Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-26 06:10:29 +01:00
Code Issues Packages Projects Releases Wiki Activity

common : use common_ prefix for common library functions (#9805)

Browse Source 
* common : use common_ prefix for common library functions

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Diego Devesa 2024-10-10 22:57:42 +02:00 committed by GitHub
parent 0e9f760eb1
commit 7eee341bee
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
45 changed files with 1284 additions and 1284 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

868
common/arg.cpp
View File

File diff suppressed because it is too large Load Diff

44
common/arg.h
View File

@ -10,7 +10,7 @@
// CLI argument parsing // CLI argument parsing
// //
struct llama_arg { struct common_arg {
std::set<enum llama_example> examples = {LLAMA_EXAMPLE_COMMON}; std::set<enum llama_example> examples = {LLAMA_EXAMPLE_COMMON};
std::vector<const char *> args; std::vector<const char *> args;
const char * value_hint = nullptr; // help text or example for arg value const char * value_hint = nullptr; // help text or example for arg value
@ -18,60 +18,60 @@ struct llama_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;
llama_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) {}
llama_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) {}
llama_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
llama_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 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) {}
llama_arg & set_examples(std::initializer_list<enum llama_example> examples); common_arg & set_examples(std::initializer_list<enum llama_example> examples);
llama_arg & set_env(const char * env); common_arg & set_env(const char * env);
llama_arg & set_sparam(); common_arg & set_sparam();
bool in_example(enum llama_example ex); bool in_example(enum llama_example ex);
bool get_value_from_env(std::string & output); bool get_value_from_env(std::string & output);
bool has_value_from_env(); bool has_value_from_env();
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<llama_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);

116
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;
@ -493,7 +493,7 @@ std::string string_from(const struct llama_context * ctx, const std::vector<llam
first = false; first = false;
} }
auto detokenized = llama_token_to_piece(ctx, token); auto detokenized = common_token_to_piece(ctx, token);
detokenized.erase( detokenized.erase(
std::remove_if( std::remove_if(
@ -524,7 +524,7 @@ std::string string_from(const struct llama_context * ctx, const struct llama_bat
first = false; first = false;
} }
auto detokenized = llama_token_to_piece(ctx, batch.token[i]); auto detokenized = common_token_to_piece(ctx, batch.token[i]);
detokenized.erase( detokenized.erase(
std::remove_if( std::remove_if(
@ -819,16 +819,16 @@ std::string fs_get_cache_file(const std::string & filename) {
// //
// Model utils // Model utils
// //
struct llama_init_result llama_init_from_gpt_params(gpt_params & params) { struct common_init_result common_init_from_params(common_params & params) {
llama_init_result iparams; common_init_result iparams;
auto mparams = llama_model_params_from_gpt_params(params); auto mparams = common_model_params_to_llama(params);
llama_model * model = nullptr; llama_model * model = nullptr;
if (!params.hf_repo.empty() && !params.hf_file.empty()) { if (!params.hf_repo.empty() && !params.hf_file.empty()) {
model = llama_load_model_from_hf(params.hf_repo.c_str(), params.hf_file.c_str(), params.model.c_str(), params.hf_token.c_str(), mparams); model = common_load_model_from_hf(params.hf_repo.c_str(), params.hf_file.c_str(), params.model.c_str(), params.hf_token.c_str(), mparams);
} else if (!params.model_url.empty()) { } else if (!params.model_url.empty()) {
model = llama_load_model_from_url(params.model_url.c_str(), params.model.c_str(), params.hf_token.c_str(), mparams); model = common_load_model_from_url(params.model_url.c_str(), params.model.c_str(), params.hf_token.c_str(), mparams);
} else { } else {
model = llama_load_model_from_file(params.model.c_str(), mparams); model = llama_load_model_from_file(params.model.c_str(), mparams);
} }
@ -863,7 +863,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
} }
} }
auto cparams = llama_context_params_from_gpt_params(params); auto cparams = common_context_params_to_llama(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) {
@ -876,7 +876,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
if (params.control_vector_layer_start <= 0) params.control_vector_layer_start = 1; if (params.control_vector_layer_start <= 0) params.control_vector_layer_start = 1;
if (params.control_vector_layer_end <= 0) params.control_vector_layer_end = llama_n_layer(model); if (params.control_vector_layer_end <= 0) params.control_vector_layer_end = llama_n_layer(model);
const auto cvec = llama_control_vector_load(params.control_vectors); const auto cvec = common_control_vector_load(params.control_vectors);
if (cvec.n_embd == -1) { if (cvec.n_embd == -1) {
llama_free(lctx); llama_free(lctx);
llama_free_model(model); llama_free_model(model);
@ -900,7 +900,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
// load and optionally apply lora adapters // load and optionally apply lora adapters
for (auto & la : params.lora_adapters) { for (auto & la : params.lora_adapters) {
llama_lora_adapter_container loaded_la; common_lora_adapter_container loaded_la;
loaded_la.path = la.path; loaded_la.path = la.path;
loaded_la.scale = la.scale; loaded_la.scale = la.scale;
loaded_la.adapter = llama_lora_adapter_init(model, la.path.c_str()); loaded_la.adapter = llama_lora_adapter_init(model, la.path.c_str());
@ -913,7 +913,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
iparams.lora_adapters.push_back(loaded_la); // copy to list of loaded adapters iparams.lora_adapters.push_back(loaded_la); // copy to list of loaded adapters
} }
if (!params.lora_init_without_apply) { if (!params.lora_init_without_apply) {
llama_lora_adapters_apply(lctx, iparams.lora_adapters); common_lora_adapters_apply(lctx, iparams.lora_adapters);
} }
if (params.sparams.ignore_eos && llama_token_eos(model) == LLAMA_TOKEN_NULL) { if (params.sparams.ignore_eos && llama_token_eos(model) == LLAMA_TOKEN_NULL) {
@ -961,7 +961,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
return iparams; return iparams;
} }
void llama_lora_adapters_apply(struct llama_context * ctx, std::vector<llama_lora_adapter_container> & lora_adapters) { void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_lora_adapter_container> & lora_adapters) {
llama_lora_adapter_clear(ctx); llama_lora_adapter_clear(ctx);
for (auto & la : lora_adapters) { for (auto & la : lora_adapters) {
if (la.scale != 0.0f) { if (la.scale != 0.0f) {
@ -970,7 +970,7 @@ void llama_lora_adapters_apply(struct llama_context * ctx, std::vector<llama_lor
} }
} }
struct llama_model_params llama_model_params_from_gpt_params(const gpt_params & params) { struct llama_model_params common_model_params_to_llama(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 llama_context_params_from_gpt_params(const gpt_params & params) { struct llama_context_params common_context_params_to_llama(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;
@ -1112,7 +1112,7 @@ static bool curl_perform_with_retry(const std::string& url, CURL* curl, int max_
return false; return false;
} }
static bool llama_download_file(const std::string & url, const std::string & path, const std::string & hf_token) { static bool common_download_file(const std::string & url, const std::string & path, const std::string & hf_token) {
// Initialize libcurl // Initialize libcurl
std::unique_ptr<CURL, decltype(&curl_easy_cleanup)> curl(curl_easy_init(), &curl_easy_cleanup); std::unique_ptr<CURL, decltype(&curl_easy_cleanup)> curl(curl_easy_init(), &curl_easy_cleanup);
@ -1182,15 +1182,15 @@ static bool llama_download_file(const std::string & url, const std::string & pat
} }
// Send a HEAD request to retrieve the etag and last-modified headers // Send a HEAD request to retrieve the etag and last-modified headers
struct llama_load_model_from_url_headers { struct common_load_model_from_url_headers {
std::string etag; std::string etag;
std::string last_modified; std::string last_modified;
}; };
llama_load_model_from_url_headers headers; common_load_model_from_url_headers headers;
{ {
typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *); typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *);
auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t { auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t {
llama_load_model_from_url_headers *headers = (llama_load_model_from_url_headers *) userdata; common_load_model_from_url_headers *headers = (common_load_model_from_url_headers *) userdata;
static std::regex header_regex("([^:]+): (.*)\r\n"); static std::regex header_regex("([^:]+): (.*)\r\n");
static std::regex etag_regex("ETag", std::regex_constants::icase); static std::regex etag_regex("ETag", std::regex_constants::icase);
@ -1326,7 +1326,7 @@ static bool llama_download_file(const std::string & url, const std::string & pat
return true; return true;
} }
struct llama_model * llama_load_model_from_url( struct llama_model * common_load_model_from_url(
const char * model_url, const char * model_url,
const char * path_model, const char * path_model,
const char * hf_token, const char * hf_token,
@ -1337,7 +1337,7 @@ struct llama_model * llama_load_model_from_url(
return NULL; return NULL;
} }
if (!llama_download_file(model_url, path_model, hf_token)) { if (!common_download_file(model_url, path_model, hf_token)) {
return NULL; return NULL;
} }
@ -1390,7 +1390,7 @@ struct llama_model * llama_load_model_from_url(
char split_url[LLAMA_CURL_MAX_URL_LENGTH] = {0}; char split_url[LLAMA_CURL_MAX_URL_LENGTH] = {0};
llama_split_path(split_url, sizeof(split_url), split_url_prefix, download_idx, n_split); llama_split_path(split_url, sizeof(split_url), split_url_prefix, download_idx, n_split);
return llama_download_file(split_url, split_path, hf_token); return common_download_file(split_url, split_path, hf_token);
}, idx)); }, idx));
} }
@ -1405,7 +1405,7 @@ struct llama_model * llama_load_model_from_url(
return llama_load_model_from_file(path_model, params); return llama_load_model_from_file(path_model, params);
} }
struct llama_model * llama_load_model_from_hf( struct llama_model * common_load_model_from_hf(
const char * repo, const char * repo,
const char * model, const char * model,
const char * path_model, const char * path_model,
@ -1425,12 +1425,12 @@ struct llama_model * llama_load_model_from_hf(
model_url += "/resolve/main/"; model_url += "/resolve/main/";
model_url += model; model_url += model;
return llama_load_model_from_url(model_url.c_str(), path_model, hf_token, params); return common_load_model_from_url(model_url.c_str(), path_model, hf_token, params);
} }
#else #else
struct llama_model * llama_load_model_from_url( struct llama_model * common_load_model_from_url(
const char * /*model_url*/, const char * /*model_url*/,
const char * /*path_model*/, const char * /*path_model*/,
const char * /*hf_token*/, const char * /*hf_token*/,
@ -1439,7 +1439,7 @@ struct llama_model * llama_load_model_from_url(
return nullptr; return nullptr;
} }
struct llama_model * llama_load_model_from_hf( struct llama_model * common_load_model_from_hf(
const char * /*repo*/, const char * /*repo*/,
const char * /*model*/, const char * /*model*/,
const char * /*path_model*/, const char * /*path_model*/,
@ -1455,11 +1455,11 @@ struct llama_model * llama_load_model_from_hf(
// Batch utils // Batch utils
// //
void llama_batch_clear(struct llama_batch & batch) { void common_batch_clear(struct llama_batch & batch) {
batch.n_tokens = 0; batch.n_tokens = 0;
} }
void llama_batch_add( void common_batch_add(
struct llama_batch & batch, struct llama_batch & batch,
llama_token id, llama_token id,
llama_pos pos, llama_pos pos,
@ -1482,15 +1482,15 @@ void llama_batch_add(
// Vocab utils // Vocab utils
// //
std::vector<llama_token> llama_tokenize( std::vector<llama_token> common_tokenize(
const struct llama_context * ctx, const struct llama_context * ctx,
const std::string & text, const std::string & text,
bool add_special, bool add_special,
bool parse_special) { bool parse_special) {
return llama_tokenize(llama_get_model(ctx), text, add_special, parse_special); return common_tokenize(llama_get_model(ctx), text, add_special, parse_special);
} }
std::vector<llama_token> llama_tokenize( std::vector<llama_token> common_tokenize(
const struct llama_model * model, const struct llama_model * model,
const std::string & text, const std::string & text,
bool add_special, bool add_special,
@ -1509,7 +1509,7 @@ std::vector<llama_token> llama_tokenize(
return result; return result;
} }
std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token, bool special) { std::string common_token_to_piece(const struct llama_context * ctx, llama_token token, bool special) {
std::string piece; std::string piece;
piece.resize(piece.capacity()); // using string internal cache, 15 bytes + '\n' piece.resize(piece.capacity()); // using string internal cache, 15 bytes + '\n'
const int n_chars = llama_token_to_piece(llama_get_model(ctx), token, &piece[0], piece.size(), 0, special); const int n_chars = llama_token_to_piece(llama_get_model(ctx), token, &piece[0], piece.size(), 0, special);
@ -1525,7 +1525,7 @@ std::string llama_token_to_piece(const struct llama_context * ctx, llama_token t
return piece; return piece;
} }
std::string llama_detokenize(llama_context * ctx, const std::vector<llama_token> & tokens, bool special) { std::string common_detokenize(llama_context * ctx, const std::vector<llama_token> & tokens, bool special) {
std::string text; std::string text;
text.resize(std::max(text.capacity(), tokens.size())); text.resize(std::max(text.capacity(), tokens.size()));
int32_t n_chars = llama_detokenize(llama_get_model(ctx), tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special); int32_t n_chars = llama_detokenize(llama_get_model(ctx), tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
@ -1545,15 +1545,15 @@ std::string llama_detokenize(llama_context * ctx, const std::vector<llama_token>
// Chat template utils // Chat template utils
// //
bool llama_chat_verify_template(const std::string & tmpl) { bool common_chat_verify_template(const std::string & tmpl) {
llama_chat_message chat[] = {{"user", "test"}}; llama_chat_message chat[] = {{"user", "test"}};
int res = llama_chat_apply_template(nullptr, tmpl.c_str(), chat, 1, true, nullptr, 0); int res = llama_chat_apply_template(nullptr, tmpl.c_str(), chat, 1, true, nullptr, 0);
return res >= 0; return res >= 0;
} }
std::string llama_chat_apply_template(const struct llama_model * model, std::string common_chat_apply_template(const struct llama_model * model,
const std::string & tmpl, const std::string & tmpl,
const std::vector<llama_chat_msg> & msgs, const std::vector<common_chat_msg> & msgs,
bool add_ass) { bool add_ass) {
int alloc_size = 0; int alloc_size = 0;
bool fallback = false; // indicate if we must fallback to default chatml bool fallback = false; // indicate if we must fallback to default chatml
@ -1595,42 +1595,42 @@ std::string llama_chat_apply_template(const struct llama_model * model,
return formatted_chat; return formatted_chat;
} }
std::string llama_chat_format_single(const struct llama_model * model, std::string common_chat_format_single(const struct llama_model * model,
const std::string & tmpl, const std::string & tmpl,
const std::vector<llama_chat_msg> & past_msg, const std::vector<common_chat_msg> & past_msg,
const llama_chat_msg & new_msg, const common_chat_msg & new_msg,
bool add_ass) { bool add_ass) {
std::ostringstream ss; std::ostringstream ss;
auto fmt_past_msg = past_msg.empty() ? "" : llama_chat_apply_template(model, tmpl, past_msg, false); auto fmt_past_msg = past_msg.empty() ? "" : common_chat_apply_template(model, tmpl, past_msg, false);
std::vector<llama_chat_msg> chat_new(past_msg); std::vector<common_chat_msg> chat_new(past_msg);
// if the past_msg ends with a newline, we must preserve it in the formatted version // if the past_msg ends with a newline, we must preserve it in the formatted version
if (add_ass && !fmt_past_msg.empty() && fmt_past_msg.back() == '\n') { if (add_ass && !fmt_past_msg.empty() && fmt_past_msg.back() == '\n') {
ss << "\n"; ss << "\n";
}; };
// format chat with new_msg // format chat with new_msg
chat_new.push_back(new_msg); chat_new.push_back(new_msg);
auto fmt_new_msg = llama_chat_apply_template(model, tmpl, chat_new, add_ass); auto fmt_new_msg = common_chat_apply_template(model, tmpl, chat_new, add_ass);
// get the diff part // get the diff part
ss << fmt_new_msg.substr(fmt_past_msg.size(), fmt_new_msg.size() - fmt_past_msg.size()); ss << fmt_new_msg.substr(fmt_past_msg.size(), fmt_new_msg.size() - fmt_past_msg.size());
return ss.str(); return ss.str();
} }
std::string llama_chat_format_example(const struct llama_model * model, std::string common_chat_format_example(const struct llama_model * model,
const std::string & tmpl) { const std::string & tmpl) {
std::vector<llama_chat_msg> msgs = { std::vector<common_chat_msg> msgs = {
{"system", "You are a helpful assistant"}, {"system", "You are a helpful assistant"},
{"user", "Hello"}, {"user", "Hello"},
{"assistant", "Hi there"}, {"assistant", "Hi there"},
{"user", "How are you?"}, {"user", "How are you?"},
}; };
return llama_chat_apply_template(model, tmpl, msgs, true); return common_chat_apply_template(model, tmpl, msgs, true);
} }
// //
// KV cache utils // KV cache utils
// //
void llama_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size) { void common_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size) {
static const char slot_chars[] = ".123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz+"; static const char slot_chars[] = ".123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz+";
printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d", printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d",
@ -1653,7 +1653,7 @@ void llama_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size) {
printf("\n=== Done dumping\n"); printf("\n=== Done dumping\n");
} }
void llama_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size) { void common_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size) {
static const char slot_chars[] = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz"; static const char slot_chars[] = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz";
printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d\n", printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d\n",
@ -1705,7 +1705,7 @@ void llama_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_siz
// Embedding utils // Embedding utils
// //
void llama_embd_normalize(const float * inp, float * out, int n, int embd_norm) { void common_embd_normalize(const float * inp, float * out, int n, int embd_norm) {
double sum = 0.0; double sum = 0.0;
switch (embd_norm) { switch (embd_norm) {
@ -1739,7 +1739,7 @@ void llama_embd_normalize(const float * inp, float * out, int n, int embd_norm)
} }
} }
float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n){ float common_embd_similarity_cos(const float * embd1, const float * embd2, int n){
double sum = 0.0; double sum = 0.0;
double sum1 = 0.0; double sum1 = 0.0;
double sum2 = 0.0; double sum2 = 0.0;
@ -1765,8 +1765,8 @@ float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n)
// Control vector utils // Control vector utils
// //
static llama_control_vector_data llama_control_vector_load_one(const llama_control_vector_load_info & load_info) { static common_control_vector_data common_control_vector_load_one(const common_control_vector_load_info & load_info) {
llama_control_vector_data result = { -1, {} }; common_control_vector_data result = { -1, {} };
ggml_context * ctx = nullptr; ggml_context * ctx = nullptr;
struct gguf_init_params meta_gguf_params = { struct gguf_init_params meta_gguf_params = {
@ -1850,11 +1850,11 @@ static llama_control_vector_data llama_control_vector_load_one(const llama_contr
return result; return result;
} }
llama_control_vector_data llama_control_vector_load(const std::vector<llama_control_vector_load_info> & load_infos) { common_control_vector_data common_control_vector_load(const std::vector<common_control_vector_load_info> & load_infos) {
llama_control_vector_data result = { -1, {} }; common_control_vector_data result = { -1, {} };
for (const auto & info : load_infos) { for (const auto & info : load_infos) {
auto cur = llama_control_vector_load_one(info); auto cur = common_control_vector_load_one(info);
if (cur.n_embd == -1) { if (cur.n_embd == -1) {
result.n_embd = -1; result.n_embd = -1;
@ -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;

