_divergence(llama_context * ctx, const gpt_params & params) { static void kl_divergence(llama_context * ctx, const common_params & params) {
if (params.logits_file.empty()) { if (params.logits_file.empty()) {
LOG_ERR("%s: you must provide a name of a file containing the log probabilities of the base model\n", __func__); LOG_ERR("%s: you must provide a name of a file containing the log probabilities of the base model\n", __func__);
return; return;
@ -1955,17 +1955,17 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
} }
int main(int argc, char ** argv) { int main(int argc, char ** argv) {
gpt_params params; common_params params;
params.n_ctx = 512; params.n_ctx = 512;
params.logits_all = true; params.logits_all = true;
params.escape = false; params.escape = false;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_PERPLEXITY)) { if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PERPLEXITY)) {
return 1; return 1;
} }
gpt_init(); common_init();
const int32_t n_ctx = params.n_ctx; const int32_t n_ctx = params.n_ctx;
@ -2004,7 +2004,7 @@ int main(int argc, char ** argv) {
llama_numa_init(params.numa); llama_numa_init(params.numa);
// load the model and apply lora adapter, if any // load the model and apply lora adapter, if any
common_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_common_params(params);
llama_model * model = llama_init.model; llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context; llama_context * ctx = llama_init.context;
@ -2023,7 +2023,7 @@ int main(int argc, char ** argv) {
// print system information // print system information
{ {
LOG_INF("\n"); LOG_INF("\n");
LOG_INF("%s\n", gpt_params_get_system_info(params).c_str()); LOG_INF("%s\n", common_params_get_system_info(params).c_str());
} }
struct results_perplexity results; struct results_perplexity results;

10
examples/retrieval/retrieval.cpp
View file

@ -112,13 +112,13 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
} }
int main(int argc, char ** argv) { int main(int argc, char ** argv) {
gpt_params params; common_params params;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_RETRIEVAL, print_usage)) { if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_RETRIEVAL, print_usage)) {
return 1; return 1;
} }
gpt_init(); common_init();
// For BERT models, batch size must be equal to ubatch size // For BERT models, batch size must be equal to ubatch size
params.n_ubatch = params.n_batch; params.n_ubatch = params.n_batch;
@ -149,7 +149,7 @@ int main(int argc, char ** argv) {
llama_numa_init(params.numa); llama_numa_init(params.numa);
// load the model // load the model
common_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_common_params(params);
llama_model * model = llama_init.model; llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context; llama_context * ctx = llama_init.context;
@ -176,7 +176,7 @@ int main(int argc, char ** argv) {
// print system information // print system information
{ {
LOG_INF("\n"); LOG_INF("\n");
LOG_INF("%s\n", gpt_params_get_system_info(params).c_str()); LOG_INF("%s\n", common_params_get_system_info(params).c_str());
} }
// max batch size // max batch size

10
examples/save-load-state/save-load-state.cpp
View file

@ -6,12 +6,12 @@
#include <cstdio> #include <cstdio>
int main(int argc, char ** argv) { int main(int argc, char ** argv) {
gpt_params params; common_params params;
params.prompt = "The quick brown fox"; params.prompt = "The quick brown fox";
params.sparams.seed = 1234; params.sparams.seed = 1234;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) { if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
return 1; return 1;
} }
@ -28,7 +28,7 @@ int main(int argc, char ** argv) {
std::string result2; std::string result2;
// init // init
common_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_common_params(params);
llama_model * model = llama_init.model; llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context; llama_context * ctx = llama_init.context;
@ -92,7 +92,7 @@ int main(int argc, char ** argv) {
llama_free(ctx); llama_free(ctx);
// make new context // make new context
auto * ctx2 = llama_new_context_with_model(model, common_context_params_from_gpt_params(params)); auto * ctx2 = llama_new_context_with_model(model, common_context_params_from_common_params(params));
llama_sampler * smpl2 = llama_sampler_chain_init(sparams); llama_sampler * smpl2 = llama_sampler_chain_init(sparams);
@ -152,7 +152,7 @@ int main(int argc, char ** argv) {
} }
// make new context // make new context
auto * ctx3 = llama_new_context_with_model(model, common_context_params_from_gpt_params(params)); auto * ctx3 = llama_new_context_with_model(model, common_context_params_from_common_params(params));
llama_sampler * smpl3 = llama_sampler_chain_init(sparams); llama_sampler * smpl3 = llama_sampler_chain_init(sparams);

44
examples/server/server.cpp
View file

@ -188,8 +188,8 @@ struct server_slot {
// sampling // sampling
json json_schema; json json_schema;
struct gpt_sampler_params sparams; struct common_sampler_params sparams;
struct gpt_sampler * smpl = nullptr; struct common_sampler * smpl = nullptr;
llama_token sampled; llama_token sampled;
@ -231,7 +231,7 @@ struct server_slot {
generated_token_probs.clear(); generated_token_probs.clear();
} }
bool has_budget(gpt_params &global_params) { bool has_budget(common_params &global_params) {
if (params.n_predict == -1 && global_params.n_predict == -1) { if (params.n_predict == -1 && global_params.n_predict == -1) {
return true; // limitless return true; // limitless
} }
@ -613,7 +613,7 @@ struct server_context {
llama_context * ctx = nullptr; llama_context * ctx = nullptr;
std::vector<common_lora_adapter_container> loras; std::vector<common_lora_adapter_container> loras;
gpt_params params; common_params params;
llama_batch batch = {}; llama_batch batch = {};
@ -655,20 +655,20 @@ struct server_context {
// Clear any sampling context // Clear any sampling context
for (server_slot & slot : slots) { for (server_slot & slot : slots) {
if (slot.smpl != nullptr) { if (slot.smpl != nullptr) {
gpt_sampler_free(slot.smpl); common_sampler_free(slot.smpl);
} }
} }
llama_batch_free(batch); llama_batch_free(batch);
} }
bool load_model(const gpt_params & params_) { bool load_model(const common_params & params_) {
params = params_; params = params_;
// dedicate one sequence to the system prompt // dedicate one sequence to the system prompt
params.n_parallel += 1; params.n_parallel += 1;
common_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_common_params(params);
model = llama_init.model; model = llama_init.model;
ctx = llama_init.context; ctx = llama_init.context;
@ -1031,7 +1031,7 @@ struct server_context {
sampler_names.emplace_back(name); sampler_names.emplace_back(name);
} }
} }
slot.sparams.samplers = gpt_sampler_types_from_names(sampler_names, false); slot.sparams.samplers = common_sampler_types_from_names(sampler_names, false);
} else { } else {
slot.sparams.samplers = default_sparams.samplers; slot.sparams.samplers = default_sparams.samplers;
} }
@ -1039,10 +1039,10 @@ struct server_context {
{ {
if (slot.smpl != nullptr) { if (slot.smpl != nullptr) {
gpt_sampler_free(slot.smpl); common_sampler_free(slot.smpl);
} }
slot.smpl = gpt_sampler_init(model, slot.sparams); slot.smpl = common_sampler_init(model, slot.sparams);
if (slot.smpl == nullptr) { if (slot.smpl == nullptr) {
// for now, the only error that may happen here is invalid grammar // for now, the only error that may happen here is invalid grammar
send_error(task, "Failed to parse grammar", ERROR_TYPE_INVALID_REQUEST); send_error(task, "Failed to parse grammar", ERROR_TYPE_INVALID_REQUEST);
@ -1224,7 +1224,7 @@ struct server_context {
std::vector<std::string> samplers; std::vector<std::string> samplers;
samplers.reserve(slot.sparams.samplers.size()); samplers.reserve(slot.sparams.samplers.size());
for (const auto & sampler : slot.sparams.samplers) { for (const auto & sampler : slot.sparams.samplers) {
samplers.emplace_back(gpt_sampler_type_to_str(sampler)); samplers.emplace_back(common_sampler_type_to_str(sampler));
} }
return json { return json {
@ -1232,7 +1232,7 @@ struct server_context {
{"n_predict", slot.n_predict}, // Server configured n_predict {"n_predict", slot.n_predict}, // Server configured n_predict
{"model", params.model_alias}, {"model", params.model_alias},
{"seed", slot.sparams.seed}, {"seed", slot.sparams.seed},
{"seed_cur", slot.smpl ? gpt_sampler_get_seed(slot.smpl) : 0}, {"seed_cur", slot.smpl ? common_sampler_get_seed(slot.smpl) : 0},