110
common/common.h
View File

@ -24,12 +24,12 @@
#define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf" #define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
struct llama_lora_adapter_info { struct common_lora_adapter_info {
std::string path; std::string path;
float scale; float scale;
}; };
struct llama_lora_adapter_container : llama_lora_adapter_info { struct common_lora_adapter_container : common_lora_adapter_info {
struct llama_lora_adapter * adapter; struct llama_lora_adapter * adapter;
}; };
@ -39,7 +39,7 @@ extern char const * LLAMA_COMMIT;
extern char const * LLAMA_COMPILER; extern char const * LLAMA_COMPILER;
extern char const * LLAMA_BUILD_TARGET; extern char const * LLAMA_BUILD_TARGET;
struct llama_control_vector_load_info; struct common_control_vector_load_info;
// //
// CPU utils // CPU utils
@ -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
@ -208,9 +208,9 @@ struct gpt_params {
std::vector<llama_model_kv_override> kv_overrides; std::vector<llama_model_kv_override> kv_overrides;
bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_lora_adapter_apply) bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_lora_adapter_apply)
std::vector<llama_lora_adapter_info> lora_adapters; // lora adapter path with user defined scale std::vector<common_lora_adapter_info> lora_adapters; // lora adapter path with user defined scale
std::vector<llama_control_vector_load_info> control_vectors; // control vector with user defined scale std::vector<common_control_vector_load_info> control_vectors; // control vector with user defined scale
int32_t verbosity = 0; int32_t verbosity = 0;
int32_t control_vector_layer_start = -1; // layer range for control vector int32_t control_vector_layer_start = -1; // layer range for control vector
@ -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]);
@ -404,29 +404,29 @@ std::string fs_get_cache_file(const std::string & filename);
// Model utils // Model utils
// //
struct llama_init_result { struct common_init_result {
struct llama_model * model = nullptr; struct llama_model * model = nullptr;
struct llama_context * context = nullptr; struct llama_context * context = nullptr;
std::vector<llama_lora_adapter_container> lora_adapters; std::vector<common_lora_adapter_container> lora_adapters;
}; };
struct llama_init_result llama_init_from_gpt_params(gpt_params & params); struct common_init_result common_init_from_params(common_params & params);
struct llama_model_params llama_model_params_from_gpt_params (const gpt_params & params); struct llama_model_params common_model_params_to_llama (const common_params & params);
struct llama_context_params llama_context_params_from_gpt_params (const gpt_params & params); struct llama_context_params common_context_params_to_llama(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 * llama_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);
struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const char * hf_token, const struct llama_model_params & params); struct llama_model * common_load_model_from_hf(const char * repo, const char * file, const char * path_model, const char * hf_token, const struct llama_model_params & params);
// clear LoRA adapters from context, then apply new list of adapters // clear LoRA adapters from context, then apply new list of adapters
void llama_lora_adapters_apply(struct llama_context * ctx, std::vector<llama_lora_adapter_container> & lora_adapters); void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_lora_adapter_container> & lora_adapters);
// Batch utils // Batch utils
void llama_batch_clear(struct llama_batch & batch); void common_batch_clear(struct llama_batch & batch);
void llama_batch_add( void common_batch_add(
struct llama_batch & batch, struct llama_batch & batch,
llama_token id, llama_token id,
llama_pos pos, llama_pos pos,
@ -439,13 +439,13 @@ void llama_batch_add(
// tokenizes a string into a vector of tokens // tokenizes a string into a vector of tokens
// should work similar to Python's `tokenizer.encode` // should work similar to Python's `tokenizer.encode`
std::vector<llama_token> llama_tokenize( std::vector<llama_token> common_tokenize(
const struct llama_context * ctx, const struct llama_context * ctx,
const std::string & text, const std::string & text,
bool add_special, bool add_special,
bool parse_special = false); bool parse_special = false);
std::vector<llama_token> llama_tokenize( std::vector<llama_token> common_tokenize(
const struct llama_model * model, const struct llama_model * model,
const std::string & text, const std::string & text,
bool add_special, bool add_special,
@ -453,7 +453,7 @@ std::vector<llama_token> llama_tokenize(
// tokenizes a token into a piece, optionally renders special/control tokens // tokenizes a token into a piece, optionally renders special/control tokens
// should work similar to Python's `tokenizer.id_to_piece` // should work similar to Python's `tokenizer.id_to_piece`
std::string llama_token_to_piece( std::string common_token_to_piece(
const struct llama_context * ctx, const struct llama_context * ctx,
llama_token token, llama_token token,
bool special = true); bool special = true);
@ -461,7 +461,7 @@ std::string llama_token_to_piece(
// detokenizes a vector of tokens into a string // detokenizes a vector of tokens into a string
// should work similar to Python's `tokenizer.decode` // should work similar to Python's `tokenizer.decode`
// optionally renders special/control tokens // optionally renders special/control tokens
std::string llama_detokenize( std::string common_detokenize(
llama_context * ctx, llama_context * ctx,
const std::vector<llama_token> & tokens, const std::vector<llama_token> & tokens,
bool special = true); bool special = true);
@ -471,31 +471,31 @@ std::string llama_detokenize(
// //
// same with llama_chat_message, but uses std::string // same with llama_chat_message, but uses std::string
struct llama_chat_msg { struct common_chat_msg {
std::string role; std::string role;
std::string content; std::string content;
}; };
// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid // Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
bool llama_chat_verify_template(const std::string & tmpl); bool common_chat_verify_template(const std::string & tmpl);
// CPP wrapper for llama_chat_apply_template // CPP wrapper for llama_chat_apply_template
// If the built-in template is not supported, we default to chatml // If the built-in template is not supported, we default to chatml
// If the custom "tmpl" is not supported, we throw an error // If the custom "tmpl" is not supported, we throw an error
std::string llama_chat_apply_template(const struct llama_model * model, std::string common_chat_apply_template(const struct llama_model * model,
const std::string & tmpl, const std::string & tmpl,
const std::vector<llama_chat_msg> & chat, const std::vector<common_chat_msg> & chat,
bool add_ass); bool add_ass);
// Format single message, while taking into account the position of that message in chat history // Format single message, while taking into account the position of that message in chat history
std::string llama_chat_format_single(const struct llama_model * model, std::string common_chat_format_single(const struct llama_model * model,
const std::string & tmpl, const std::string & tmpl,
const std::vector<llama_chat_msg> & past_msg, const std::vector<common_chat_msg> & past_msg,
const llama_chat_msg & new_msg, const common_chat_msg & new_msg,
bool add_ass); bool add_ass);
// Returns an example of formatted chat // Returns an example of formatted chat
std::string llama_chat_format_example(const struct llama_model * model, std::string common_chat_format_example(const struct llama_model * model,
const std::string & tmpl); const std::string & tmpl);
// //
@ -503,31 +503,31 @@ std::string llama_chat_format_example(const struct llama_model * model,
// //
// Dump the KV cache view with the number of sequences per cell. // Dump the KV cache view with the number of sequences per cell.
void llama_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size = 80); void common_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size = 80);
// Dump the KV cache view showing individual sequences in each cell (long output). // Dump the KV cache view showing individual sequences in each cell (long output).
void llama_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size = 40); void common_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
// //
// Embedding utils // Embedding utils
// //
void llama_embd_normalize(const float * inp, float * out, int n, int embd_norm = 2); void common_embd_normalize(const float * inp, float * out, int n, int embd_norm = 2);
float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n); float common_embd_similarity_cos(const float * embd1, const float * embd2, int n);
// //
// Control vector utils // Control vector utils
// //
struct llama_control_vector_data { struct common_control_vector_data {
int n_embd; int n_embd;
// stores data for layers [1, n_layer] where n_layer = data.size() / n_embd // stores data for layers [1, n_layer] where n_layer = data.size() / n_embd
std::vector<float> data; std::vector<float> data;
}; };
struct llama_control_vector_load_info { struct common_control_vector_load_info {
float strength; float strength;
std::string fname; std::string fname;
@ -535,7 +535,7 @@ struct llama_control_vector_load_info {
// Load control vectors, scale each by strength, and add them together. // Load control vectors, scale each by strength, and add them together.
// On error, returns {-1, empty} // On error, returns {-1, empty}
llama_control_vector_data llama_control_vector_load(const std::vector<llama_control_vector_load_info> & load_infos); common_control_vector_data common_control_vector_load(const std::vector<common_control_vector_load_info> & load_infos);
// //
// Split utils // Split utils
@ -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)

72
common/ngram-cache.cpp
View File

@ -8,7 +8,7 @@
#include <fstream> #include <fstream>
#include <thread> #include <thread>
void llama_ngram_cache_update(llama_ngram_cache & ngram_cache, int ngram_min, int ngram_max, void common_ngram_cache_update(common_ngram_cache & ngram_cache, int ngram_min, int ngram_max,
std::vector<llama_token> & inp, int nnew, bool print_progress) { std::vector<llama_token> & inp, int nnew, bool print_progress) {
const int64_t t_start_ms = ggml_time_ms(); const int64_t t_start_ms = ggml_time_ms();
const int64_t inp_size = inp.size(); const int64_t inp_size = inp.size();
@ -20,16 +20,16 @@ void llama_ngram_cache_update(llama_ngram_cache & ngram_cache, int ngram_min, in
const int64_t i_start = std::max(inp_size - nnew, ngram_size); const int64_t i_start = std::max(inp_size - nnew, ngram_size);
for (int64_t i = i_start; i < inp_size; ++i) { for (int64_t i = i_start; i < inp_size; ++i) {
const int64_t ngram_start = i - ngram_size; const int64_t ngram_start = i - ngram_size;
llama_ngram ngram(&inp[ngram_start], ngram_size); common_ngram ngram(&inp[ngram_start], ngram_size);
const llama_token token = inp[i]; const llama_token token = inp[i];
llama_ngram_cache::iterator part_it = ngram_cache.find(ngram); common_ngram_cache::iterator part_it = ngram_cache.find(ngram);
if (part_it == ngram_cache.end()) { if (part_it == ngram_cache.end()) {
llama_ngram_cache_part part; common_ngram_cache_part part;
part.emplace(token, 1); part.emplace(token, 1);
ngram_cache.emplace(ngram, part); ngram_cache.emplace(ngram, part);
} else { } else {
llama_ngram_cache_part::iterator token_count_it = part_it->second.find(token); common_ngram_cache_part::iterator token_count_it = part_it->second.find(token);
if (token_count_it == part_it->second.end()) { if (token_count_it == part_it->second.end()) {
part_it->second.emplace(token, 1); part_it->second.emplace(token, 1);
} else { } else {
@ -62,12 +62,12 @@ constexpr int draft_min_sample_size_strict[LLAMA_NGRAM_MAX] = { 4, 3, 2, 2};
constexpr int draft_min_percent_strict[LLAMA_NGRAM_MAX] = {75, 66, 66, 66}; constexpr int draft_min_percent_strict[LLAMA_NGRAM_MAX] = {75, 66, 66, 66};
// Helper function that tries to draft a token from only the static ngram cache: // Helper function that tries to draft a token from only the static ngram cache:
static llama_token try_draft(llama_ngram_cache & nc_static, const llama_ngram ngram_static) { static llama_token try_draft(common_ngram_cache & nc_static, const common_ngram ngram_static) {
llama_ngram_cache::iterator part_static_it = nc_static.find(ngram_static); common_ngram_cache::iterator part_static_it = nc_static.find(ngram_static);
if (part_static_it == nc_static.end()) { if (part_static_it == nc_static.end()) {
return -1; return -1;
} }
const llama_ngram_cache_part part_static = part_static_it->second; const common_ngram_cache_part part_static = part_static_it->second;
int max_count_static = 0; int max_count_static = 0;
int sum_count_static = 0; int sum_count_static = 0;
@ -95,19 +95,19 @@ static llama_token try_draft(llama_ngram_cache & nc_static, const llama_ngram ng
// Try to draft a token from primary cache (context/dynamic), validate with static cache: // Try to draft a token from primary cache (context/dynamic), validate with static cache:
static llama_token try_draft( static llama_token try_draft(
llama_ngram_cache & nc_primary, const std::vector<llama_ngram> & ngrams_primary, llama_ngram_cache_part & part_static, common_ngram_cache & nc_primary, const std::vector<common_ngram> & ngrams_primary, common_ngram_cache_part & part_static,
const int * min_sample_size, const int * min_percent) { const int * min_sample_size, const int * min_percent) {
llama_token drafted_token = -1; llama_token drafted_token = -1;
for (int i = ngrams_primary.size()-1; i >= 0 && drafted_token == -1; --i) { for (int i = ngrams_primary.size()-1; i >= 0 && drafted_token == -1; --i) {
const llama_ngram ngram_primary = ngrams_primary[i]; const common_ngram ngram_primary = ngrams_primary[i];
llama_ngram_cache::iterator part_primary_it = nc_primary.find(ngram_primary); common_ngram_cache::iterator part_primary_it = nc_primary.find(ngram_primary);
if (part_primary_it == nc_primary.end()) { if (part_primary_it == nc_primary.end()) {
continue; continue;
} }
const llama_ngram_cache_part part_primary = part_primary_it->second; const common_ngram_cache_part part_primary = part_primary_it->second;
int max_count_primary = 0; int max_count_primary = 0;
int max_count_static = 0; int max_count_static = 0;
@ -117,7 +117,7 @@ static llama_token try_draft(
for (std::pair<llama_token, int> token_count_primary : part_primary) { for (std::pair<llama_token, int> token_count_primary : part_primary) {
const llama_token token = token_count_primary.first; const llama_token token = token_count_primary.first;
llama_ngram_cache_part::iterator token_count_static_it = part_static.find(token); common_ngram_cache_part::iterator token_count_static_it = part_static.find(token);
const int32_t count_primary = token_count_primary.second; const int32_t count_primary = token_count_primary.second;
const int32_t count_static = token_count_static_it != part_static.end() ? 100*token_count_static_it->second : 1; const int32_t count_static = token_count_static_it != part_static.end() ? 100*token_count_static_it->second : 1;
@ -142,9 +142,9 @@ static llama_token try_draft(
return drafted_token; return drafted_token;
} }
void llama_ngram_cache_draft( void common_ngram_cache_draft(
std::vector<llama_token> & inp, std::vector<llama_token> & draft, int n_draft, int ngram_min, int ngram_max, std::vector<llama_token> & inp, std::vector<llama_token> & draft, int n_draft, int ngram_min, int ngram_max,
llama_ngram_cache & nc_context, llama_ngram_cache & nc_dynamic, llama_ngram_cache & nc_static common_ngram_cache & nc_context, common_ngram_cache & nc_dynamic, common_ngram_cache & nc_static
) { ) {
GGML_ASSERT(draft.size() == 1); GGML_ASSERT(draft.size() == 1);
const int inp_size = inp.size(); const int inp_size = inp.size();
@ -157,21 +157,21 @@ void llama_ngram_cache_draft(
llama_token drafted_token = -1; llama_token drafted_token = -1;
const int ngram_start_static = inp_size-LLAMA_NGRAM_STATIC + draft.size()-1; const int ngram_start_static = inp_size-LLAMA_NGRAM_STATIC + draft.size()-1;
llama_ngram ngram_static; common_ngram ngram_static;
for (int j = ngram_start_static; j < ngram_start_static + LLAMA_NGRAM_STATIC; ++j) { for (int j = ngram_start_static; j < ngram_start_static + LLAMA_NGRAM_STATIC; ++j) {
ngram_static.tokens[j-ngram_start_static] = get_token(inp, draft, j); ngram_static.tokens[j-ngram_start_static] = get_token(inp, draft, j);
} }
llama_ngram_cache::iterator part_static_it = nc_static.find(ngram_static); common_ngram_cache::iterator part_static_it = nc_static.find(ngram_static);
llama_ngram_cache_part part_static; common_ngram_cache_part part_static;
if (part_static_it != nc_static.end()) { if (part_static_it != nc_static.end()) {
part_static = part_static_it->second; part_static = part_static_it->second;
} }
// cd = context + dynamic // cd = context + dynamic
std::vector<llama_ngram> ngrams_cd; std::vector<common_ngram> ngrams_cd;
for (int ngram_size_cd = ngram_min; ngram_size_cd <= ngram_max; ++ngram_size_cd) { for (int ngram_size_cd = ngram_min; ngram_size_cd <= ngram_max; ++ngram_size_cd) {
const int ngram_start_cd = inp_size-ngram_size_cd + draft.size()-1; const int ngram_start_cd = inp_size-ngram_size_cd + draft.size()-1;
llama_ngram ngram_cd; common_ngram ngram_cd;
for (int j = ngram_start_cd; j < ngram_start_cd + ngram_size_cd; ++j) { for (int j = ngram_start_cd; j < ngram_start_cd + ngram_size_cd; ++j) {
ngram_cd.tokens[j-ngram_start_cd] = get_token(inp, draft, j); ngram_cd.tokens[j-ngram_start_cd] = get_token(inp, draft, j);
} }
@ -196,16 +196,16 @@ void llama_ngram_cache_draft(
} }
} }
void llama_ngram_cache_save(llama_ngram_cache & ngram_cache, std::string & filename) { void common_ngram_cache_save(common_ngram_cache & ngram_cache, std::string & filename) {
std::ofstream file_out(filename, std::ios::binary); std::ofstream file_out(filename, std::ios::binary);
for (std::pair<llama_ngram, llama_ngram_cache_part> item : ngram_cache) { for (std::pair<common_ngram, common_ngram_cache_part> item : ngram_cache) {
const llama_ngram ngram = item.first; const common_ngram ngram = item.first;
llama_ngram_cache_part token_counts = item.second; common_ngram_cache_part token_counts = item.second;
GGML_ASSERT(!token_counts.empty()); GGML_ASSERT(!token_counts.empty());
const int32_t ntokens = token_counts.size(); const int32_t ntokens = token_counts.size();
GGML_ASSERT(ntokens > 0); GGML_ASSERT(ntokens > 0);
file_out.write(reinterpret_cast<const char *>(&ngram), sizeof(llama_ngram)); file_out.write(reinterpret_cast<const char *>(&ngram), sizeof(common_ngram));
file_out.write(reinterpret_cast<const char *>(&ntokens), sizeof(int32_t)); file_out.write(reinterpret_cast<const char *>(&ntokens), sizeof(int32_t));
for (std::pair<llama_token, int32_t> item2 : token_counts) { for (std::pair<llama_token, int32_t> item2 : token_counts) {
const llama_token token = item2.first; const llama_token token = item2.first;
@ -219,14 +219,14 @@ void llama_ngram_cache_save(llama_ngram_cache & ngram_cache, std::string & filen
} }
llama_ngram_cache llama_ngram_cache_load(std::string & filename) { common_ngram_cache common_ngram_cache_load(std::string & filename) {
std::ifstream hashmap_file(filename, std::ios::binary); std::ifstream hashmap_file(filename, std::ios::binary);
if (!hashmap_file) { if (!hashmap_file) {
throw std::ifstream::failure("Unable to open file " + filename); throw std::ifstream::failure("Unable to open file " + filename);
} }
llama_ngram_cache ngram_cache; common_ngram_cache ngram_cache;
llama_ngram ngram; common_ngram ngram;
int32_t ntokens; int32_t ntokens;
llama_token token; llama_token token;
int32_t count; int32_t count;
@ -235,11 +235,11 @@ llama_ngram_cache llama_ngram_cache_load(std::string & filename) {
char * ntokensc = reinterpret_cast<char*>(&ntokens); char * ntokensc = reinterpret_cast<char*>(&ntokens);
char * tokenc = reinterpret_cast<char*>(&token); char * tokenc = reinterpret_cast<char*>(&token);
char * countc = reinterpret_cast<char*>(&count); char * countc = reinterpret_cast<char*>(&count);
while(hashmap_file.read(ngramc, sizeof(llama_ngram))) { while(hashmap_file.read(ngramc, sizeof(common_ngram))) {
GGML_ASSERT(!hashmap_file.eof()); GGML_ASSERT(!hashmap_file.eof());
GGML_ASSERT(hashmap_file.read(ntokensc, sizeof(int32_t))); GGML_ASSERT(hashmap_file.read(ntokensc, sizeof(int32_t)));
GGML_ASSERT(ntokens > 0); GGML_ASSERT(ntokens > 0);
llama_ngram_cache_part token_counts; common_ngram_cache_part token_counts;
for (int i = 0; i < ntokens; ++i) { for (int i = 0; i < ntokens; ++i) {
GGML_ASSERT(!hashmap_file.eof()); GGML_ASSERT(!hashmap_file.eof());
@ -257,12 +257,12 @@ llama_ngram_cache llama_ngram_cache_load(std::string & filename) {
return ngram_cache; return ngram_cache;
} }
void llama_ngram_cache_merge(llama_ngram_cache & ngram_cache_target, llama_ngram_cache & ngram_cache_add) { void common_ngram_cache_merge(common_ngram_cache & ngram_cache_target, common_ngram_cache & ngram_cache_add) {
for (std::pair<llama_ngram, llama_ngram_cache_part> ngram_part : ngram_cache_add) { for (std::pair<common_ngram, common_ngram_cache_part> ngram_part : ngram_cache_add) {
const llama_ngram ngram = ngram_part.first; const common_ngram ngram = ngram_part.first;
llama_ngram_cache_part part = ngram_part.second; common_ngram_cache_part part = ngram_part.second;
llama_ngram_cache::iterator part_merged_it = ngram_cache_target.find(ngram); common_ngram_cache::iterator part_merged_it = ngram_cache_target.find(ngram);
if (part_merged_it == ngram_cache_target.end()) { if (part_merged_it == ngram_cache_target.end()) {
ngram_cache_target.emplace(ngram, part); ngram_cache_target.emplace(ngram, part);
continue; continue;
@ -273,7 +273,7 @@ void llama_ngram_cache_merge(llama_ngram_cache & ngram_cache_target, llama_ngram
const int32_t count = token_count.second; const int32_t count = token_count.second;
GGML_ASSERT(count > 0); GGML_ASSERT(count > 0);
llama_ngram_cache_part::iterator token_count_merged_it = part_merged_it->second.find(token); common_ngram_cache_part::iterator token_count_merged_it = part_merged_it->second.find(token);
if (token_count_merged_it == part_merged_it->second.end()) { if (token_count_merged_it == part_merged_it->second.end()) {
part_merged_it->second.emplace(token, count); part_merged_it->second.emplace(token, count);
continue; continue;