{"temperature", slot.sparams.temp}, {"temperature", slot.sparams.temp},
{"dynatemp_range", slot.sparams.dynatemp_range}, {"dynatemp_range", slot.sparams.dynatemp_range},
{"dynatemp_exponent", slot.sparams.dynatemp_exponent}, {"dynatemp_exponent", slot.sparams.dynatemp_exponent},
@ -2092,7 +2092,7 @@ struct server_context {
GGML_ASSERT(slot.n_prompt_tokens < slot.n_ctx); GGML_ASSERT(slot.n_prompt_tokens < slot.n_ctx);
} }
gpt_sampler_reset(slot.smpl); common_sampler_reset(slot.smpl);
if (!slot.params.cache_prompt) { if (!slot.params.cache_prompt) {
slot.n_past_se = 0; slot.n_past_se = 0;
@ -2105,7 +2105,7 @@ struct server_context {
// push the prompt into the sampling context (do not apply grammar) // push the prompt into the sampling context (do not apply grammar)
for (int i = 0; i < slot.n_past; ++i) { for (int i = 0; i < slot.n_past; ++i) {
gpt_sampler_accept(slot.smpl, slot.cache_tokens[i], false); common_sampler_accept(slot.smpl, slot.cache_tokens[i], false);
} }
} }
} }
@ -2159,7 +2159,7 @@ struct server_context {
slot.n_past_se = 0; slot.n_past_se = 0;
slot.ga_i = 0; slot.ga_i = 0;
// TODO: is the system prompt ever in the sampling context? // TODO: is the system prompt ever in the sampling context?
gpt_sampler_reset(slot.smpl); common_sampler_reset(slot.smpl);
} }
// remove the non-common part from the cache // remove the non-common part from the cache
@ -2322,9 +2322,9 @@ struct server_context {
} }
completion_token_output result; completion_token_output result;
const llama_token id = gpt_sampler_sample(slot.smpl, ctx, slot.i_batch - i); const llama_token id = common_sampler_sample(slot.smpl, ctx, slot.i_batch - i);
gpt_sampler_accept(slot.smpl, id, true); common_sampler_accept(slot.smpl, id, true);
slot.n_decoded += 1; slot.n_decoded += 1;
if (slot.n_decoded == 1) { if (slot.n_decoded == 1) {
@ -2335,7 +2335,7 @@ struct server_context {
result.tok = id; result.tok = id;
const auto * cur_p = gpt_sampler_get_candidates(slot.smpl); const auto * cur_p = common_sampler_get_candidates(slot.smpl);
for (size_t i = 0; i < (size_t) slot.sparams.n_probs; ++i) { for (size_t i = 0; i < (size_t) slot.sparams.n_probs; ++i) {
result.probs.push_back({ result.probs.push_back({
@ -2399,13 +2399,13 @@ inline void signal_handler(int signal) {
int main(int argc, char ** argv) { int main(int argc, char ** argv) {
// own arguments required by this example // own arguments required by this example
gpt_params params; common_params params;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_SERVER)) { if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_SERVER)) {
return 1; return 1;
} }
gpt_init(); common_init();
// enabling this will output extra debug information in the HTTP responses from the server // enabling this will output extra debug information in the HTTP responses from the server
// see format_final_response_oaicompat() // see format_final_response_oaicompat()
@ -2427,7 +2427,7 @@ int main(int argc, char ** argv) {
LOG_INF("system info: n_threads = %d, n_threads_batch = %d, total_threads = %d\n", params.cpuparams.n_threads, params.cpuparams_batch.n_threads, std::thread::hardware_concurrency()); LOG_INF("system info: n_threads = %d, n_threads_batch = %d, total_threads = %d\n", params.cpuparams.n_threads, params.cpuparams_batch.n_threads, std::thread::hardware_concurrency());
LOG_INF("\n"); LOG_INF("\n");
LOG_INF("%s\n", gpt_params_get_system_info(params).c_str()); LOG_INF("%s\n", common_params_get_system_info(params).c_str());
LOG_INF("\n"); LOG_INF("\n");
std::unique_ptr<httplib::Server> svr; std::unique_ptr<httplib::Server> svr;

10
examples/simple/simple.cpp
View file

@ -12,16 +12,16 @@ static void print_usage(int, char ** argv) {
} }
int main(int argc, char ** argv) { int main(int argc, char ** argv) {
gpt_params params; common_params params;
params.prompt = "Hello my name is"; params.prompt = "Hello my name is";
params.n_predict = 32; params.n_predict = 32;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, print_usage)) { if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, print_usage)) {
return 1; return 1;
} }
gpt_init(); common_init();
// total length of the sequence including the prompt // total length of the sequence including the prompt
const int n_predict = params.n_predict; const int n_predict = params.n_predict;
@ -33,7 +33,7 @@ int main(int argc, char ** argv) {
// initialize the model // initialize the model
llama_model_params model_params = common_model_params_from_gpt_params(params); llama_model_params model_params = common_model_params_from_common_params(params);
llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params); llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
@ -44,7 +44,7 @@ int main(int argc, char ** argv) {
// initialize the context // initialize the context
llama_context_params ctx_params = common_context_params_from_gpt_params(params); llama_context_params ctx_params = common_context_params_from_common_params(params);
llama_context * ctx = llama_new_context_with_model(model, ctx_params); llama_context * ctx = llama_new_context_with_model(model, ctx_params);

50
examples/speculative/speculative.cpp
View file

@ -26,20 +26,20 @@ struct seq_draft {
std::vector<llama_token> tokens; std::vector<llama_token> tokens;
std::vector<std::vector<llama_token_data>> dists; std::vector<std::vector<llama_token_data>> dists;
struct gpt_sampler * smpl = nullptr; struct common_sampler * smpl = nullptr;
}; };
int main(int argc, char ** argv) { int main(int argc, char ** argv) {
gpt_params params; common_params params;
// needed to get candidate probs even for temp <= 0.0 // needed to get candidate probs even for temp <= 0.0
params.sparams.n_probs = 128; params.sparams.n_probs = 128;
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_SPECULATIVE)) { if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_SPECULATIVE)) {
return 1; return 1;
} }
gpt_init(); common_init();
if (params.model_draft.empty()) { if (params.model_draft.empty()) {
LOG_ERR("%s: --model-draft is required\n", __func__); LOG_ERR("%s: --model-draft is required\n", __func__);
@ -66,7 +66,7 @@ int main(int argc, char ** argv) {
llama_context * ctx_dft = NULL; llama_context * ctx_dft = NULL;
// load the target model // load the target model
common_init_result llama_init_tgt = llama_init_from_gpt_params(params); common_init_result llama_init_tgt = common_init_from_common_params(params);
model_tgt = llama_init_tgt.model; model_tgt = llama_init_tgt.model;
ctx_tgt = llama_init_tgt.context; ctx_tgt = llama_init_tgt.context;
@ -78,7 +78,7 @@ int main(int argc, char ** argv) {
} }