38
common/ngram-cache.h
View File

@ -12,22 +12,22 @@
// Data structures to map n-grams to empirical token probabilities: // Data structures to map n-grams to empirical token probabilities:
struct llama_ngram { struct common_ngram {
llama_token tokens[LLAMA_NGRAM_MAX]; llama_token tokens[LLAMA_NGRAM_MAX];
llama_ngram() { common_ngram() {
for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) { for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
tokens[i] = -1; tokens[i] = -1;
} }
} }
llama_ngram(const llama_token * input, const int ngram_size) { common_ngram(const llama_token * input, const int ngram_size) {
for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) { for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
tokens[i] = i < ngram_size ? input[i] : -1; tokens[i] = i < ngram_size ? input[i] : -1;
} }
} }
bool operator==(const llama_ngram & other) const { bool operator==(const common_ngram & other) const {
for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) { for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
if (tokens[i] != other.tokens[i]) { if (tokens[i] != other.tokens[i]) {
return false; return false;
@ -37,28 +37,28 @@ struct llama_ngram {
} }
}; };
struct llama_token_hash_function { struct common_token_hash_function {
size_t operator()(const llama_token token) const { size_t operator()(const llama_token token) const {
// see https://probablydance.com/2018/06/16/fibonacci-hashing-the-optimization-that-the-world-forgot-or-a-better-alternative-to-integer-modulo/ // see https://probablydance.com/2018/06/16/fibonacci-hashing-the-optimization-that-the-world-forgot-or-a-better-alternative-to-integer-modulo/
return token * 11400714819323198485llu; return token * 11400714819323198485llu;
} }
}; };
struct llama_ngram_hash_function { struct common_ngram_hash_function {
size_t operator()(const llama_ngram & ngram) const { size_t operator()(const common_ngram & ngram) const {
size_t hash = llama_token_hash_function{}(ngram.tokens[0]); size_t hash = common_token_hash_function{}(ngram.tokens[0]);
for (int i = 1; i < LLAMA_NGRAM_MAX; ++i) { for (int i = 1; i < LLAMA_NGRAM_MAX; ++i) {
hash ^= llama_token_hash_function{}(ngram.tokens[i]); hash ^= common_token_hash_function{}(ngram.tokens[i]);
} }
return hash; return hash;
} }
}; };
// token -> number of times token has been seen // token -> number of times token has been seen
typedef std::unordered_map<llama_token, int32_t> llama_ngram_cache_part; typedef std::unordered_map<llama_token, int32_t> common_ngram_cache_part;
// n-gram -> empirical distribution of following tokens // n-gram -> empirical distribution of following tokens
typedef std::unordered_map<llama_ngram, llama_ngram_cache_part, llama_ngram_hash_function> llama_ngram_cache; typedef std::unordered_map<common_ngram, common_ngram_cache_part, common_ngram_hash_function> common_ngram_cache;
// Update an ngram cache with tokens. // Update an ngram cache with tokens.
@ -70,8 +70,8 @@ typedef std::unordered_map<llama_ngram, llama_ngram_cache_part, llama_ngram_hash
// //
// In order to get correct results inp_data can ONLY BE APPENDED TO. // In order to get correct results inp_data can ONLY BE APPENDED TO.
// Changes in the middle need a complete rebuild. // Changes in the middle need a complete rebuild.
void llama_ngram_cache_update( void common_ngram_cache_update(
llama_ngram_cache & ngram_cache, int ngram_min, int ngram_max, std::vector<llama_token> & inp_data, int nnew, bool print_progress); common_ngram_cache & ngram_cache, int ngram_min, int ngram_max, std::vector<llama_token> & inp_data, int nnew, bool print_progress);
// Try to draft tokens from ngram caches. // Try to draft tokens from ngram caches.
// inp: the tokens generated so far. // inp: the tokens generated so far.
@ -81,21 +81,21 @@ void llama_ngram_cache_update(
// nc_context: ngram cache based on current context. // nc_context: ngram cache based on current context.
// nc_dynamic: ngram cache based on previous user generations. // nc_dynamic: ngram cache based on previous user generations.
// nc_static: ngram cache generated from a large text corpus, used for validation. // nc_static: ngram cache generated from a large text corpus, used for validation.
void llama_ngram_cache_draft( void common_ngram_cache_draft(
std::vector<llama_token> & inp, std::vector<llama_token> & draft, int n_draft, int ngram_min, int ngram_max, std::vector<llama_token> & inp, std::vector<llama_token> & draft, int n_draft, int ngram_min, int ngram_max,
llama_ngram_cache & nc_context, llama_ngram_cache & nc_dynamic, llama_ngram_cache & nc_static); common_ngram_cache & nc_context, common_ngram_cache & nc_dynamic, common_ngram_cache & nc_static);
// Save an ngram cache to a file. // Save an ngram cache to a file.
// ngram_cache: the ngram cache to save. // ngram_cache: the ngram cache to save.
// filename: the path under which to save the ngram cache. // filename: the path under which to save the ngram cache.
void llama_ngram_cache_save(llama_ngram_cache & ngram_cache, std::string & filename); void common_ngram_cache_save(common_ngram_cache & ngram_cache, std::string & filename);
// Load an ngram cache saved with llama_ngram_cache_save. // Load an ngram cache saved with common_ngram_cache_save.
// filename: the path from which to load the ngram cache. // filename: the path from which to load the ngram cache.
// returns: an ngram cache containing the information saved to filename. // returns: an ngram cache containing the information saved to filename.
llama_ngram_cache llama_ngram_cache_load(std::string & filename); common_ngram_cache common_ngram_cache_load(std::string & filename);
// Merge two ngram caches. // Merge two ngram caches.
// ngram_cache_target: the ngram cache to which to add the information from ngram_cache_add. // ngram_cache_target: the ngram cache to which to add the information from ngram_cache_add.
// ngram_cache_add: the ngram cache to add to ngram_cache_target. // ngram_cache_add: the ngram cache to add to ngram_cache_target.
void llama_ngram_cache_merge(llama_ngram_cache & ngram_cache_target, llama_ngram_cache & ngram_cache_add); void common_ngram_cache_merge(common_ngram_cache & ngram_cache_target, common_ngram_cache & ngram_cache_add);

136
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) {
@ -358,63 +358,63 @@ std::string gpt_sampler_prev_str(gpt_sampler * gsmpl, llama_context * ctx_main,
GGML_ASSERT(id != LLAMA_TOKEN_NULL && "null token in the sampling history - should not happen"); GGML_ASSERT(id != LLAMA_TOKEN_NULL && "null token in the sampling history - should not happen");
result += llama_token_to_piece(ctx_main, id); result += common_token_to_piece(ctx_main, id);
} }
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);

20
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 = llama_model_params_from_gpt_params(params); llama_model_params model_params = common_model_params_to_llama(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 = llama_context_params_from_gpt_params(params); llama_context_params ctx_params = common_context_params_to_llama(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());
@ -92,7 +92,7 @@ int main(int argc, char ** argv) {
// warm up // warm up
{ {
for (int i = 0; i < 16; ++i) { for (int i = 0; i < 16; ++i) {
llama_batch_add(batch, 0, i, { 0 }, false); common_batch_add(batch, 0, i, { 0 }, false);
} }
if (!decode_helper(ctx, batch, ctx_params.n_batch)) { if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
@ -122,11 +122,11 @@ int main(int argc, char ** argv) {
continue; continue;
} }
llama_batch_clear(batch); common_batch_clear(batch);
for (int i = 0; i < pp; ++i) { for (int i = 0; i < pp; ++i) {
for (int j = 0; j < (is_pp_shared ? 1 : pl); ++j) { for (int j = 0; j < (is_pp_shared ? 1 : pl); ++j) {
llama_batch_add(batch, 0, i, { j }, false); common_batch_add(batch, 0, i, { j }, false);
} }
} }
batch.logits[batch.n_tokens - 1] = true; batch.logits[batch.n_tokens - 1] = true;
@ -151,10 +151,10 @@ int main(int argc, char ** argv) {
const auto t_tg_start = ggml_time_us(); const auto t_tg_start = ggml_time_us();
for (int i = 0; i < tg; ++i) { for (int i = 0; i < tg; ++i) {
llama_batch_clear(batch); common_batch_clear(batch);
for (int j = 0; j < pl; ++j) { for (int j = 0; j < pl; ++j) {
llama_batch_add(batch, 0, pp + i, { j }, true); common_batch_add(batch, 0, pp + i, { j }, true);
} }
if (!decode_helper(ctx, batch, ctx_params.n_batch)) { if (!decode_helper(ctx, batch, ctx_params.n_batch)) {

28
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 = llama_model_params_from_gpt_params(params); llama_model_params model_params = common_model_params_to_llama(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);
@ -51,13 +51,13 @@ int main(int argc, char ** argv) {
// tokenize the prompt // tokenize the prompt
std::vector<llama_token> tokens_list; std::vector<llama_token> tokens_list;
tokens_list = ::llama_tokenize(model, params.prompt, true); tokens_list = common_tokenize(model, params.prompt, true);
const int n_kv_req = tokens_list.size() + (n_predict - tokens_list.size())*n_parallel; const int n_kv_req = tokens_list.size() + (n_predict - tokens_list.size())*n_parallel;
// initialize the context // initialize the context
llama_context_params ctx_params = llama_context_params_from_gpt_params(params); llama_context_params ctx_params = common_context_params_to_llama(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);
@ -94,7 +94,7 @@ int main(int argc, char ** argv) {
LOG("\n"); LOG("\n");
for (auto id : tokens_list) { for (auto id : tokens_list) {
LOG("%s", llama_token_to_piece(ctx, id).c_str()); LOG("%s", common_token_to_piece(ctx, id).c_str());
} }
// create a llama_batch // create a llama_batch
@ -108,7 +108,7 @@ int main(int argc, char ** argv) {
// evaluate the initial prompt // evaluate the initial prompt
for (size_t i = 0; i < tokens_list.size(); ++i) { for (size_t i = 0; i < tokens_list.size(); ++i) {
llama_batch_add(batch, tokens_list[i], i, seq_ids, false); common_batch_add(batch, tokens_list[i], i, seq_ids, false);
} }
GGML_ASSERT(batch.n_tokens == (int) tokens_list.size()); GGML_ASSERT(batch.n_tokens == (int) tokens_list.size());
@ -123,8 +123,8 @@ int main(int argc, char ** argv) {
decoder_start_token_id = llama_token_bos(model); decoder_start_token_id = llama_token_bos(model);
} }
llama_batch_clear(batch); common_batch_clear(batch);
llama_batch_add(batch, decoder_start_token_id, 0, seq_ids, false); common_batch_add(batch, decoder_start_token_id, 0, seq_ids, false);
} }
// llama_decode will output logits only for the last token of the prompt // llama_decode will output logits only for the last token of the prompt
@ -161,7 +161,7 @@ int main(int argc, char ** argv) {
while (n_cur <= n_predict) { while (n_cur <= n_predict) {
// prepare the next batch // prepare the next batch
llama_batch_clear(batch); common_batch_clear(batch);
// sample the next token for each parallel sequence / stream // sample the next token for each parallel sequence / stream
for (int32_t i = 0; i < n_parallel; ++i) { for (int32_t i = 0; i < n_parallel; ++i) {
@ -185,15 +185,15 @@ int main(int argc, char ** argv) {
// if there is only one stream, we print immediately to stdout // if there is only one stream, we print immediately to stdout
if (n_parallel == 1) { if (n_parallel == 1) {
LOG("%s", llama_token_to_piece(ctx, new_token_id).c_str()); LOG("%s", common_token_to_piece(ctx, new_token_id).c_str());
} }
streams[i] += llama_token_to_piece(ctx, new_token_id); streams[i] += common_token_to_piece(ctx, new_token_id);
i_batch[i] = batch.n_tokens; i_batch[i] = batch.n_tokens;
// push this new token for next evaluation // push this new token for next evaluation
llama_batch_add(batch, new_token_id, n_cur, { i }, true); common_batch_add(batch, new_token_id, n_cur, { i }, true);
n_decode += 1; n_decode += 1;
} }

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)) {

16
examples/cvector-generator/cvector-generator.cpp
View File

@ -31,7 +31,7 @@ template <class Iter>
static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) { static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
std::string ret; std::string ret;
for (; begin != end; ++begin) { for (; begin != end; ++begin) {
ret += llama_token_to_piece(ctx, *begin); ret += common_token_to_piece(ctx, *begin);
} }
return ret; return ret;
@ -272,8 +272,8 @@ struct tokenized_prompt {
tokenized_prompt(llama_context * ctx, std::string pos, std::string neg) { tokenized_prompt(llama_context * ctx, std::string pos, std::string neg) {
const bool add_bos = llama_add_bos_token(llama_get_model(ctx)); const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
tokens_pos = ::llama_tokenize(ctx, pos, add_bos, true); tokens_pos = common_tokenize(ctx, pos, add_bos, true);
tokens_neg = ::llama_tokenize(ctx, neg, add_bos, true); tokens_neg = common_tokenize(ctx, neg, add_bos, true);
max_seq_len = std::max(tokens_pos.size(), tokens_neg.size()); max_seq_len = std::max(tokens_pos.size(), tokens_neg.size());
padding_seq(ctx, tokens_pos, max_seq_len); padding_seq(ctx, tokens_pos, max_seq_len);
padding_seq(ctx, tokens_neg, max_seq_len); padding_seq(ctx, tokens_neg, max_seq_len);
@ -281,7 +281,7 @@ struct tokenized_prompt {
void padding_seq(llama_context * ctx, std::vector<llama_token> & tokens, size_t len) { void padding_seq(llama_context * ctx, std::vector<llama_token> & tokens, size_t len) {
// TODO: customize padding token // TODO: customize padding token
std::vector<llama_token> pad_tokens = ::llama_tokenize(ctx, " ", false); std::vector<llama_token> pad_tokens = common_tokenize(ctx, " ", false);
llama_token pad_tok = pad_tokens.back(); llama_token pad_tok = pad_tokens.back();
while (tokens.size() < len) { while (tokens.size() < len) {
tokens.push_back(pad_tok); tokens.push_back(pad_tok);
@ -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
llama_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_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;

24
examples/embedding/embedding.cpp
View File

@ -28,7 +28,7 @@ static std::vector<std::string> split_lines(const std::string & s, const std::st
static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, llama_seq_id seq_id) { static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, llama_seq_id seq_id) {
size_t n_tokens = tokens.size(); size_t n_tokens = tokens.size();
for (size_t i = 0; i < n_tokens; i++) { for (size_t i = 0; i < n_tokens; i++) {
llama_batch_add(batch, tokens[i], i, { seq_id }, true); common_batch_add(batch, tokens[i], i, { seq_id }, true);
} }
} }
@ -74,18 +74,18 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
} }
float * out = output + embd_pos * n_embd; float * out = output + embd_pos * n_embd;
llama_embd_normalize(embd, out, n_embd, embd_norm); common_embd_normalize(embd, out, n_embd, embd_norm);
} }
} }
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
llama_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_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
@ -135,7 +135,7 @@ int main(int argc, char ** argv) {
// tokenize the prompts and trim // tokenize the prompts and trim
std::vector<std::vector<int32_t>> inputs; std::vector<std::vector<int32_t>> inputs;
for (const auto & prompt : prompts) { for (const auto & prompt : prompts) {
auto inp = ::llama_tokenize(ctx, prompt, true, true); auto inp = common_tokenize(ctx, prompt, true, true);
if (inp.size() > n_batch) { if (inp.size() > n_batch) {
LOG_ERR("%s: number of tokens in input line (%lld) exceeds batch size (%lld), increase batch size and re-run\n", LOG_ERR("%s: number of tokens in input line (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
__func__, (long long int) inp.size(), (long long int) n_batch); __func__, (long long int) inp.size(), (long long int) n_batch);
@ -159,7 +159,7 @@ int main(int argc, char ** argv) {
LOG_INF("%s: prompt %d: '%s'\n", __func__, i, prompts[i].c_str()); LOG_INF("%s: prompt %d: '%s'\n", __func__, i, prompts[i].c_str());
LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, inputs[i].size()); LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, inputs[i].size());
for (int j = 0; j < (int) inputs[i].size(); j++) { for (int j = 0; j < (int) inputs[i].size(); j++) {
LOG("%6d -> '%s'\n", inputs[i][j], llama_token_to_piece(ctx, inputs[i][j]).c_str()); LOG("%6d -> '%s'\n", inputs[i][j], common_token_to_piece(ctx, inputs[i][j]).c_str());
} }
LOG("\n\n"); LOG("\n\n");
} }
@ -199,7 +199,7 @@ int main(int argc, char ** argv) {
batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize); batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
e += pooling_type == LLAMA_POOLING_TYPE_NONE ? batch.n_tokens : s; e += pooling_type == LLAMA_POOLING_TYPE_NONE ? batch.n_tokens : s;
s = 0; s = 0;
llama_batch_clear(batch); common_batch_clear(batch);
} }
// add to batch // add to batch
@ -263,7 +263,7 @@ int main(int argc, char ** argv) {
LOG("\n"); LOG("\n");
for (int i = 0; i < n_prompts; i++) { for (int i = 0; i < n_prompts; i++) {
for (int j = 0; j < n_prompts; j++) { for (int j = 0; j < n_prompts; j++) {
float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd); float sim = common_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
LOG("%6.2f ", sim); LOG("%6.2f ", sim);
} }
LOG("%1.10s", prompts[i].c_str()); LOG("%1.10s", prompts[i].c_str());
@ -296,7 +296,7 @@ int main(int argc, char ** argv) {
for (int i = 0;;) { // at least two iteration (n_embd_count > 1) for (int i = 0;;) { // at least two iteration (n_embd_count > 1)
LOG(" ["); LOG(" [");
for (int j = 0;;) { // at least two iteration (n_embd_count > 1) for (int j = 0;;) { // at least two iteration (n_embd_count > 1)
float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd); float sim = common_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
LOG("%6.2f", sim); LOG("%6.2f", sim);
j++; j++;
if (j < n_embd_count) LOG(", "); else break; if (j < n_embd_count) LOG(", "); else break;