params.cpuparams_batch.n_threads = params.draft_cpuparams_batch.n_threads; params.cpuparams_batch.n_threads = params.draft_cpuparams_batch.n_threads;
common_init_result llama_init_dft = llama_init_from_gpt_params(params); common_init_result llama_init_dft = common_init_from_common_params(params);
model_dft = llama_init_dft.model; model_dft = llama_init_dft.model;
ctx_dft = llama_init_dft.context; ctx_dft = llama_init_dft.context;
@ -178,7 +178,7 @@ int main(int argc, char ** argv) {
bool has_eos = false; bool has_eos = false;
// target model sampling context (reuse the llama_context's sampling instance) // target model sampling context (reuse the llama_context's sampling instance)
struct gpt_sampler * smpl = gpt_sampler_init(model_tgt, params.sparams); struct common_sampler * smpl = common_sampler_init(model_tgt, params.sparams);
struct llama_sampler * softmax = llama_sampler_init_softmax(); struct llama_sampler * softmax = llama_sampler_init_softmax();
@ -186,8 +186,8 @@ int main(int argc, char ** argv) {
std::vector<seq_draft> drafts(n_seq_dft); std::vector<seq_draft> drafts(n_seq_dft);
for (int s = 0; s < n_seq_dft; ++s) { for (int s = 0; s < n_seq_dft; ++s) {
// allocate gpt_sampler for each draft sequence // allocate llama_sampler for each draft sequence
drafts[s].smpl = gpt_sampler_init(model_dft, params.sparams); drafts[s].smpl = common_sampler_init(model_dft, params.sparams);
} }
llama_batch batch_dft = llama_batch_init(params.n_ctx, 0, 1); llama_batch batch_dft = llama_batch_init(params.n_ctx, 0, 1);
@ -229,9 +229,9 @@ int main(int argc, char ** argv) {
bool accept = false; bool accept = false;
if (params.sparams.temp > 0) { if (params.sparams.temp > 0) {
// stochastic verification // stochastic verification
gpt_sampler_sample(smpl, ctx_tgt, drafts[s_keep].i_batch_tgt[i_dft], true); common_sampler_sample(smpl, ctx_tgt, drafts[s_keep].i_batch_tgt[i_dft], true);
auto & dist_tgt = *gpt_sampler_get_candidates(smpl); auto & dist_tgt = *common_sampler_get_candidates(smpl);
float p_tgt = 0.0f; float p_tgt = 0.0f;
float p_dft = 0.0f; float p_dft = 0.0f;
@ -278,7 +278,7 @@ int main(int argc, char ** argv) {
accept = true; accept = true;
token_id = drafts[s].tokens[i_dft]; token_id = drafts[s].tokens[i_dft];
token_str = common_token_to_piece(ctx_tgt, token_id); token_str = common_token_to_piece(ctx_tgt, token_id);
gpt_sampler_accept(smpl, token_id, true); common_sampler_accept(smpl, token_id, true);
LOG_DBG("draft token %d of sequence %d (%d, '%s') accepted\n", i_dft, s, token_id, token_str.c_str()); LOG_DBG("draft token %d of sequence %d (%d, '%s') accepted\n", i_dft, s, token_id, token_str.c_str());
break; break;
@ -349,7 +349,7 @@ int main(int argc, char ** argv) {
const int idx = dist(rng); const int idx = dist(rng);
token_id = dist_tgt.data[idx].id; token_id = dist_tgt.data[idx].id;
gpt_sampler_accept(smpl, token_id, true); common_sampler_accept(smpl, token_id, true);
token_str = common_token_to_piece(ctx_tgt, token_id); token_str = common_token_to_piece(ctx_tgt, token_id);
} }
} else { } else {