14
examples/eval-callback/eval-callback.cpp
View File

@ -126,10 +126,10 @@ 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 = ::llama_tokenize(ctx, params.prompt, add_bos); std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, add_bos);
if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) { if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
LOG_ERR("%s : failed to eval\n", __func__); LOG_ERR("%s : failed to eval\n", __func__);
@ -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
llama_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_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");
} }

6
examples/export-lora/export-lora.cpp
View File

@ -128,7 +128,7 @@ struct lora_merge_ctx {
lora_merge_ctx( lora_merge_ctx(
std::string & base_fname, std::string & base_fname,
std::vector<llama_lora_adapter_info> & lora_files, std::vector<common_lora_adapter_info> & lora_files,
std::string & outfile, std::string & outfile,
int n_threads) : base_model(base_fname, 0), n_threads(n_threads), fout(outfile, std::ios::binary) { int n_threads) : base_model(base_fname, 0), n_threads(n_threads), fout(outfile, std::ios::binary) {
fout.exceptions(std::ofstream::failbit); // fail fast on write errors fout.exceptions(std::ofstream::failbit); // fail fast on write errors
@ -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;
} }

14
examples/gen-docs/gen-docs.cpp
View File

@ -11,7 +11,7 @@ static void write_table_header(std::ofstream & file) {
file << "| -------- | ----------- |\n"; file << "| -------- | ----------- |\n";
} }
static void write_table_entry(std::ofstream & file, const llama_arg & opt) { static void write_table_entry(std::ofstream & file, const common_arg & opt) {
file << "| `"; file << "| `";
// args // args
for (const auto & arg : opt.args) { for (const auto & arg : opt.args) {
@ -40,7 +40,7 @@ static void write_table_entry(std::ofstream & file, const llama_arg & opt) {
file << "` | " << md_help << " |\n"; file << "` | " << md_help << " |\n";
} }
static void write_table(std::ofstream & file, std::vector<llama_arg *> & opts) { static void write_table(std::ofstream & file, std::vector<common_arg *> & opts) {
write_table_header(file); write_table_header(file);
for (const auto & opt : opts) { for (const auto & opt : opts) {
write_table_entry(file, *opt); write_table_entry(file, *opt);
@ -50,12 +50,12 @@ static void write_table(std::ofstream & file, std::vector<llama_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<llama_arg *> common_options; std::vector<common_arg *> common_options;
std::vector<llama_arg *> sparam_options; std::vector<common_arg *> sparam_options;
std::vector<llama_arg *> specific_options; std::vector<common_arg *> specific_options;
for (auto & opt : ctx_arg.options) { for (auto & opt : ctx_arg.options) {
// in case multiple LLAMA_EXAMPLE_* are set, we prioritize the LLAMA_EXAMPLE_* matching current example // in case multiple LLAMA_EXAMPLE_* are set, we prioritize the LLAMA_EXAMPLE_* matching current example
if (opt.is_sparam) { if (opt.is_sparam) {

36
examples/gritlm/gritlm.cpp
View File

@ -15,11 +15,11 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
llama_batch batch = llama_batch_init(llama_n_batch(ctx), 0, 1); llama_batch batch = llama_batch_init(llama_n_batch(ctx), 0, 1);
for (uint64_t i = 0; i < sentences.size(); i++) { for (uint64_t i = 0; i < sentences.size(); i++) {
llama_batch_clear(batch); common_batch_clear(batch);
const std::string input_string = instruction + sentences[i]; const std::string input_string = instruction + sentences[i];
std::vector<llama_token> inputs = llama_tokenize(model, input_string, true, false); std::vector<llama_token> inputs = common_tokenize(model, input_string, true, false);
const int32_t n_toks = inputs.size(); const int32_t n_toks = inputs.size();
@ -28,7 +28,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
// inputs.push_back(llama_token_eos(model)); // inputs.push_back(llama_token_eos(model));
// we want to ignore instruction tokens for mean pooling // we want to ignore instruction tokens for mean pooling
const int32_t n_inst = llama_tokenize(model, instruction, true, false).size(); const int32_t n_inst = common_tokenize(model, instruction, true, false).size();
#ifdef GRIT_DEBUG #ifdef GRIT_DEBUG
// debug tokens - should be matching as referenced in the GritLM sample // debug tokens - should be matching as referenced in the GritLM sample
@ -40,7 +40,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
// add input to batch (this increments n_tokens) // add input to batch (this increments n_tokens)
for (int32_t j = 0; j < n_toks; j++) { for (int32_t j = 0; j < n_toks; j++) {
llama_batch_add(batch, inputs[j], j, { 0 }, j >= n_inst); common_batch_add(batch, inputs[j], j, { 0 }, j >= n_inst);
} }
// clear previous kv_cache values (irrelevant for embeddings) // clear previous kv_cache values (irrelevant for embeddings)
@ -75,7 +75,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
} }
std::vector<float> emb_norm(emb_unorm.size()); std::vector<float> emb_norm(emb_unorm.size());
llama_embd_normalize(emb_unorm.data(), emb_norm.data(), n_embd); common_embd_normalize(emb_unorm.data(), emb_norm.data(), n_embd);
result.push_back(emb_norm); result.push_back(emb_norm);
#ifdef GRIT_DEBUG #ifdef GRIT_DEBUG
@ -105,16 +105,16 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1); llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
std::vector<llama_token> inputs = llama_tokenize(model, prompt, false, true); std::vector<llama_token> inputs = common_tokenize(model, prompt, false, true);
int32_t i_current_token = 0; int32_t i_current_token = 0;
while (true) { while (true) {
llama_batch_clear(bat); common_batch_clear(bat);
{ {
const int32_t n_inputs = inputs.size(); const int32_t n_inputs = inputs.size();
for (int32_t i = 0; i < n_inputs; i++) { for (int32_t i = 0; i < n_inputs; i++) {
llama_batch_add(bat, inputs[i], i_current_token++, { 0 }, i == n_inputs - 1); common_batch_add(bat, inputs[i], i_current_token++, { 0 }, i == n_inputs - 1);
} }
} }
inputs.clear(); inputs.clear();
@ -127,7 +127,7 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
break; break;
} }
std::string piece = llama_token_to_piece(ctx, token); std::string piece = common_token_to_piece(ctx, token);
if (stream) { if (stream) {
std::printf("%s", piece.c_str()); std::printf("%s", piece.c_str());
std::fflush(stdout); std::fflush(stdout);
@ -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 = llama_model_params_from_gpt_params(params); llama_model_params mparams = common_model_params_to_llama(params);
llama_context_params cparams = llama_context_params_from_gpt_params(params); llama_context_params cparams = common_context_params_to_llama(params);
llama_backend_init(); llama_backend_init();
@ -199,10 +199,10 @@ int main(int argc, char * argv[]) {
const int n_embd = llama_n_embd(model); const int n_embd = llama_n_embd(model);
const float cosine_sim_q0_d0 = llama_embd_similarity_cos(q_rep[0].data(), d_rep[0].data(), n_embd); const float cosine_sim_q0_d0 = common_embd_similarity_cos(q_rep[0].data(), d_rep[0].data(), n_embd);
const float cosine_sim_q0_d1 = llama_embd_similarity_cos(q_rep[0].data(), d_rep[1].data(), n_embd); const float cosine_sim_q0_d1 = common_embd_similarity_cos(q_rep[0].data(), d_rep[1].data(), n_embd);
const float cosine_sim_q1_d0 = llama_embd_similarity_cos(q_rep[1].data(), d_rep[0].data(), n_embd); const float cosine_sim_q1_d0 = common_embd_similarity_cos(q_rep[1].data(), d_rep[0].data(), n_embd);
const float cosine_sim_q1_d1 = llama_embd_similarity_cos(q_rep[1].data(), d_rep[1].data(), n_embd); const float cosine_sim_q1_d1 = common_embd_similarity_cos(q_rep[1].data(), d_rep[1].data(), n_embd);
std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[0].c_str(), cosine_sim_q0_d0); std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[0].c_str(), cosine_sim_q0_d0);
std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[1].c_str(), cosine_sim_q0_d1); std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[1].c_str(), cosine_sim_q0_d1);

18
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);
@ -436,7 +436,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
auto tim1 = std::chrono::high_resolution_clock::now(); auto tim1 = std::chrono::high_resolution_clock::now();
LOG_INF("%s: tokenizing the input ..\n", __func__); LOG_INF("%s: tokenizing the input ..\n", __func__);
std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, true); std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, true);
auto tim2 = std::chrono::high_resolution_clock::now(); auto tim2 = std::chrono::high_resolution_clock::now();
LOG_INF("%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count()); LOG_INF("%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
@ -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
llama_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_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)) {

66
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__);
llama_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_params(params);
model = llama_init.model; model = llama_init.model;
ctx = llama_init.context; ctx = llama_init.context;
@ -195,15 +195,15 @@ 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));
std::vector<llama_token> embd_inp; std::vector<llama_token> embd_inp;
std::vector<llama_token> embd_end; std::vector<llama_token> embd_end;
std::vector<llama_token> inp_pfx = ::llama_tokenize(ctx, params.input_prefix, false); std::vector<llama_token> inp_pfx = common_tokenize(ctx, params.input_prefix, false);
std::vector<llama_token> inp_sfx = ::llama_tokenize(ctx, params.input_suffix, false); std::vector<llama_token> inp_sfx = common_tokenize(ctx, params.input_suffix, false);
GGML_ASSERT(llama_token_prefix(model) >= 0); GGML_ASSERT(llama_token_prefix(model) >= 0);
GGML_ASSERT(llama_token_suffix(model) >= 0); GGML_ASSERT(llama_token_suffix(model) >= 0);
@ -257,13 +257,13 @@ int main(int argc, char ** argv) {
LOG_INF("%s: prompt: '%s'\n", __func__, params.prompt.c_str()); LOG_INF("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size()); LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
for (int i = 0; i < (int) embd_inp.size(); i++) { for (int i = 0; i < (int) embd_inp.size(); i++) {
LOG_INF("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str()); LOG_INF("%6d -> '%s'\n", embd_inp[i], common_token_to_piece(ctx, embd_inp[i]).c_str());
} }
if (params.n_keep > 0) { if (params.n_keep > 0) {
LOG_INF("%s: static prompt based on n_keep: '", __func__); LOG_INF("%s: static prompt based on n_keep: '", __func__);
for (int i = 0; i < params.n_keep; i++) { for (int i = 0; i < params.n_keep; i++) {
LOG_CNT("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str()); LOG_CNT("%s", common_token_to_piece(ctx, embd_inp[i]).c_str());
} }
LOG_CNT("'\n"); LOG_CNT("'\n");
} }
@ -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) {
@ -446,7 +446,7 @@ int main(int argc, char ** argv) {
// display text // display text
if (input_echo) { if (input_echo) {
for (auto id : embd) { for (auto id : embd) {
const std::string token_str = llama_token_to_piece(ctx, id); const std::string token_str = common_token_to_piece(ctx, id);
LOG("%s", token_str.c_str()); LOG("%s", token_str.c_str());
if (embd.size() > 1) { if (embd.size() > 1) {
@ -465,10 +465,10 @@ 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", llama_token_to_piece(ctx, llama_token_eot(model)).c_str()); LOG("%s", common_token_to_piece(ctx, llama_token_eot(model)).c_str());
} }
LOG("\n"); LOG("\n");
console::set_display(console::user_input); console::set_display(console::user_input);
@ -505,8 +505,8 @@ int main(int argc, char ** argv) {
} }
// tokenize new prefix and suffix // tokenize new prefix and suffix
std::vector<llama_token> inp_pfx = ::llama_tokenize(ctx, params.input_prefix, false); std::vector<llama_token> inp_pfx = common_tokenize(ctx, params.input_prefix, false);
std::vector<llama_token> inp_sfx = ::llama_tokenize(ctx, params.input_suffix, false); std::vector<llama_token> inp_sfx = common_tokenize(ctx, params.input_suffix, false);
inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model)); inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model)); inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
@ -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) {
@ -579,7 +579,7 @@ int main(int argc, char ** argv) {
const size_t original_size = embd_inp.size(); const size_t original_size = embd_inp.size();
const auto line_inp = ::llama_tokenize(ctx, buffer, false); const auto line_inp = common_tokenize(ctx, buffer, false);
LOG_DBG("input tokens: %s\n", string_from(ctx, line_inp).c_str()); LOG_DBG("input tokens: %s\n", string_from(ctx, line_inp).c_str());
embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end()); embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
@ -587,7 +587,7 @@ int main(int argc, char ** argv) {
for (size_t i = original_size; i < embd_inp.size(); ++i) { for (size_t i = original_size; i < embd_inp.size(); ++i) {
const llama_token token = embd_inp[i]; const llama_token token = embd_inp[i];
output_tokens.push_back(token); output_tokens.push_back(token);
output_ss << llama_token_to_piece(ctx, token); output_ss << common_token_to_piece(ctx, token);
} }
n_remain -= line_inp.size(); n_remain -= line_inp.size();
@ -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;
} }
@ -620,17 +620,17 @@ int main(int argc, char ** argv) {
} }
} }
if (!params.interactive && n_remain <= 0) { if (!params.interactive && n_remain <= 0) {
LOG("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str()); LOG("%s", common_token_to_piece(ctx, llama_token_eot(model)).c_str());
} }
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;

22
examples/llama.android/llama/src/main/cpp/llama-android.cpp
View File

@ -186,11 +186,11 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
for (nri = 0; nri < nr; nri++) { for (nri = 0; nri < nr; nri++) {
LOGi("Benchmark prompt processing (pp)"); LOGi("Benchmark prompt processing (pp)");
llama_batch_clear(*batch); common_batch_clear(*batch);
const int n_tokens = pp; const int n_tokens = pp;
for (i = 0; i < n_tokens; i++) { for (i = 0; i < n_tokens; i++) {
llama_batch_add(*batch, 0, i, { 0 }, false); common_batch_add(*batch, 0, i, { 0 }, false);
} }
batch->logits[batch->n_tokens - 1] = true; batch->logits[batch->n_tokens - 1] = true;
@ -210,9 +210,9 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
const auto t_tg_start = ggml_time_us(); const auto t_tg_start = ggml_time_us();
for (i = 0; i < tg; i++) { for (i = 0; i < tg; i++) {
llama_batch_clear(*batch); common_batch_clear(*batch);
for (j = 0; j < pl; j++) { for (j = 0; j < pl; j++) {
llama_batch_add(*batch, 0, i, { j }, true); common_batch_add(*batch, 0, i, { j }, true);
} }
LOGi("llama_decode() text generation: %d", i); LOGi("llama_decode() text generation: %d", i);
@ -357,7 +357,7 @@ Java_android_llama_cpp_LLamaAndroid_completion_1init(
const auto context = reinterpret_cast<llama_context *>(context_pointer); const auto context = reinterpret_cast<llama_context *>(context_pointer);
const auto batch = reinterpret_cast<llama_batch *>(batch_pointer); const auto batch = reinterpret_cast<llama_batch *>(batch_pointer);
const auto tokens_list = llama_tokenize(context, text, 1); const auto tokens_list = common_tokenize(context, text, 1);
auto n_ctx = llama_n_ctx(context); auto n_ctx = llama_n_ctx(context);
auto n_kv_req = tokens_list.size() + (n_len - tokens_list.size()); auto n_kv_req = tokens_list.size() + (n_len - tokens_list.size());
@ -369,14 +369,14 @@ Java_android_llama_cpp_LLamaAndroid_completion_1init(
} }
for (auto id : tokens_list) { for (auto id : tokens_list) {
LOGi("%s", llama_token_to_piece(context, id).c_str()); LOGi("%s", common_token_to_piece(context, id).c_str());
} }
llama_batch_clear(*batch); common_batch_clear(*batch);
// evaluate the initial prompt // evaluate the initial prompt
for (auto i = 0; i < tokens_list.size(); i++) { for (auto i = 0; i < tokens_list.size(); i++) {
llama_batch_add(*batch, tokens_list[i], i, { 0 }, false); common_batch_add(*batch, tokens_list[i], i, { 0 }, false);
} }
// llama_decode will output logits only for the last token of the prompt // llama_decode will output logits only for the last token of the prompt
@ -419,7 +419,7 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop(
return nullptr; return nullptr;
} }
auto new_token_chars = llama_token_to_piece(context, new_token_id); auto new_token_chars = common_token_to_piece(context, new_token_id);
cached_token_chars += new_token_chars; cached_token_chars += new_token_chars;
jstring new_token = nullptr; jstring new_token = nullptr;
@ -431,8 +431,8 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop(
new_token = env->NewStringUTF(""); new_token = env->NewStringUTF("");
} }
llama_batch_clear(*batch); common_batch_clear(*batch);
llama_batch_add(*batch, new_token_id, n_cur, { 0 }, true); common_batch_add(*batch, new_token_id, n_cur, { 0 }, true);
env->CallVoidMethod(intvar_ncur, la_int_var_inc); env->CallVoidMethod(intvar_ncur, la_int_var_inc);

44
examples/llava/llava-cli.cpp
View File

@ -37,21 +37,21 @@ static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){ static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
std::string str2 = str; std::string str2 = str;
std::vector<llama_token> embd_inp = ::llama_tokenize(ctx_llama, str2, add_bos, true); std::vector<llama_token> embd_inp = common_tokenize(ctx_llama, str2, add_bos, true);
eval_tokens(ctx_llama, embd_inp, n_batch, n_past); eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
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>";
} else { } else {
ret = llama_token_to_piece(ctx_llama, id); ret = common_token_to_piece(ctx_llama, id);
} }
eval_id(ctx_llama, id, n_past); eval_id(ctx_llama, id, n_past);
return ret.c_str(); return ret.c_str();
@ -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;
@ -159,16 +159,16 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
user_prompt = prompt.substr(image_pos + std::string("<image>").length()); user_prompt = prompt.substr(image_pos + std::string("<image>").length());
LOG_INF("system_prompt: %s\n", system_prompt.c_str()); LOG_INF("system_prompt: %s\n", system_prompt.c_str());
if (params->verbose_prompt) { if (params->verbose_prompt) {
auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, system_prompt, true, true); auto tmp = common_tokenize(ctx_llava->ctx_llama, system_prompt, true, true);
for (int i = 0; i < (int) tmp.size(); i++) { for (int i = 0; i < (int) tmp.size(); i++) {
LOG_INF("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str()); LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
} }
} }
LOG_INF("user_prompt: %s\n", user_prompt.c_str()); LOG_INF("user_prompt: %s\n", user_prompt.c_str());
if (params->verbose_prompt) { if (params->verbose_prompt) {
auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true); auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
for (int i = 0; i < (int) tmp.size(); i++) { for (int i = 0; i < (int) tmp.size(); i++) {
LOG_INF("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str()); LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
} }
} }
} else { } else {
@ -176,9 +176,9 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
system_prompt = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:"; system_prompt = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:";
user_prompt = prompt + "\nASSISTANT:"; user_prompt = prompt + "\nASSISTANT:";
if (params->verbose_prompt) { if (params->verbose_prompt) {
auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true); auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
for (int i = 0; i < (int) tmp.size(); i++) { for (int i = 0; i < (int) tmp.size(); i++) {
LOG_INF("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str()); LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
} }
} }
} }
@ -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 = llama_model_params_from_gpt_params(*params); llama_model_params model_params = common_model_params_to_llama(*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 = llama_context_params_from_gpt_params(*params); llama_context_params ctx_params = common_context_params_to_llama(*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);

40
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 = llama_model_params_from_gpt_params(*params); llama_model_params model_params = common_model_params_to_llama(*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 = llama_context_params_from_gpt_params(*params); llama_context_params ctx_params = common_context_params_to_llama(*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;
@ -114,7 +114,7 @@ static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){ static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
std::string str2 = str; std::string str2 = str;
std::vector<llama_token> embd_inp = ::llama_tokenize(ctx_llama, str2, add_bos, true); std::vector<llama_token> embd_inp = common_tokenize(ctx_llama, str2, add_bos, true);
return eval_tokens(ctx_llama, embd_inp, n_batch, n_past); return eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
} }
@ -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);
@ -162,22 +162,22 @@ 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>";
} else { } else {
ret = llama_token_to_piece(ctx_llama, id); ret = common_token_to_piece(ctx_llama, id);
} }
eval_id(ctx_llama, id, n_past); eval_id(ctx_llama, id, n_past);
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");

48
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
llama_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_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;
@ -65,7 +65,7 @@ int main(int argc, char ** argv) {
std::vector<llama_token> inp; std::vector<llama_token> inp;
std::vector<llama_token> all; std::vector<llama_token> all;
inp = ::llama_tokenize(ctx, params.prompt, true, true); inp = common_tokenize(ctx, params.prompt, true, true);
all = inp; all = inp;
const int max_context_size = llama_n_ctx(ctx); const int max_context_size = llama_n_ctx(ctx);
@ -79,7 +79,7 @@ int main(int argc, char ** argv) {
LOG("\n\n"); LOG("\n\n");
for (auto id : inp) { for (auto id : inp) {
LOG("%s", llama_token_to_piece(ctx, id).c_str()); LOG("%s", common_token_to_piece(ctx, id).c_str());
} }
fflush(stderr); fflush(stderr);
@ -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);
@ -156,12 +156,12 @@ 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 = llama_token_to_piece(ctx, id); const std::string token_str = common_token_to_piece(ctx, id);
LOG("%s", token_str.c_str()); LOG("%s", token_str.c_str());
fflush(stdout); fflush(stdout);
@ -172,7 +172,7 @@ int main(int argc, char ** argv) {
// debug // debug
if (dump_kv_cache) { if (dump_kv_cache) {
llama_kv_cache_view_update(ctx, &kvc_view); llama_kv_cache_view_update(ctx, &kvc_view);
llama_kv_cache_dump_view_seqs(kvc_view, 40); common_kv_cache_dump_view_seqs(kvc_view, 40);
} }
// build the mask from https://lmsys.org/blog/2023-11-21-lookahead-decoding/ // build the mask from https://lmsys.org/blog/2023-11-21-lookahead-decoding/
@ -201,10 +201,10 @@ int main(int argc, char ** argv) {
// V V V V V V // V V V V V V
// id // id
{ {
llama_batch_clear(batch); common_batch_clear(batch);
// current token - first token of the first level // current token - first token of the first level
llama_batch_add(batch, id, n_past, seq_id_all, true); common_batch_add(batch, id, n_past, seq_id_all, true);
// verification n-grams - queue this before the lookahead tokens for less KV cache fragmentation // verification n-grams - queue this before the lookahead tokens for less KV cache fragmentation
{ {
@ -229,7 +229,7 @@ int main(int argc, char ** argv) {
ngrams_cur[g].tokens [j + 1] = t; ngrams_cur[g].tokens [j + 1] = t;
ngrams_cur[g].i_batch[j + 1] = batch.n_tokens; ngrams_cur[g].i_batch[j + 1] = batch.n_tokens;
llama_batch_add(batch, t, n_past + j + 1, { W + 1 + g }, true); common_batch_add(batch, t, n_past + j + 1, { W + 1 + g }, true);
} }
} }
} }
@ -241,13 +241,13 @@ int main(int argc, char ** argv) {
seq_id_look[j] = i + j + 1; seq_id_look[j] = i + j + 1;
} }
llama_batch_add(batch, tokens_j[0][i], n_past + i, seq_id_look, false); common_batch_add(batch, tokens_j[0][i], n_past + i, seq_id_look, false);
} }
// fill the rest of the levels // fill the rest of the levels
for (int j = 1; j < N - 1; j++) { for (int j = 1; j < N - 1; j++) {
for (int i = 0; i < W; i++) { for (int i = 0; i < W; i++) {
llama_batch_add(batch, tokens_j[j][i], n_past + j + i, { i + 1 }, j == N - 2); common_batch_add(batch, tokens_j[j][i], n_past + j + i, { i + 1 }, j == N - 2);
} }
} }
} }
@ -281,13 +281,13 @@ 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
{ {
const std::string token_str = llama_token_to_piece(ctx, id); const std::string token_str = common_token_to_piece(ctx, id);
if (v == 0) { if (v == 0) {
LOG("%s", token_str.c_str()); LOG("%s", token_str.c_str());
@ -327,7 +327,7 @@ int main(int argc, char ** argv) {
// print known n-grams starting with token id (debug) // print known n-grams starting with token id (debug)
if (0 && v == 0) { if (0 && v == 0) {
if (ngrams_observed.cnt[id] > 0) { if (ngrams_observed.cnt[id] > 0) {
LOG("\n - %d n-grams starting with '%s'\n", ngrams_observed.cnt[id], llama_token_to_piece(ctx, id).c_str()); LOG("\n - %d n-grams starting with '%s'\n", ngrams_observed.cnt[id], common_token_to_piece(ctx, id).c_str());
} }
for (int i = 0; i < ngrams_observed.cnt[id]; i++) { for (int i = 0; i < ngrams_observed.cnt[id]; i++) {
@ -336,7 +336,7 @@ int main(int argc, char ** argv) {
const int idx = id*(N - 1)*G + i*(N - 1); const int idx = id*(N - 1)*G + i*(N - 1);
for (int j = 0; j < N - 1; j++) { for (int j = 0; j < N - 1; j++) {
const std::string token_str = llama_token_to_piece(ctx, ngrams_observed.tokens[idx + j]); const std::string token_str = common_token_to_piece(ctx, ngrams_observed.tokens[idx + j]);
LOG("%s", token_str.c_str()); LOG("%s", token_str.c_str());
} }
@ -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);

14
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
llama_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_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;
@ -31,15 +31,15 @@ int main(int argc, char ** argv){
// tokenize the prompt // tokenize the prompt
std::vector<llama_token> inp; std::vector<llama_token> inp;
inp = ::llama_tokenize(ctx, params.prompt, true, true); inp = common_tokenize(ctx, params.prompt, true, true);
fprintf(stderr, "%s: tokenization done\n", __func__); fprintf(stderr, "%s: tokenization done\n", __func__);
llama_ngram_cache ngram_cache; common_ngram_cache ngram_cache;
llama_ngram_cache_update(ngram_cache, LLAMA_NGRAM_STATIC, LLAMA_NGRAM_STATIC, inp, inp.size(), true); common_ngram_cache_update(ngram_cache, LLAMA_NGRAM_STATIC, LLAMA_NGRAM_STATIC, inp, inp.size(), true);
fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.lookup_cache_static.c_str()); fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.lookup_cache_static.c_str());
llama_ngram_cache_save(ngram_cache, params.lookup_cache_static); common_ngram_cache_save(ngram_cache, params.lookup_cache_static);
return 0; return 0;
} }

8
examples/lookup/lookup-merge.cpp
View File

@ -33,15 +33,15 @@ int main(int argc, char ** argv){
} }
fprintf(stderr, "lookup-merge: loading file %s\n", args[0].c_str()); fprintf(stderr, "lookup-merge: loading file %s\n", args[0].c_str());
llama_ngram_cache ngram_cache_merged = llama_ngram_cache_load(args[0]); common_ngram_cache ngram_cache_merged = common_ngram_cache_load(args[0]);
for (size_t i = 1; i < args.size()-1; ++i) { for (size_t i = 1; i < args.size()-1; ++i) {
fprintf(stderr, "lookup-merge: loading file %s\n", args[i].c_str()); fprintf(stderr, "lookup-merge: loading file %s\n", args[i].c_str());
llama_ngram_cache ngram_cache = llama_ngram_cache_load(args[i]); common_ngram_cache ngram_cache = common_ngram_cache_load(args[i]);
llama_ngram_cache_merge(ngram_cache_merged, ngram_cache); common_ngram_cache_merge(ngram_cache_merged, ngram_cache);
} }
fprintf(stderr, "lookup-merge: saving file %s\n", args.back().c_str()); fprintf(stderr, "lookup-merge: saving file %s\n", args.back().c_str());
llama_ngram_cache_save(ngram_cache_merged, args.back()); common_ngram_cache_save(ngram_cache_merged, args.back());
} }

28
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;
@ -28,18 +28,18 @@ int main(int argc, char ** argv){
llama_numa_init(params.numa); llama_numa_init(params.numa);
// load the model // load the model
llama_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_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;
// tokenize the prompt // tokenize the prompt
std::vector<llama_token> inp; std::vector<llama_token> inp;
inp = ::llama_tokenize(ctx, params.prompt, true, true); inp = common_tokenize(ctx, params.prompt, true, true);
llama_ngram_cache ngram_cache_context; common_ngram_cache ngram_cache_context;
llama_ngram_cache ngram_cache_dynamic; common_ngram_cache ngram_cache_dynamic;
llama_ngram_cache ngram_cache_static; common_ngram_cache ngram_cache_static;
int64_t t_draft_flat_us = 0; int64_t t_draft_flat_us = 0;
int64_t t_draft_us = 0; int64_t t_draft_us = 0;
@ -48,7 +48,7 @@ int main(int argc, char ** argv){
if (!params.lookup_cache_static.empty()) { if (!params.lookup_cache_static.empty()) {
try { try {
ngram_cache_static = llama_ngram_cache_load(params.lookup_cache_static); ngram_cache_static = common_ngram_cache_load(params.lookup_cache_static);
} catch (std::ifstream::failure const &) { } catch (std::ifstream::failure const &) {
LOG_ERR("failed to open static lookup cache: %s", params.lookup_cache_static.c_str()); LOG_ERR("failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
exit(1); exit(1);
@ -57,7 +57,7 @@ int main(int argc, char ** argv){
if (!params.lookup_cache_dynamic.empty()) { if (!params.lookup_cache_dynamic.empty()) {
try { try {
ngram_cache_dynamic = llama_ngram_cache_load(params.lookup_cache_dynamic); ngram_cache_dynamic = common_ngram_cache_load(params.lookup_cache_dynamic);
} catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
} }
@ -86,7 +86,7 @@ int main(int argc, char ** argv){
{ {
const int64_t t_start_draft_us = ggml_time_us(); const int64_t t_start_draft_us = ggml_time_us();
llama_ngram_cache_draft(pseudo_output, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static); common_ngram_cache_draft(pseudo_output, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static);
t_draft_us += ggml_time_us() - t_start_draft_us; t_draft_us += ggml_time_us() - t_start_draft_us;
} }
@ -105,7 +105,7 @@ int main(int argc, char ** argv){
{ {
const int64_t t_start_draft_us = ggml_time_us(); const int64_t t_start_draft_us = ggml_time_us();
llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, pseudo_output, 1, false); common_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, pseudo_output, 1, false);
t_draft_us += ggml_time_us() - t_start_draft_us; t_draft_us += ggml_time_us() - t_start_draft_us;
} }
} }
@ -115,7 +115,7 @@ int main(int argc, char ** argv){
pseudo_output.push_back(inp_slice[pseudo_output.size()]); pseudo_output.push_back(inp_slice[pseudo_output.size()]);
{ {
const int64_t t_start_draft_us = ggml_time_us(); const int64_t t_start_draft_us = ggml_time_us();
llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, pseudo_output, 1, false); common_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, pseudo_output, 1, false);
t_draft_us += ggml_time_us() - t_start_draft_us; t_draft_us += ggml_time_us() - t_start_draft_us;
} }
} }
@ -133,7 +133,7 @@ int main(int argc, char ** argv){
} }
// After each chunk, update the dynamic ngram cache with the context ngram cache: // After each chunk, update the dynamic ngram cache with the context ngram cache:
llama_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context); common_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
ngram_cache_context.clear(); ngram_cache_context.clear();
} }

54
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;
@ -31,29 +31,29 @@ int main(int argc, char ** argv){
llama_numa_init(params.numa); llama_numa_init(params.numa);
// load the model // load the model
llama_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_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;
// tokenize the prompt // tokenize the prompt
std::vector<llama_token> inp; std::vector<llama_token> inp;
inp = ::llama_tokenize(ctx, params.prompt, true, true); inp = common_tokenize(ctx, params.prompt, true, true);
llama_ngram_cache ngram_cache_context; common_ngram_cache ngram_cache_context;
llama_ngram_cache ngram_cache_dynamic; common_ngram_cache ngram_cache_dynamic;
llama_ngram_cache ngram_cache_static; common_ngram_cache ngram_cache_static;
int64_t t_draft_flat_us = 0; int64_t t_draft_flat_us = 0;
int64_t t_draft_us = 0; int64_t t_draft_us = 0;
{ {
// Fill up context ngram cache with tokens from user input: // Fill up context ngram cache with tokens from user input:
const int64_t t_start_draft_us = ggml_time_us(); const int64_t t_start_draft_us = ggml_time_us();
llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, inp.size(), false); common_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, inp.size(), false);
if (!params.lookup_cache_static.empty()) { if (!params.lookup_cache_static.empty()) {
try { try {
ngram_cache_static = llama_ngram_cache_load(params.lookup_cache_static); ngram_cache_static = common_ngram_cache_load(params.lookup_cache_static);
} catch (std::ifstream::failure const &) { } catch (std::ifstream::failure const &) {
LOG_ERR("failed to open static lookup cache: %s", params.lookup_cache_static.c_str()); LOG_ERR("failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
exit(1); exit(1);
@ -62,7 +62,7 @@ int main(int argc, char ** argv){
if (!params.lookup_cache_dynamic.empty()) { if (!params.lookup_cache_dynamic.empty()) {
try { try {
ngram_cache_dynamic = llama_ngram_cache_load(params.lookup_cache_dynamic); ngram_cache_dynamic = common_ngram_cache_load(params.lookup_cache_dynamic);
} catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
} }
@ -80,7 +80,7 @@ int main(int argc, char ** argv){
LOG("\n\n"); LOG("\n\n");
for (auto id : inp) { for (auto id : inp) {
LOG("%s", llama_token_to_piece(ctx, id).c_str()); LOG("%s", common_token_to_piece(ctx, id).c_str());
} }
fflush(stderr); fflush(stderr);
@ -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;
@ -117,7 +117,7 @@ int main(int argc, char ** argv){
// debug // debug
if (dump_kv_cache) { if (dump_kv_cache) {
llama_kv_cache_view_update(ctx, &kvc_view); llama_kv_cache_view_update(ctx, &kvc_view);
llama_kv_cache_dump_view_seqs(kvc_view, 40); common_kv_cache_dump_view_seqs(kvc_view, 40);
} }
// print current draft sequence // print current draft sequence
@ -126,11 +126,11 @@ 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 = llama_token_to_piece(ctx, id); const std::string token_str = common_token_to_piece(ctx, id);
if (!params.use_color) { if (!params.use_color) {
LOG("%s", token_str.c_str()); LOG("%s", token_str.c_str());
@ -152,7 +152,7 @@ int main(int argc, char ** argv){
{ {
// Update context ngram cache with the newly accepted token: // Update context ngram cache with the newly accepted token:
const int64_t t_start_draft_us = ggml_time_us(); const int64_t t_start_draft_us = ggml_time_us();
llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, 1, false); common_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, 1, false);
t_draft_us += ggml_time_us() - t_start_draft_us; t_draft_us += ggml_time_us() - t_start_draft_us;
} }
@ -178,7 +178,7 @@ int main(int argc, char ** argv){
{ {
// Update context ngram cache with the newly accepted token: // Update context ngram cache with the newly accepted token:
const int64_t t_start_draft_us = ggml_time_us(); const int64_t t_start_draft_us = ggml_time_us();
llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, 1, false); common_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, 1, false);
t_draft_us += ggml_time_us() - t_start_draft_us; t_draft_us += ggml_time_us() - t_start_draft_us;
} }
break; break;
@ -192,18 +192,18 @@ int main(int argc, char ** argv){
// clean the cache of draft tokens that weren't accepted // clean the cache of draft tokens that weren't accepted
llama_kv_cache_seq_rm(ctx, 0, n_past, -1); llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
llama_batch_clear(batch_tgt); common_batch_clear(batch_tgt);
llama_batch_add(batch_tgt, draft[0], n_past, { 0 }, true); common_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);
// Draft already contains a single token sampled from the model: // Draft already contains a single token sampled from the model:
GGML_ASSERT(draft.size() == 1); GGML_ASSERT(draft.size() == 1);
GGML_ASSERT(draft[0] == inp.back()); GGML_ASSERT(draft[0] == inp.back());
const int64_t t_start_draft_us = ggml_time_us(); const int64_t t_start_draft_us = ggml_time_us();
llama_ngram_cache_draft(inp, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static); common_ngram_cache_draft(inp, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static);
for (size_t i = 1; i < draft.size(); ++i) { for (size_t i = 1; i < draft.size(); ++i) {
llama_batch_add(batch_tgt, draft[i], n_past + i, { 0 }, true); common_batch_add(batch_tgt, draft[i], n_past + i, { 0 }, true);
} }
t_draft_us += ggml_time_us() - t_start_draft_us; t_draft_us += ggml_time_us() - t_start_draft_us;
@ -218,8 +218,8 @@ int main(int argc, char ** argv){
auto t_dec_end = ggml_time_us(); auto t_dec_end = ggml_time_us();
// Update dynamic ngram cache with context ngram cache and save it to disk: // Update dynamic ngram cache with context ngram cache and save it to disk:
llama_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context); common_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
llama_ngram_cache_save(ngram_cache_dynamic, params.lookup_cache_dynamic); common_ngram_cache_save(ngram_cache_dynamic, params.lookup_cache_dynamic);
LOG("\n\n"); LOG("\n\n");
@ -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);