@ -357,9 +357,9 @@ int main(int argc, char ** argv) {
// sample from the target model // sample from the target model
LOG_DBG("sampling target: s_keep = %3d, i_dft = %3d, i_batch_tgt = %3d\n", s_keep, i_dft, drafts[s_keep].i_batch_tgt[i_dft]); LOG_DBG("sampling target: s_keep = %3d, i_dft = %3d, i_batch_tgt = %3d\n", s_keep, i_dft, drafts[s_keep].i_batch_tgt[i_dft]);
token_id = gpt_sampler_sample(smpl, ctx_tgt, drafts[s_keep].i_batch_tgt[i_dft]); token_id = common_sampler_sample(smpl, ctx_tgt, drafts[s_keep].i_batch_tgt[i_dft]);
gpt_sampler_accept(smpl, token_id, true); common_sampler_accept(smpl, token_id, true);
token_str = common_token_to_piece(ctx_tgt, token_id); token_str = common_token_to_piece(ctx_tgt, token_id);
@ -446,9 +446,9 @@ int main(int argc, char ** argv) {
} }
if (drafts[0].smpl) { if (drafts[0].smpl) {
gpt_sampler_free(drafts[0].smpl); common_sampler_free(drafts[0].smpl);
} }
drafts[0].smpl = gpt_sampler_clone(smpl); drafts[0].smpl = common_sampler_clone(smpl);
int n_seq_cur = 1; int n_seq_cur = 1;
int n_past_cur = n_past_dft; int n_past_cur = n_past_dft;
@ -477,9 +477,9 @@ int main(int argc, char ** argv) {
continue; continue;
} }
gpt_sampler_sample(drafts[s].smpl, ctx_dft, drafts[s].i_batch_dft, true); common_sampler_sample(drafts[s].smpl, ctx_dft, drafts[s].i_batch_dft, true);
const auto * cur_p = gpt_sampler_get_candidates(drafts[s].smpl); const auto * cur_p = common_sampler_get_candidates(drafts[s].smpl);
for (int k = 0; k < std::min(n_seq_dft + 3, (int) cur_p->size); ++k) { for (int k = 0; k < std::min(n_seq_dft + 3, (int) cur_p->size); ++k) {
LOG_DBG(" - draft candidate %3d for seq %3d, pos %3d: %6d (%8.3f) '%s'\n", LOG_DBG(" - draft candidate %3d for seq %3d, pos %3d: %6d (%8.3f) '%s'\n",
@ -518,9 +518,9 @@ int main(int argc, char ** argv) {
drafts[n_seq_cur].i_batch_tgt = drafts[s].i_batch_tgt; drafts[n_seq_cur].i_batch_tgt = drafts[s].i_batch_tgt;
if (drafts[n_seq_cur].smpl) { if (drafts[n_seq_cur].smpl) {
gpt_sampler_free(drafts[n_seq_cur].smpl); common_sampler_free(drafts[n_seq_cur].smpl);
} }
drafts[n_seq_cur].smpl = gpt_sampler_clone(drafts[s].smpl); drafts[n_seq_cur].smpl = common_sampler_clone(drafts[s].smpl);
sa.push_back(n_seq_cur); sa.push_back(n_seq_cur);
@ -536,7 +536,7 @@ int main(int argc, char ** argv) {
const int s = sa[is]; const int s = sa[is];
gpt_sampler_accept(drafts[s].smpl, id, true); common_sampler_accept(drafts[s].smpl, id, true);
drafts[s].tokens.push_back(id); drafts[s].tokens.push_back(id);
// save cur_p.data into drafts[s].dists // save cur_p.data into drafts[s].dists
@ -617,11 +617,11 @@ int main(int argc, char ** argv) {
LOG_INF("\n"); LOG_INF("\n");
LOG_INF("target:\n\n"); LOG_INF("target:\n\n");
gpt_perf_print(ctx_tgt, smpl); common_perf_print(ctx_tgt, smpl);
gpt_sampler_free(smpl); common_sampler_free(smpl);
for (int s = 0; s < n_seq_dft; ++s) { for (int s = 0; s < n_seq_dft; ++s) {
gpt_sampler_free(drafts[s].smpl); common_sampler_free(drafts[s].smpl);
} }
llama_sampler_free(softmax); llama_sampler_free(softmax);