92
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);
} }
@ -127,22 +127,22 @@ static void sigint_handler(int signo) {
} }
#endif #endif
static std::string chat_add_and_format(struct llama_model * model, std::vector<llama_chat_msg> & chat_msgs, const std::string & role, const std::string & content) { static std::string chat_add_and_format(struct llama_model * model, std::vector<common_chat_msg> & chat_msgs, const std::string & role, const std::string & content) {
llama_chat_msg new_msg{role, content}; common_chat_msg new_msg{role, content};
auto formatted = llama_chat_format_single(model, g_params->chat_template, chat_msgs, new_msg, role == "user"); auto formatted = common_chat_format_single(model, g_params->chat_template, chat_msgs, new_msg, role == "user");
chat_msgs.push_back({role, content}); chat_msgs.push_back({role, content});
LOG_DBG("formatted: '%s'\n", formatted.c_str()); LOG_DBG("formatted: '%s'\n", formatted.c_str());
return formatted; return formatted;
} }
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;
@ -187,9 +187,9 @@ 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<llama_chat_msg> chat_msgs; std::vector<common_chat_msg> chat_msgs;
g_model = &model; g_model = &model;
g_ctx = &ctx; g_ctx = &ctx;
@ -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__);
llama_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_params(params);
model = llama_init.model; model = llama_init.model;
ctx = llama_init.context; ctx = llama_init.context;
@ -246,7 +246,7 @@ int main(int argc, char ** argv) {
// print chat template example in conversation mode // print chat template example in conversation mode
if (params.conversation) { if (params.conversation) {
if (params.enable_chat_template) { if (params.enable_chat_template) {
LOG_INF("%s: chat template example:\n%s\n", __func__, llama_chat_format_example(model, params.chat_template).c_str()); LOG_INF("%s: chat template example:\n%s\n", __func__, common_chat_format_example(model, params.chat_template).c_str());
} else { } else {
LOG_INF("%s: in-suffix/prefix is specified, chat template will be disabled\n", __func__); LOG_INF("%s: in-suffix/prefix is specified, chat template will be disabled\n", __func__);
} }
@ -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");
} }
@ -296,7 +296,7 @@ int main(int argc, char ** argv) {
: params.prompt; : params.prompt;
if (params.interactive_first || !params.prompt.empty() || session_tokens.empty()) { if (params.interactive_first || !params.prompt.empty() || session_tokens.empty()) {
LOG_DBG("tokenize the prompt\n"); LOG_DBG("tokenize the prompt\n");
embd_inp = ::llama_tokenize(ctx, prompt, true, true); embd_inp = common_tokenize(ctx, prompt, true, true);
} else { } else {
LOG_DBG("use session tokens\n"); LOG_DBG("use session tokens\n");
embd_inp = session_tokens; embd_inp = session_tokens;
@ -379,13 +379,13 @@ int main(int argc, char ** argv) {
LOG_INF("%s: prompt: '%s'\n", __func__, params.prompt.c_str()); LOG_INF("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size()); LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
for (int i = 0; i < (int) embd_inp.size(); i++) { for (int i = 0; i < (int) embd_inp.size(); i++) {
LOG_INF("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str()); LOG_INF("%6d -> '%s'\n", embd_inp[i], common_token_to_piece(ctx, embd_inp[i]).c_str());
} }
if (params.n_keep > add_bos) { if (params.n_keep > add_bos) {
LOG_INF("%s: static prompt based on n_keep: '", __func__); LOG_INF("%s: static prompt based on n_keep: '", __func__);
for (int i = 0; i < params.n_keep; i++) { for (int i = 0; i < params.n_keep; i++) {
LOG_CNT("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str()); LOG_CNT("%s", common_token_to_piece(ctx, embd_inp[i]).c_str());
} }
LOG_CNT("'\n"); LOG_CNT("'\n");
} }
@ -415,9 +415,9 @@ int main(int argc, char ** argv) {
for (const auto & antiprompt : params.antiprompt) { for (const auto & antiprompt : params.antiprompt) {
LOG_INF("Reverse prompt: '%s'\n", antiprompt.c_str()); LOG_INF("Reverse prompt: '%s'\n", antiprompt.c_str());
if (params.verbose_prompt) { if (params.verbose_prompt) {
auto tmp = ::llama_tokenize(ctx, antiprompt, false, true); auto tmp = common_tokenize(ctx, antiprompt, false, true);
for (int i = 0; i < (int) tmp.size(); i++) { for (int i = 0; i < (int) tmp.size(); i++) {
LOG_INF("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str()); LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx, tmp[i]).c_str());
} }
} }
} }
@ -430,9 +430,9 @@ int main(int argc, char ** argv) {
if (!params.input_prefix.empty()) { if (!params.input_prefix.empty()) {
LOG_INF("Input prefix: '%s'\n", params.input_prefix.c_str()); LOG_INF("Input prefix: '%s'\n", params.input_prefix.c_str());
if (params.verbose_prompt) { if (params.verbose_prompt) {
auto tmp = ::llama_tokenize(ctx, params.input_prefix, true, true); auto tmp = common_tokenize(ctx, params.input_prefix, true, true);
for (int i = 0; i < (int) tmp.size(); i++) { for (int i = 0; i < (int) tmp.size(); i++) {
LOG_INF("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str()); LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx, tmp[i]).c_str());
} }
} }
} }
@ -440,23 +440,23 @@ int main(int argc, char ** argv) {
if (!params.input_suffix.empty()) { if (!params.input_suffix.empty()) {
LOG_INF("Input suffix: '%s'\n", params.input_suffix.c_str()); LOG_INF("Input suffix: '%s'\n", params.input_suffix.c_str());
if (params.verbose_prompt) { if (params.verbose_prompt) {
auto tmp = ::llama_tokenize(ctx, params.input_suffix, false, true); auto tmp = common_tokenize(ctx, params.input_suffix, false, true);
for (int i = 0; i < (int) tmp.size(); i++) { for (int i = 0; i < (int) tmp.size(); i++) {
LOG_INF("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str()); LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx, tmp[i]).c_str());
} }
} }
} }
} }
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);
@ -521,7 +521,7 @@ int main(int argc, char ** argv) {
antiprompt_ids.reserve(params.antiprompt.size()); antiprompt_ids.reserve(params.antiprompt.size());
for (const std::string & antiprompt : params.antiprompt) { for (const std::string & antiprompt : params.antiprompt) {
antiprompt_ids.emplace_back(::llama_tokenize(ctx, antiprompt, false, true)); antiprompt_ids.emplace_back(::common_tokenize(ctx, antiprompt, false, true));
} }
if (llama_model_has_encoder(model)) { if (llama_model_has_encoder(model)) {
@ -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) {
@ -714,7 +714,7 @@ int main(int argc, char ** argv) {
// display text // display text
if (input_echo && display) { if (input_echo && display) {
for (auto id : embd) { for (auto id : embd) {
const std::string token_str = llama_token_to_piece(ctx, id, params.special); const std::string token_str = common_token_to_piece(ctx, id, params.special);
// Console/Stream Output // Console/Stream Output
LOG("%s", token_str.c_str()); LOG("%s", token_str.c_str());
@ -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) {
@ -782,13 +782,13 @@ 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) {
if (!params.antiprompt.empty()) { if (!params.antiprompt.empty()) {
// tokenize and inject first reverse prompt // tokenize and inject first reverse prompt
const auto first_antiprompt = ::llama_tokenize(ctx, params.antiprompt.front(), false, true); const auto first_antiprompt = common_tokenize(ctx, params.antiprompt.front(), false, true);
embd_inp.insert(embd_inp.end(), first_antiprompt.begin(), first_antiprompt.end()); embd_inp.insert(embd_inp.end(), first_antiprompt.begin(), first_antiprompt.end());
is_antiprompt = true; is_antiprompt = true;
} }
@ -803,8 +803,8 @@ 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 << llama_token_to_piece(ctx, id, false); assistant_ss << common_token_to_piece(ctx, id, false);
} }
if (n_past > 0 && is_interacting) { if (n_past > 0 && is_interacting) {
@ -862,9 +862,9 @@ int main(int argc, char ** argv) {
? chat_add_and_format(model, chat_msgs, "user", std::move(buffer)) ? chat_add_and_format(model, chat_msgs, "user", std::move(buffer))
: std::move(buffer); : std::move(buffer);
// TODO: one inconvenient of current chat template implementation is that we can't distinguish between user input and special tokens (prefix/postfix) // TODO: one inconvenient of current chat template implementation is that we can't distinguish between user input and special tokens (prefix/postfix)
const auto line_pfx = ::llama_tokenize(ctx, params.input_prefix, false, true); const auto line_pfx = common_tokenize(ctx, params.input_prefix, false, true);
const auto line_inp = ::llama_tokenize(ctx, user_inp, false, format_chat); const auto line_inp = common_tokenize(ctx, user_inp, false, format_chat);
const auto line_sfx = ::llama_tokenize(ctx, params.input_suffix, false, true); const auto line_sfx = common_tokenize(ctx, params.input_suffix, false, true);
LOG_DBG("input tokens: %s\n", string_from(ctx, line_inp).c_str()); LOG_DBG("input tokens: %s\n", string_from(ctx, line_inp).c_str());
@ -882,7 +882,7 @@ int main(int argc, char ** argv) {
for (size_t i = original_size; i < embd_inp.size(); ++i) { for (size_t i = original_size; i < embd_inp.size(); ++i) {
const llama_token token = embd_inp[i]; const llama_token token = embd_inp[i];
output_tokens.push_back(token); output_tokens.push_back(token);
output_ss << llama_token_to_piece(ctx, token); output_ss << common_token_to_piece(ctx, token);
} }
// reset assistant message // reset assistant message
@ -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);

36
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
llama_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_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;
@ -160,11 +160,11 @@ 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;
tokens_system = ::llama_tokenize(ctx, k_system, true); tokens_system = common_tokenize(ctx, k_system, true);
const int32_t n_tokens_system = tokens_system.size(); const int32_t n_tokens_system = tokens_system.size();
llama_seq_id g_seq_id = 0; llama_seq_id g_seq_id = 0;
@ -189,7 +189,7 @@ int main(int argc, char ** argv) {
LOG_INF("%s: Evaluating the system prompt ...\n", __func__); LOG_INF("%s: Evaluating the system prompt ...\n", __func__);
for (int32_t i = 0; i < n_tokens_system; ++i) { for (int32_t i = 0; i < n_tokens_system; ++i) {
llama_batch_add(batch, tokens_system[i], i, { 0 }, false); common_batch_add(batch, tokens_system[i], i, { 0 }, false);
} }
if (llama_decode(ctx, batch) != 0) { if (llama_decode(ctx, batch) != 0) {
@ -210,10 +210,10 @@ int main(int argc, char ** argv) {
while (true) { while (true) {
if (dump_kv_cache) { if (dump_kv_cache) {
llama_kv_cache_view_update(ctx, &kvc_view); llama_kv_cache_view_update(ctx, &kvc_view);
llama_kv_cache_dump_view_seqs(kvc_view, 40); common_kv_cache_dump_view_seqs(kvc_view, 40);
} }
llama_batch_clear(batch); common_batch_clear(batch);
// decode any currently ongoing sequences // decode any currently ongoing sequences
for (auto & client : clients) { for (auto & client : clients) {
@ -223,7 +223,7 @@ int main(int argc, char ** argv) {
client.i_batch = batch.n_tokens; client.i_batch = batch.n_tokens;
llama_batch_add(batch, client.sampled, n_tokens_system + client.n_prompt + client.n_decoded, { client.id + 1 }, true); common_batch_add(batch, client.sampled, n_tokens_system + client.n_prompt + client.n_decoded, { client.id + 1 }, true);
client.n_decoded += 1; client.n_decoded += 1;
} }
@ -252,14 +252,14 @@ 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;
tokens_prompt = ::llama_tokenize(ctx, client.prompt, false); tokens_prompt = common_tokenize(ctx, client.prompt, false);
for (size_t i = 0; i < tokens_prompt.size(); ++i) { for (size_t i = 0; i < tokens_prompt.size(); ++i) {
llama_batch_add(batch, tokens_prompt[i], i + n_tokens_system, { client.id + 1 }, false); common_batch_add(batch, tokens_prompt[i], i + n_tokens_system, { client.id + 1 }, false);
} }
// extract the logits only for the last token // extract the logits only for the last token
@ -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
@ -350,7 +350,7 @@ int main(int argc, char ** argv) {
client.t_start_gen = ggml_time_us(); client.t_start_gen = ggml_time_us();
} }
const std::string token_str = llama_token_to_piece(ctx, id); const std::string token_str = common_token_to_piece(ctx, id);
client.response += token_str; client.response += token_str;
client.sampled = id; client.sampled = id;

28
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 = llama_model_params_from_gpt_params(params); llama_model_params model_params = common_model_params_to_llama(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 = llama_context_params_from_gpt_params(params); llama_context_params ctx_params = common_context_params_to_llama(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;
@ -92,10 +92,10 @@ int main(int argc, char ** argv) {
// tokenize the prompt // tokenize the prompt
std::vector<llama_token> tokens_list; std::vector<llama_token> tokens_list;
tokens_list = ::llama_tokenize(ctx, params.prompt, true); tokens_list = common_tokenize(ctx, params.prompt, true);
// tokenize the prefix and use it as a sink // tokenize the prefix and use it as a sink
const int n_tokens_prefix = ::llama_tokenize(ctx, prompt_prefix, true).size(); const int n_tokens_prefix = common_tokenize(ctx, prompt_prefix, true).size();
const int n_tokens_all = tokens_list.size(); const int n_tokens_all = tokens_list.size();
@ -137,10 +137,10 @@ int main(int argc, char ** argv) {
n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1; n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
} }
llama_batch_clear(batch); common_batch_clear(batch);
for (int j = 0; j < n_batch && i + j < n_tokens_all; j++) { for (int j = 0; j < n_batch && i + j < n_tokens_all; j++) {
llama_batch_add(batch, tokens_list[i + j], n_past++, { 0 }, false); common_batch_add(batch, tokens_list[i + j], n_past++, { 0 }, false);
} }
if (i + n_batch >= n_tokens_all) { if (i + n_batch >= n_tokens_all) {
@ -171,10 +171,10 @@ int main(int argc, char ** argv) {
n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1; n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
llama_batch_clear(batch); common_batch_clear(batch);
for (int j = 0; j < n_batch && i + j < n_tokens_all; j++) { for (int j = 0; j < n_batch && i + j < n_tokens_all; j++) {
llama_batch_add(batch, tokens_list[i + j], n_past++, { 0 }, false); common_batch_add(batch, tokens_list[i + j], n_past++, { 0 }, false);
} }
if (i + n_batch >= n_tokens_all) { if (i + n_batch >= n_tokens_all) {
@ -229,15 +229,15 @@ int main(int argc, char ** argv) {
break; break;
} }
LOG("%s", llama_token_to_piece(ctx, new_token_id).c_str()); LOG("%s", common_token_to_piece(ctx, new_token_id).c_str());
n_decode += 1; n_decode += 1;
// prepare the next batch // prepare the next batch
llama_batch_clear(batch); common_batch_clear(batch);
// push this new token for next evaluation // push this new token for next evaluation
llama_batch_add(batch, new_token_id, n_past++, { 0 }, true); common_batch_add(batch, new_token_id, n_past++, { 0 }, true);
} }
n_cur += 1; n_cur += 1;

58
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()) {
@ -339,7 +339,7 @@ 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]`
@ -350,7 +350,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
LOG_INF("%s: tokenizing the input ..\n", __func__); LOG_INF("%s: tokenizing the input ..\n", __func__);
std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, true); std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, true);
const int n_ctx = llama_n_ctx(ctx); const int n_ctx = llama_n_ctx(ctx);
@ -474,7 +474,7 @@ 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);
} }
@ -502,7 +502,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
auto tim1 = std::chrono::high_resolution_clock::now(); auto tim1 = std::chrono::high_resolution_clock::now();
LOG_INF("%s: tokenizing the input ..\n", __func__); LOG_INF("%s: tokenizing the input ..\n", __func__);
std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, true); std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, true);
auto tim2 = std::chrono::high_resolution_clock::now(); auto tim2 = std::chrono::high_resolution_clock::now();
LOG_INF("%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count()); LOG_INF("%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
@ -772,7 +772,7 @@ 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
@ -853,7 +853,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
hs_cur.gold_ending_idx = std::stoi( prompt_lines[idx*6+1] ); hs_cur.gold_ending_idx = std::stoi( prompt_lines[idx*6+1] );
for (size_t j = 0; j < 4; j++) { for (size_t j = 0; j < 4; j++) {
hs_cur.ending[j] = prompt_lines[idx*6+2+j]; hs_cur.ending[j] = prompt_lines[idx*6+2+j];
hs_cur.seq_tokens[j] = ::llama_tokenize(ctx, hs_cur.context + " " + hs_cur.ending[j], true); hs_cur.seq_tokens[j] = common_tokenize(ctx, hs_cur.context + " " + hs_cur.ending[j], true);
} }
// determine the common prefix of the endings // determine the common prefix of the endings
@ -910,7 +910,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
size_t i1 = i0; size_t i1 = i0;
size_t i_logits = 0; // this tells us how many logits were needed before this point in the batch size_t i_logits = 0; // this tells us how many logits were needed before this point in the batch
llama_batch_clear(batch); common_batch_clear(batch);
// batch as much tasks as possible into the available context // batch as much tasks as possible into the available context
// each task has 4 unique sequence ids - one for each ending // each task has 4 unique sequence ids - one for each ending
@ -926,7 +926,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
} }
for (size_t i = 0; i < hs_cur.common_prefix; ++i) { for (size_t i = 0; i < hs_cur.common_prefix; ++i) {
llama_batch_add(batch, hs_cur.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3 }, false); common_batch_add(batch, hs_cur.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3 }, false);
} }
batch.logits[batch.n_tokens - 1] = true; // we need logits for the last token of the common prefix batch.logits[batch.n_tokens - 1] = true; // we need logits for the last token of the common prefix
n_logits += 1; n_logits += 1;
@ -936,7 +936,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
// TODO: don't evaluate the last token of each sequence // TODO: don't evaluate the last token of each sequence
for (size_t i = hs_cur.common_prefix; i < seq_tokens_size; ++i) { for (size_t i = hs_cur.common_prefix; i < seq_tokens_size; ++i) {
const bool needs_logits = i < seq_tokens_size - 1; const bool needs_logits = i < seq_tokens_size - 1;
llama_batch_add(batch, hs_cur.seq_tokens[s][i], i, { s0 + s }, needs_logits); common_batch_add(batch, hs_cur.seq_tokens[s][i], i, { s0 + s }, needs_logits);
n_logits += needs_logits; n_logits += needs_logits;
} }
} }
@ -1112,7 +1112,7 @@ 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;
@ -1146,8 +1146,8 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
LOG_INF("%s : tokenizing selected tasks\n", __func__); LOG_INF("%s : tokenizing selected tasks\n", __func__);
for (auto & task : data) { for (auto & task : data) {
task.seq_tokens[0] = ::llama_tokenize(ctx, task.first + task.choices[0] + task.second, true); task.seq_tokens[0] = common_tokenize(ctx, task.first + task.choices[0] + task.second, true);
task.seq_tokens[1] = ::llama_tokenize(ctx, task.first + task.choices[1] + task.second, true); task.seq_tokens[1] = common_tokenize(ctx, task.first + task.choices[1] + task.second, true);
task.common_prefix = 0; task.common_prefix = 0;
for (size_t k = 0; k < task.seq_tokens[0].size(); k++) { for (size_t k = 0; k < task.seq_tokens[0].size(); k++) {
@ -1162,8 +1162,8 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
task.seq_tokens[0].size() - task.common_prefix + task.seq_tokens[0].size() - task.common_prefix +
task.seq_tokens[1].size() - task.common_prefix; task.seq_tokens[1].size() - task.common_prefix;
task.n_base1 = ::llama_tokenize(ctx, task.first + task.choices[0], true).size(); task.n_base1 = common_tokenize(ctx, task.first + task.choices[0], true).size();
task.n_base2 = ::llama_tokenize(ctx, task.first + task.choices[1], true).size(); task.n_base2 = common_tokenize(ctx, task.first + task.choices[1], true).size();
} }
LOG_INF("%s : calculating winogrande score over selected tasks.\n", __func__); LOG_INF("%s : calculating winogrande score over selected tasks.\n", __func__);
@ -1195,7 +1195,7 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
size_t i1 = i0; size_t i1 = i0;
size_t i_logits = 0; size_t i_logits = 0;
llama_batch_clear(batch); common_batch_clear(batch);
while (n_cur + (int) data[i1].required_tokens <= n_ctx) { while (n_cur + (int) data[i1].required_tokens <= n_ctx) {
int n_logits = 0; int n_logits = 0;
@ -1205,7 +1205,7 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
} }
for (size_t i = 0; i < data[i1].common_prefix; ++i) { for (size_t i = 0; i < data[i1].common_prefix; ++i) {
llama_batch_add(batch, data[i1].seq_tokens[0][i], i, { s0 + 0, s0 + 1 }, false); common_batch_add(batch, data[i1].seq_tokens[0][i], i, { s0 + 0, s0 + 1 }, false);
} }
batch.logits[batch.n_tokens - 1] = true; batch.logits[batch.n_tokens - 1] = true;
n_logits += 1; n_logits += 1;
@ -1213,7 +1213,7 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
for (int s = 0; s < 2; ++s) { for (int s = 0; s < 2; ++s) {
// TODO: end before the last token, no need to predict past the end of the sequences // TODO: end before the last token, no need to predict past the end of the sequences
for (size_t i = data[i1].common_prefix; i < data[i1].seq_tokens[s].size(); ++i) { for (size_t i = data[i1].common_prefix; i < data[i1].seq_tokens[s].size(); ++i) {
llama_batch_add(batch, data[i1].seq_tokens[s][i], i, { s0 + s }, true); common_batch_add(batch, data[i1].seq_tokens[s][i], i, { s0 + s }, true);
n_logits += 1; n_logits += 1;
} }
} }
@ -1370,7 +1370,7 @@ static bool multiple_choice_prepare_one_task(llama_context * ctx, multiple_choic
} }
return false; return false;
} }
task.seq_tokens.emplace_back(::llama_tokenize(ctx, task.question + " " + answer, true)); task.seq_tokens.emplace_back(::common_tokenize(ctx, task.question + " " + answer, true));
} }
auto min_len = task.seq_tokens.front().size(); auto min_len = task.seq_tokens.front().size();
for (auto& seq : task.seq_tokens) { for (auto& seq : task.seq_tokens) {
@ -1414,7 +1414,7 @@ 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;
@ -1548,7 +1548,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
size_t i1 = i0; size_t i1 = i0;
size_t i_logits = 0; // this tells us how many logits were needed before this point in the batch size_t i_logits = 0; // this tells us how many logits were needed before this point in the batch
llama_batch_clear(batch); common_batch_clear(batch);
// batch as much tasks as possible into the available context // batch as much tasks as possible into the available context
// each task has 4 unique sequence ids - one for each ending // each task has 4 unique sequence ids - one for each ending
@ -1571,7 +1571,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
for (size_t i = 0; i < cur_task.common_prefix; ++i) { for (size_t i = 0; i < cur_task.common_prefix; ++i) {
//llama_batch_add(batch, cur_task.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3}, false); //llama_batch_add(batch, cur_task.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3}, false);
llama_batch_add(batch, cur_task.seq_tokens[0][i], i, batch_indeces, false); common_batch_add(batch, cur_task.seq_tokens[0][i], i, batch_indeces, false);
} }
batch.logits[batch.n_tokens - 1] = true; // we need logits for the last token of the common prefix batch.logits[batch.n_tokens - 1] = true; // we need logits for the last token of the common prefix
n_logits += 1; n_logits += 1;
@ -1581,7 +1581,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
// TODO: don't evaluate the last token of each sequence // TODO: don't evaluate the last token of each sequence
for (size_t i = cur_task.common_prefix; i < seq_tokens_size; ++i) { for (size_t i = cur_task.common_prefix; i < seq_tokens_size; ++i) {
const bool needs_logits = i < seq_tokens_size - 1; const bool needs_logits = i < seq_tokens_size - 1;
llama_batch_add(batch, cur_task.seq_tokens[s][i], i, { s0 + s }, needs_logits); common_batch_add(batch, cur_task.seq_tokens[s][i], i, { s0 + s }, needs_logits);
n_logits += needs_logits; n_logits += needs_logits;
} }
} }
@ -1695,7 +1695,7 @@ 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;
@ -1968,17 +1968,17 @@ 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;
@ -2017,7 +2017,7 @@ 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
llama_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_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;
@ -2036,7 +2036,7 @@ 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;