28
tests/test-arg-parser.cpp
View file

@ -10,12 +10,12 @@
#include <cassert> #include <cassert>
int main(void) { int main(void) {
gpt_params params; common_params params;
printf("test-arg-parser: make sure there is no duplicated arguments in any examples\n\n"); printf("test-arg-parser: make sure there is no duplicated arguments in any examples\n\n");
for (int ex = 0; ex < LLAMA_EXAMPLE_COUNT; ex++) { for (int ex = 0; ex < LLAMA_EXAMPLE_COUNT; ex++) {
try { try {
auto ctx_arg = gpt_params_parser_init(params, (enum llama_example)ex); auto ctx_arg = common_params_parser_init(params, (enum llama_example)ex);
std::unordered_set<std::string> seen_args; std::unordered_set<std::string> seen_args;
std::unordered_set<std::string> seen_env_vars; std::unordered_set<std::string> seen_env_vars;
for (const auto & opt : ctx_arg.options) { for (const auto & opt : ctx_arg.options) {
@ -58,44 +58,44 @@ int main(void) {
// missing value // missing value
argv = {"binary_name", "-m"}; argv = {"binary_name", "-m"};
assert(false == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON)); assert(false == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
// wrong value (int) // wrong value (int)
argv = {"binary_name", "-ngl", "hello"}; argv = {"binary_name", "-ngl", "hello"};
assert(false == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON)); assert(false == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
// wrong value (enum) // wrong value (enum)
argv = {"binary_name", "-sm", "hello"}; argv = {"binary_name", "-sm", "hello"};
assert(false == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON)); assert(false == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
// non-existence arg in specific example (--draft cannot be used outside llama-speculative) // non-existence arg in specific example (--draft cannot be used outside llama-speculative)
argv = {"binary_name", "--draft", "123"}; argv = {"binary_name", "--draft", "123"};
assert(false == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_SERVER)); assert(false == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_SERVER));
printf("test-arg-parser: test valid usage\n\n"); printf("test-arg-parser: test valid usage\n\n");
argv = {"binary_name", "-m", "model_file.gguf"}; argv = {"binary_name", "-m", "model_file.gguf"};
assert(true == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON)); assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
assert(params.model == "model_file.gguf"); assert(params.model == "model_file.gguf");
argv = {"binary_name", "-t", "1234"}; argv = {"binary_name", "-t", "1234"};
assert(true == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON)); assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
assert(params.cpuparams.n_threads == 1234); assert(params.cpuparams.n_threads == 1234);
argv = {"binary_name", "--verbose"}; argv = {"binary_name", "--verbose"};
assert(true == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON)); assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
assert(params.verbosity > 1); assert(params.verbosity > 1);
argv = {"binary_name", "-m", "abc.gguf", "--predict", "6789", "--batch-size", "9090"}; argv = {"binary_name", "-m", "abc.gguf", "--predict", "6789", "--batch-size", "9090"};
assert(true == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON)); assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
assert(params.model == "abc.gguf"); assert(params.model == "abc.gguf");
assert(params.n_predict == 6789); assert(params.n_predict == 6789);
assert(params.n_batch == 9090); assert(params.n_batch == 9090);
// --draft cannot be used outside llama-speculative // --draft cannot be used outside llama-speculative
argv = {"binary_name", "--draft", "123"}; argv = {"binary_name", "--draft", "123"};
assert(true == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_SPECULATIVE)); assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_SPECULATIVE));
assert(params.n_draft == 123); assert(params.n_draft == 123);
// skip this part on windows, because setenv is not supported // skip this part on windows, because setenv is not supported
@ -106,12 +106,12 @@ int main(void) {
setenv("LLAMA_ARG_THREADS", "blah", true); setenv("LLAMA_ARG_THREADS", "blah", true);
argv = {"binary_name"}; argv = {"binary_name"};
assert(false == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON)); assert(false == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
setenv("LLAMA_ARG_MODEL", "blah.gguf", true); setenv("LLAMA_ARG_MODEL", "blah.gguf", true);
setenv("LLAMA_ARG_THREADS", "1010", true); setenv("LLAMA_ARG_THREADS", "1010", true);
argv = {"binary_name"}; argv = {"binary_name"};
assert(true == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON)); assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
assert(params.model == "blah.gguf"); assert(params.model == "blah.gguf");
assert(params.cpuparams.n_threads == 1010); assert(params.cpuparams.n_threads == 1010);
@ -121,7 +121,7 @@ int main(void) {
setenv("LLAMA_ARG_MODEL", "blah.gguf", true); setenv("LLAMA_ARG_MODEL", "blah.gguf", true);
setenv("LLAMA_ARG_THREADS", "1010", true); setenv("LLAMA_ARG_THREADS", "1010", true);
argv = {"binary_name", "-m", "overwritten.gguf"}; argv = {"binary_name", "-m", "overwritten.gguf"};
assert(true == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON)); assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
assert(params.model == "overwritten.gguf"); assert(params.model == "overwritten.gguf");
assert(params.cpuparams.n_threads == 1010); assert(params.cpuparams.n_threads == 1010);
#endif // _WIN32 #endif // _WIN32

4
tests/test-log.cpp
View file

@ -24,8 +24,8 @@ int main() {
} }
if (rand () % 10 < 5) { if (rand () % 10 < 5) {
gpt_log_set_timestamps(gpt_log_main(), rand() % 2); common_log_set_timestamps(common_log_main(), rand() % 2);
gpt_log_set_prefix (gpt_log_main(), rand() % 2); common_log_set_prefix (common_log_main(), rand() % 2);
} }
} }
}); });