26
examples/retrieval/retrieval.cpp
View File

@ -77,7 +77,7 @@ static std::vector<chunk> chunk_file(const std::string & filename, int chunk_siz
static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, llama_seq_id seq_id) { static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, llama_seq_id seq_id) {
size_t n_tokens = tokens.size(); size_t n_tokens = tokens.size();
for (size_t i = 0; i < n_tokens; i++) { for (size_t i = 0; i < n_tokens; i++) {
llama_batch_add(batch, tokens[i], i, { seq_id }, true); common_batch_add(batch, tokens[i], i, { seq_id }, true);
} }
} }
@ -107,18 +107,18 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
} }
float * out = output + batch.seq_id[i][0] * n_embd; float * out = output + batch.seq_id[i][0] * n_embd;
llama_embd_normalize(embd, out, n_embd); common_embd_normalize(embd, out, n_embd);
} }
} }
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
llama_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_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
@ -185,7 +185,7 @@ int main(int argc, char ** argv) {
// tokenize the prompts and trim // tokenize the prompts and trim
for (auto & chunk : chunks) { for (auto & chunk : chunks) {
auto inp = ::llama_tokenize(ctx, chunk.textdata, true, false); auto inp = common_tokenize(ctx, chunk.textdata, true, false);
if (inp.size() > n_batch) { if (inp.size() > n_batch) {
LOG_ERR("%s: chunk size (%lld) exceeds batch size (%lld), increase batch size and re-run\n", LOG_ERR("%s: chunk size (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
__func__, (long long int) inp.size(), (long long int) n_batch); __func__, (long long int) inp.size(), (long long int) n_batch);
@ -204,7 +204,7 @@ int main(int argc, char ** argv) {
LOG_INF("%s: prompt %d: '%s'\n", __func__, i, chunks[i].textdata.c_str()); LOG_INF("%s: prompt %d: '%s'\n", __func__, i, chunks[i].textdata.c_str());
LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, chunks[i].tokens.size()); LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, chunks[i].tokens.size());
for (int j = 0; j < (int) chunks[i].tokens.size(); j++) { for (int j = 0; j < (int) chunks[i].tokens.size(); j++) {
LOG_INF("%6d -> '%s'\n", chunks[i].tokens[j], llama_token_to_piece(ctx, chunks[i].tokens[j]).c_str()); LOG_INF("%6d -> '%s'\n", chunks[i].tokens[j], common_token_to_piece(ctx, chunks[i].tokens[j]).c_str());
} }
LOG_INF("\n\n"); LOG_INF("\n\n");
} }
@ -232,7 +232,7 @@ int main(int argc, char ** argv) {
if (batch.n_tokens + n_toks > n_batch) { if (batch.n_tokens + n_toks > n_batch) {
float * out = emb + p * n_embd; float * out = emb + p * n_embd;
batch_decode(ctx, batch, out, s, n_embd); batch_decode(ctx, batch, out, s, n_embd);
llama_batch_clear(batch); common_batch_clear(batch);
p += s; p += s;
s = 0; s = 0;
} }
@ -260,20 +260,20 @@ int main(int argc, char ** argv) {
while (true) { while (true) {
LOG("Enter query: "); LOG("Enter query: ");
std::getline(std::cin, query); std::getline(std::cin, query);
std::vector<int32_t> query_tokens = llama_tokenize(ctx, query, true); std::vector<int32_t> query_tokens = common_tokenize(ctx, query, true);
batch_add_seq(query_batch, query_tokens, 0); batch_add_seq(query_batch, query_tokens, 0);
std::vector<float> query_emb(n_embd, 0); std::vector<float> query_emb(n_embd, 0);
batch_decode(ctx, query_batch, query_emb.data(), 1, n_embd); batch_decode(ctx, query_batch, query_emb.data(), 1, n_embd);
llama_batch_clear(query_batch); common_batch_clear(query_batch);
// compute cosine similarities // compute cosine similarities
{ {
std::vector<std::pair<int, float>> similarities; std::vector<std::pair<int, float>> similarities;
for (int i = 0; i < n_chunks; i++) { for (int i = 0; i < n_chunks; i++) {
float sim = llama_embd_similarity_cos(chunks[i].embedding.data(), query_emb.data(), n_embd); float sim = common_embd_similarity_cos(chunks[i].embedding.data(), query_emb.data(), n_embd);
similarities.push_back(std::make_pair(i, sim)); similarities.push_back(std::make_pair(i, sim));
} }

18
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
llama_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_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;
@ -46,7 +46,7 @@ int main(int argc, char ** argv) {
llama_sampler_chain_add(smpl, llama_sampler_init_dist(params.sparams.seed)); llama_sampler_chain_add(smpl, llama_sampler_init_dist(params.sparams.seed));
// tokenize prompt // tokenize prompt
auto tokens = llama_tokenize(ctx, params.prompt, true); auto tokens = common_tokenize(ctx, params.prompt, true);
// evaluate prompt // evaluate prompt
llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), n_past, 0)); llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), n_past, 0));
@ -72,7 +72,7 @@ int main(int argc, char ** argv) {
for (auto i = 0; i < params.n_predict; i++) { for (auto i = 0; i < params.n_predict; i++) {
auto next_token = llama_sampler_sample(smpl, ctx, -1); auto next_token = llama_sampler_sample(smpl, ctx, -1);
auto next_token_str = llama_token_to_piece(ctx, next_token); auto next_token_str = common_token_to_piece(ctx, next_token);
printf("%s", next_token_str.c_str()); printf("%s", next_token_str.c_str());
result0 += next_token_str; result0 += next_token_str;
@ -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, llama_context_params_from_gpt_params(params)); auto * ctx2 = llama_new_context_with_model(model, common_context_params_to_llama(params));
llama_sampler * smpl2 = llama_sampler_chain_init(sparams); llama_sampler * smpl2 = llama_sampler_chain_init(sparams);
@ -128,7 +128,7 @@ int main(int argc, char ** argv) {
// second run // second run
for (auto i = 0; i < params.n_predict; i++) { for (auto i = 0; i < params.n_predict; i++) {
auto next_token = llama_sampler_sample(smpl2, ctx2, -1); auto next_token = llama_sampler_sample(smpl2, ctx2, -1);
auto next_token_str = llama_token_to_piece(ctx2, next_token); auto next_token_str = common_token_to_piece(ctx2, next_token);
printf("%s", next_token_str.c_str()); printf("%s", next_token_str.c_str());
result1 += next_token_str; result1 += next_token_str;
@ -152,7 +152,7 @@ int main(int argc, char ** argv) {
} }
// make new context // make new context
auto * ctx3 = llama_new_context_with_model(model, llama_context_params_from_gpt_params(params)); auto * ctx3 = llama_new_context_with_model(model, common_context_params_to_llama(params));
llama_sampler * smpl3 = llama_sampler_chain_init(sparams); llama_sampler * smpl3 = llama_sampler_chain_init(sparams);
@ -216,7 +216,7 @@ int main(int argc, char ** argv) {
// third run with seq 1 instead of 0 // third run with seq 1 instead of 0
for (auto i = 0; i < params.n_predict; i++) { for (auto i = 0; i < params.n_predict; i++) {
auto next_token = llama_sampler_sample(smpl3, ctx3, -1); auto next_token = llama_sampler_sample(smpl3, ctx3, -1);
auto next_token_str = llama_token_to_piece(ctx3, next_token); auto next_token_str = common_token_to_piece(ctx3, next_token);
printf("%s", next_token_str.c_str()); printf("%s", next_token_str.c_str());
result2 += next_token_str; result2 += next_token_str;

80
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
} }
@ -611,9 +611,9 @@ struct server_response {
struct server_context { struct server_context {
llama_model * model = nullptr; llama_model * model = nullptr;
llama_context * ctx = nullptr; llama_context * ctx = nullptr;
std::vector<llama_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;
llama_init_result llama_init = llama_init_from_gpt_params(params); common_init_result llama_init = common_init_from_params(params);
model = llama_init.model; model = llama_init.model;
ctx = llama_init.context; ctx = llama_init.context;
@ -771,10 +771,10 @@ struct server_context {
std::vector<llama_token> p; std::vector<llama_token> p;
if (first) { if (first) {
p = ::llama_tokenize(ctx, s, add_special, TMP_FORCE_SPECIAL); p = common_tokenize(ctx, s, add_special, TMP_FORCE_SPECIAL);
first = false; first = false;
} else { } else {
p = ::llama_tokenize(ctx, s, false, TMP_FORCE_SPECIAL); p = common_tokenize(ctx, s, false, TMP_FORCE_SPECIAL);
} }
prompt_tokens.insert(prompt_tokens.end(), p.begin(), p.end()); prompt_tokens.insert(prompt_tokens.end(), p.begin(), p.end());
@ -788,7 +788,7 @@ struct server_context {
} }
} else { } else {
auto s = json_prompt.template get<std::string>(); auto s = json_prompt.template get<std::string>();
prompt_tokens = ::llama_tokenize(ctx, s, add_special, TMP_FORCE_SPECIAL); prompt_tokens = common_tokenize(ctx, s, add_special, TMP_FORCE_SPECIAL);
} }
return prompt_tokens; return prompt_tokens;
@ -999,7 +999,7 @@ struct server_context {
slot.sparams.logit_bias.push_back({tok, bias}); slot.sparams.logit_bias.push_back({tok, bias});
} }
} else if (el[0].is_string()) { } else if (el[0].is_string()) {
auto toks = llama_tokenize(model, el[0].get<std::string>(), false); auto toks = common_tokenize(model, el[0].get<std::string>(), false);
for (auto tok : toks) { for (auto tok : toks) {
slot.sparams.logit_bias.push_back({tok, bias}); slot.sparams.logit_bias.push_back({tok, bias});
} }
@ -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);
@ -1073,7 +1073,7 @@ struct server_context {
system_tokens.clear(); system_tokens.clear();
if (!system_prompt.empty()) { if (!system_prompt.empty()) {
system_tokens = ::llama_tokenize(ctx, system_prompt, true); system_tokens = common_tokenize(ctx, system_prompt, true);
const int32_t n_batch = llama_n_batch(ctx); const int32_t n_batch = llama_n_batch(ctx);
const int32_t n_tokens_prompt = system_tokens.size(); const int32_t n_tokens_prompt = system_tokens.size();
@ -1081,10 +1081,10 @@ struct server_context {
for (int32_t i = 0; i < n_tokens_prompt; i += n_batch) { for (int32_t i = 0; i < n_tokens_prompt; i += n_batch) {
const int32_t n_tokens = std::min(n_batch, n_tokens_prompt - i); const int32_t n_tokens = std::min(n_batch, n_tokens_prompt - i);
llama_batch_clear(batch); common_batch_clear(batch);
for (int32_t j = 0; j < n_tokens; ++j) { for (int32_t j = 0; j < n_tokens; ++j) {
llama_batch_add(batch, system_tokens[i + j], i + j, { 0 }, false); common_batch_add(batch, system_tokens[i + j], i + j, { 0 }, false);
} }
if (llama_decode(ctx, batch) != 0) { if (llama_decode(ctx, batch) != 0) {
@ -1113,7 +1113,7 @@ struct server_context {
bool process_token(completion_token_output & result, server_slot & slot) { bool process_token(completion_token_output & result, server_slot & slot) {
// remember which tokens were sampled - used for repetition penalties during sampling // remember which tokens were sampled - used for repetition penalties during sampling
const std::string token_str = llama_token_to_piece(ctx, result.tok, params.special); const std::string token_str = common_token_to_piece(ctx, result.tok, params.special);
slot.sampled = result.tok; slot.sampled = result.tok;
// search stop word and delete it // search stop word and delete it
@ -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},
@ -1297,7 +1297,7 @@ struct server_context {
}; };
if (slot.sparams.n_probs > 0) { if (slot.sparams.n_probs > 0) {
const std::vector<llama_token> to_send_toks = llama_tokenize(ctx, tkn.text_to_send, false); const std::vector<llama_token> to_send_toks = common_tokenize(ctx, tkn.text_to_send, false);
const size_t probs_pos = std::min(slot.n_sent_token_probs, slot.generated_token_probs.size()); const size_t probs_pos = std::min(slot.n_sent_token_probs, slot.generated_token_probs.size());
const size_t probs_stop_pos = std::min(slot.n_sent_token_probs + to_send_toks.size(), slot.generated_token_probs.size()); const size_t probs_stop_pos = std::min(slot.n_sent_token_probs + to_send_toks.size(), slot.generated_token_probs.size());
@ -1347,7 +1347,7 @@ struct server_context {
if (slot.sparams.n_probs > 0) { if (slot.sparams.n_probs > 0) {
std::vector<completion_token_output> probs; std::vector<completion_token_output> probs;
if (!slot.params.stream && slot.stopped_word) { if (!slot.params.stream && slot.stopped_word) {
const std::vector<llama_token> stop_word_toks = llama_tokenize(ctx, slot.stopping_word, false); const std::vector<llama_token> stop_word_toks = common_tokenize(ctx, slot.stopping_word, false);
size_t safe_offset = std::min(slot.generated_token_probs.size(), stop_word_toks.size()); size_t safe_offset = std::min(slot.generated_token_probs.size(), stop_word_toks.size());
probs = std::vector<completion_token_output>( probs = std::vector<completion_token_output>(
@ -1401,7 +1401,7 @@ struct server_context {
continue; continue;
} }
llama_embd_normalize(embd, embd_res.data(), n_embd); common_embd_normalize(embd, embd_res.data(), n_embd);
res.data = json { res.data = json {
{"embedding", embd_res}, {"embedding", embd_res},
@ -1835,7 +1835,7 @@ struct server_context {
} break; } break;
case SERVER_TASK_TYPE_SET_LORA: case SERVER_TASK_TYPE_SET_LORA:
{ {
llama_lora_adapters_apply(ctx, loras); common_lora_adapters_apply(ctx, loras);
server_task_result result; server_task_result result;
result.id = task.id; result.id = task.id;
result.stop = true; result.stop = true;
@ -1921,7 +1921,7 @@ struct server_context {
} }
// start populating the batch for this iteration // start populating the batch for this iteration
llama_batch_clear(batch); common_batch_clear(batch);
// frist, add sampled tokens from any ongoing sequences // frist, add sampled tokens from any ongoing sequences
for (auto & slot : slots) { for (auto & slot : slots) {
@ -1935,7 +1935,7 @@ struct server_context {
// TODO: we always have to take into account the "system_tokens" // TODO: we always have to take into account the "system_tokens"
// this is not great and needs to be improved somehow // this is not great and needs to be improved somehow
llama_batch_add(batch, slot.sampled, system_tokens.size() + slot_npast, { slot.id + 1 }, true); common_batch_add(batch, slot.sampled, system_tokens.size() + slot_npast, { slot.id + 1 }, true);
slot.n_past += 1; slot.n_past += 1;
@ -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
@ -2184,7 +2184,7 @@ struct server_context {
} }
} }
llama_batch_add(batch, prompt_tokens[slot.n_past], system_tokens.size() + slot_npast, { slot.id + 1 }, false); common_batch_add(batch, prompt_tokens[slot.n_past], system_tokens.size() + slot_npast, { slot.id + 1 }, false);
if (slot.params.cache_prompt) { if (slot.params.cache_prompt) {
slot.cache_tokens.push_back(prompt_tokens[slot.n_past]); slot.cache_tokens.push_back(prompt_tokens[slot.n_past]);
@ -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;
@ -3014,7 +3014,7 @@ int main(int argc, char ** argv) {
if (with_pieces) { if (with_pieces) {
for (const auto& token : tokens) { for (const auto& token : tokens) {
std::string piece = llama_token_to_piece(ctx_server.ctx, token); std::string piece = common_token_to_piece(ctx_server.ctx, token);
json piece_json; json piece_json;
// Check if the piece is valid UTF-8 // Check if the piece is valid UTF-8
@ -3357,7 +3357,7 @@ int main(int argc, char ** argv) {
} }
// print sample chat example to make it clear which template is used // print sample chat example to make it clear which template is used
LOG_INF("%s: chat template, built_in: %d, chat_example: '%s'\n", __func__, params.chat_template.empty(), llama_chat_format_example(ctx_server.model, params.chat_template).c_str()); LOG_INF("%s: chat template, built_in: %d, chat_example: '%s'\n", __func__, params.chat_template.empty(), common_chat_format_example(ctx_server.model, params.chat_template).c_str());
ctx_server.queue_tasks.on_new_task(std::bind( ctx_server.queue_tasks.on_new_task(std::bind(
&server_context::process_single_task, &ctx_server, std::placeholders::_1)); &server_context::process_single_task, &ctx_server, std::placeholders::_1));

8
examples/server/utils.hpp
View File

@ -57,7 +57,7 @@ static T json_value(const json & body, const std::string & key, const T & defaul
// Format given chat. If tmpl is empty, we take the template from model metadata // Format given chat. If tmpl is empty, we take the template from model metadata
inline std::string format_chat(const struct llama_model * model, const std::string & tmpl, const std::vector<json> & messages) { inline std::string format_chat(const struct llama_model * model, const std::string & tmpl, const std::vector<json> & messages) {
std::vector<llama_chat_msg> chat; std::vector<common_chat_msg> chat;
for (size_t i = 0; i < messages.size(); ++i) { for (size_t i = 0; i < messages.size(); ++i) {
const auto & curr_msg = messages[i]; const auto & curr_msg = messages[i];
@ -84,7 +84,7 @@ inline std::string format_chat(const struct llama_model * model, const std::stri
chat.push_back({role, content}); chat.push_back({role, content});
} }
const auto formatted_chat = llama_chat_apply_template(model, tmpl, chat, true); const auto formatted_chat = common_chat_apply_template(model, tmpl, chat, true);
LOG_DBG("formatted_chat: '%s'\n", formatted_chat.c_str()); LOG_DBG("formatted_chat: '%s'\n", formatted_chat.c_str());
return formatted_chat; return formatted_chat;
@ -246,7 +246,7 @@ template <class Iter>
static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) { static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
std::string ret; std::string ret;
for (; begin != end; ++begin) { for (; begin != end; ++begin) {
ret += llama_token_to_piece(ctx, *begin); ret += common_token_to_piece(ctx, *begin);
} }
return ret; return ret;
@ -254,7 +254,7 @@ static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
// format incomplete utf-8 multibyte character for output // format incomplete utf-8 multibyte character for output
static std::string tokens_to_output_formatted_string(const llama_context * ctx, const llama_token token) { static std::string tokens_to_output_formatted_string(const llama_context * ctx, const llama_token token) {
std::string out = token == -1 ? "" : llama_token_to_piece(ctx, token); std::string out = token == -1 ? "" : common_token_to_piece(ctx, token);
// if the size is 1 and first bit is 1, meaning it's a partial character // if the size is 1 and first bit is 1, meaning it's a partial character
// (size > 1 meaning it's already a known token) // (size > 1 meaning it's already a known token)

80
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
llama_init_result llama_init_tgt = llama_init_from_gpt_params(params); common_init_result llama_init_tgt = common_init_from_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;
llama_init_result llama_init_dft = llama_init_from_gpt_params(params); common_init_result llama_init_dft = common_init_from_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;
@ -124,8 +124,8 @@ int main(int argc, char ** argv) {
if (std::strcmp(token_text_tgt, token_text_dft) != 0) { if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
LOG_ERR("%s: draft model vocab must match target model to use speculation but ", __func__); LOG_ERR("%s: draft model vocab must match target model to use speculation but ", __func__);
LOG_ERR("token %d content differs - target '%s', draft '%s'\n", i, LOG_ERR("token %d content differs - target '%s', draft '%s'\n", i,
llama_token_to_piece(ctx_tgt, i).c_str(), common_token_to_piece(ctx_tgt, i).c_str(),
llama_token_to_piece(ctx_dft, i).c_str()); common_token_to_piece(ctx_dft, i).c_str());
return 1; return 1;
} }
} }
@ -134,7 +134,7 @@ int main(int argc, char ** argv) {
// Tokenize the prompt // Tokenize the prompt
std::vector<llama_token> inp; std::vector<llama_token> inp;
inp = ::llama_tokenize(ctx_tgt, params.prompt, true, true); inp = common_tokenize(ctx_tgt, params.prompt, true, true);
const int max_context_size = llama_n_ctx(ctx_tgt); const int max_context_size = llama_n_ctx(ctx_tgt);
const int max_tokens_list_size = max_context_size - 4; const int max_tokens_list_size = max_context_size - 4;
@ -147,7 +147,7 @@ int main(int argc, char ** argv) {
LOG("\n\n"); LOG("\n\n");
for (auto id : inp) { for (auto id : inp) {
LOG("%s", llama_token_to_piece(ctx_tgt, id).c_str()); LOG("%s", common_token_to_piece(ctx_tgt, id).c_str());
} }
const int n_input = inp.size(); const int n_input = inp.size();
@ -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;
@ -277,13 +277,13 @@ int main(int argc, char ** argv) {
s_keep = s; s_keep = s;
accept = true; accept = true;
token_id = drafts[s].tokens[i_dft]; token_id = drafts[s].tokens[i_dft];
token_str = llama_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;
} else { } else {
LOG_DBG("draft token %d of sequence %d (%d, '%s') rejected\n", i_dft, s, drafts[s].tokens[i_dft], llama_token_to_piece(ctx_tgt, drafts[s].tokens[i_dft]).c_str()); LOG_DBG("draft token %d of sequence %d (%d, '%s') rejected\n", i_dft, s, drafts[s].tokens[i_dft], common_token_to_piece(ctx_tgt, drafts[s].tokens[i_dft]).c_str());
drafts[s].active = false; drafts[s].active = false;
// calculate residual probability // calculate residual probability
@ -349,19 +349,19 @@ 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 = llama_token_to_piece(ctx_tgt, token_id); token_str = common_token_to_piece(ctx_tgt, token_id);
} }
} else { } else {
// greedy verification // greedy verification
// 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 = llama_token_to_piece(ctx_tgt, token_id); token_str = common_token_to_piece(ctx_tgt, token_id);
for (int s = 0; s < n_seq_dft; ++s) { for (int s = 0; s < n_seq_dft; ++s) {
if (!drafts[s].active) { if (!drafts[s].active) {
@ -431,8 +431,8 @@ int main(int argc, char ** argv) {
drafts[0].dists.push_back(std::vector<llama_token_data>()); drafts[0].dists.push_back(std::vector<llama_token_data>());
drafts[0].i_batch_tgt.push_back(0); drafts[0].i_batch_tgt.push_back(0);
llama_batch_clear(batch_dft); common_batch_clear(batch_dft);
llama_batch_add (batch_dft, token_id, n_past_dft, { 0 }, true); common_batch_add (batch_dft, token_id, n_past_dft, { 0 }, true);
llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, -1); llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, -1);
// LOG_DBG("dft batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_dft, batch_dft).c_str()); // LOG_DBG("dft batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_dft, batch_dft).c_str());
@ -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;
@ -461,8 +461,8 @@ int main(int argc, char ** argv) {
drafts[0].drafting = true; drafts[0].drafting = true;
drafts[0].i_batch_dft = 0; drafts[0].i_batch_dft = 0;
llama_batch_clear(batch_tgt); common_batch_clear(batch_tgt);
llama_batch_add (batch_tgt, drafts[0].tokens[0], n_past_tgt, { 0 }, true); common_batch_add (batch_tgt, drafts[0].tokens[0], n_past_tgt, { 0 }, true);
// sample n_draft tokens from the draft model using tree-based sampling // sample n_draft tokens from the draft model using tree-based sampling
for (int i = 0; i < n_draft; ++i) { for (int i = 0; i < n_draft; ++i) {
@ -477,13 +477,13 @@ 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",
k, s, i, cur_p->data[k].id, cur_p->data[k].p, llama_token_to_piece(ctx_dft, cur_p->data[k].id).c_str()); k, s, i, cur_p->data[k].id, cur_p->data[k].p, common_token_to_piece(ctx_dft, cur_p->data[k].id).c_str());
} }
std::vector<int> sa(1, s); std::vector<int> sa(1, s);
@ -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
@ -545,12 +545,12 @@ int main(int argc, char ** argv) {
// add unique drafted tokens to the target batch // add unique drafted tokens to the target batch
drafts[s].i_batch_tgt.push_back(batch_tgt.n_tokens); drafts[s].i_batch_tgt.push_back(batch_tgt.n_tokens);
llama_batch_add(batch_tgt, id, n_past_tgt + i + 1, { s }, true); common_batch_add(batch_tgt, id, n_past_tgt + i + 1, { s }, true);
// add the token to the batch for batched decoding with the draft model // add the token to the batch for batched decoding with the draft model
drafts[s].i_batch_dft = batch_dft.n_tokens; drafts[s].i_batch_dft = batch_dft.n_tokens;
llama_batch_add(batch_dft, id, n_past_cur, { s }, true); common_batch_add(batch_dft, id, n_past_cur, { s }, true);
if (batch_tgt.n_tokens > n_draft) { if (batch_tgt.n_tokens > n_draft) {
drafts[s].drafting = false; drafts[s].drafting = false;
@ -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);

4
examples/tokenize/tokenize.cpp
View File

@ -365,7 +365,7 @@ int main(int raw_argc, char ** raw_argv) {
const bool parse_special = !no_parse_special; const bool parse_special = !no_parse_special;
std::vector<llama_token> tokens; std::vector<llama_token> tokens;
tokens = ::llama_tokenize(model, prompt, add_bos, parse_special); tokens = common_tokenize(model, prompt, add_bos, parse_special);
if (printing_ids) { if (printing_ids) {
printf("["); printf("[");
@ -380,7 +380,7 @@ int main(int raw_argc, char ** raw_argv) {
} else { } else {
bool invalid_utf8 = false; bool invalid_utf8 = false;
printf("%6d -> '", tokens[i]); printf("%6d -> '", tokens[i]);
write_utf8_cstr_to_stdout(llama_token_to_piece(ctx, tokens[i]).c_str(), invalid_utf8); write_utf8_cstr_to_stdout(common_token_to_piece(ctx, tokens[i]).c_str(), invalid_utf8);
if (invalid_utf8) { if (invalid_utf8) {
printf("' (utf-8 decode failure)\n"); printf("' (utf-8 decode failure)\n");
} else { } else {

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

10
tests/test-chat-template.cpp
View File

@ -140,11 +140,11 @@ int main(void) {
// test llama_chat_format_single for system message // test llama_chat_format_single for system message
printf("\n\n=== llama_chat_format_single (system message) ===\n\n"); printf("\n\n=== llama_chat_format_single (system message) ===\n\n");
std::vector<llama_chat_msg> chat2; std::vector<common_chat_msg> chat2;
llama_chat_msg sys_msg{"system", "You are a helpful assistant"}; common_chat_msg sys_msg{"system", "You are a helpful assistant"};
auto fmt_sys = [&](std::string tmpl) { auto fmt_sys = [&](std::string tmpl) {
auto output = llama_chat_format_single(nullptr, tmpl, chat2, sys_msg, false); auto output = common_chat_format_single(nullptr, tmpl, chat2, sys_msg, false);
printf("fmt_sys(%s) : %s\n", tmpl.c_str(), output.c_str()); printf("fmt_sys(%s) : %s\n", tmpl.c_str(), output.c_str());
printf("-------------------------\n"); printf("-------------------------\n");
return output; return output;
@ -160,10 +160,10 @@ int main(void) {
chat2.push_back({"system", "You are a helpful assistant"}); chat2.push_back({"system", "You are a helpful assistant"});
chat2.push_back({"user", "Hello"}); chat2.push_back({"user", "Hello"});
chat2.push_back({"assistant", "I am assistant"}); chat2.push_back({"assistant", "I am assistant"});
llama_chat_msg new_msg{"user", "How are you"}; common_chat_msg new_msg{"user", "How are you"};
auto fmt_single = [&](std::string tmpl) { auto fmt_single = [&](std::string tmpl) {
auto output = llama_chat_format_single(nullptr, tmpl, chat2, new_msg, true); auto output = common_chat_format_single(nullptr, tmpl, chat2, new_msg, true);
printf("fmt_single(%s) : %s\n", tmpl.c_str(), output.c_str()); printf("fmt_single(%s) : %s\n", tmpl.c_str(), output.c_str());
printf("-------------------------\n"); printf("-------------------------\n");
return output; return output;

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);
} }
} }
}); });

14
tests/test-tokenizer-0.cpp
View File

@ -202,7 +202,7 @@ int main(int argc, char **argv) {
for (int i = 0; i < nthread; i++) { for (int i = 0; i < nthread; i++) {
threads[i] = std::thread([&, i]() { threads[i] = std::thread([&, i]() {
for (const auto & test_kv : k_tests) { for (const auto & test_kv : k_tests) {
const std::vector<llama_token> res = llama_tokenize(ctx, test_kv.first, add_special, false); const std::vector<llama_token> res = common_tokenize(ctx, test_kv.first, add_special, false);
// here only print the result of the first thread // here only print the result of the first thread
// because the other threads are running the same tests // because the other threads are running the same tests
@ -212,7 +212,7 @@ int main(int argc, char **argv) {
printf("\n"); printf("\n");
printf("src: '%s'\n", test_kv.first.c_str()); printf("src: '%s'\n", test_kv.first.c_str());
printf("res: '%s'\n", llama_detokenize(ctx, res).c_str()); printf("res: '%s'\n", common_detokenize(ctx, res).c_str());
printf("tok: "); printf("tok: ");
for (const auto & tok : res) { for (const auto & tok : res) {
printf("%d ", tok); printf("%d ", tok);
@ -229,16 +229,16 @@ int main(int argc, char **argv) {
if (!correct) { if (!correct) {
fprintf(stderr, "%s : failed test: '%s'\n", __func__, test_kv.first.c_str()); fprintf(stderr, "%s : failed test: '%s'\n", __func__, test_kv.first.c_str());
fprintf(stderr, "%s : detokenized to: '%s' instead of '%s'\n", __func__, fprintf(stderr, "%s : detokenized to: '%s' instead of '%s'\n", __func__,
llama_detokenize(ctx, res).c_str(), common_detokenize(ctx, res).c_str(),
llama_detokenize(ctx, test_kv.second).c_str()); common_detokenize(ctx, test_kv.second).c_str());
fprintf(stderr, "%s : expected tokens: ", __func__); fprintf(stderr, "%s : expected tokens: ", __func__);
for (const auto & t : test_kv.second) { for (const auto & t : test_kv.second) {
fprintf(stderr, "%6d '%s', ", t, llama_token_to_piece(ctx, t).c_str()); fprintf(stderr, "%6d '%s', ", t, common_token_to_piece(ctx, t).c_str());
} }
fprintf(stderr, "\n"); fprintf(stderr, "\n");
fprintf(stderr, "%s : got tokens: ", __func__); fprintf(stderr, "%s : got tokens: ", __func__);
for (const auto & t : res) { for (const auto & t : res) {
fprintf(stderr, "%6d '%s', ", t, llama_token_to_piece(ctx, t).c_str()); fprintf(stderr, "%6d '%s', ", t, common_token_to_piece(ctx, t).c_str());
} }
fprintf(stderr, "\n"); fprintf(stderr, "\n");
@ -273,7 +273,7 @@ int main(int argc, char **argv) {
{ {
const auto t_start = ggml_time_us(); const auto t_start = ggml_time_us();
res = llama_tokenize(ctx, text, add_special, false); res = common_tokenize(ctx, text, add_special, false);
const auto t_end = ggml_time_us(); const auto t_end = ggml_time_us();

10
tests/test-tokenizer-1-bpe.cpp
View File

@ -78,10 +78,10 @@ int main(int argc, char **argv) {
const int n_vocab = llama_n_vocab(model); const int n_vocab = llama_n_vocab(model);
for (int i = 0; i < n_vocab; ++i) { for (int i = 0; i < n_vocab; ++i) {
std::string str = llama_detokenize(ctx, std::vector<int>(1, i)); std::string str = common_detokenize(ctx, std::vector<int>(1, i));
try { try {
auto cps = unicode_cpts_from_utf8(str); auto cps = unicode_cpts_from_utf8(str);
std::vector<llama_token> tokens = llama_tokenize(ctx, str, false, true); std::vector<llama_token> tokens = common_tokenize(ctx, str, false, true);
if (ignore_merges && tokens.size() > 1) { if (ignore_merges && tokens.size() > 1) {
fprintf(stderr, fprintf(stderr,
"%s : error: token %d detokenizes to '%s'(%zu) but " "%s : error: token %d detokenizes to '%s'(%zu) but "
@ -94,7 +94,7 @@ int main(int argc, char **argv) {
fprintf(stderr, "]\n"); fprintf(stderr, "]\n");
return 2; return 2;
} }
std::string check = llama_detokenize(ctx, tokens); std::string check = common_detokenize(ctx, tokens);
if (check != str) { if (check != str) {
fprintf(stderr, "%s : error: token %d detokenizes to '%s'(%zu) but tokenization of this detokenizes to '%s'(%zu)\n", fprintf(stderr, "%s : error: token %d detokenizes to '%s'(%zu) but tokenization of this detokenizes to '%s'(%zu)\n",
__func__, i, str.c_str(), str.length(), check.c_str(), check.length()); __func__, i, str.c_str(), str.length(), check.c_str(), check.length());
@ -123,8 +123,8 @@ int main(int argc, char **argv) {
} }
std::string str = unicode_cpt_to_utf8(cp); std::string str = unicode_cpt_to_utf8(cp);
std::vector<llama_token> tokens = llama_tokenize(ctx, str, false); std::vector<llama_token> tokens = common_tokenize(ctx, str, false);
std::string check = llama_detokenize(ctx, tokens); std::string check = common_detokenize(ctx, tokens);
if (cp != 9601 && str != check) { if (cp != 9601 && str != check) {
fprintf(stderr, "error: codepoint 0x%x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n", fprintf(stderr, "error: codepoint 0x%x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
cp, check.c_str(), check.length(), str.c_str(), str.length()); cp, check.c_str(), check.length(), str.c_str(), str.length());

10
tests/test-tokenizer-1-spm.cpp
View File

@ -66,9 +66,9 @@ int main(int argc, char ** argv) {
const int n_vocab = llama_n_vocab(model); const int n_vocab = llama_n_vocab(model);
for (int i = 0; i < n_vocab; ++i) { for (int i = 0; i < n_vocab; ++i) {
std::string str = llama_detokenize(ctx, std::vector<int>(1, i), true); std::string str = common_detokenize(ctx, std::vector<int>(1, i), true);
std::vector<llama_token> tokens = llama_tokenize(ctx, str, false, true); std::vector<llama_token> tokens = common_tokenize(ctx, str, false, true);
std::string check = llama_detokenize(ctx, tokens); std::string check = common_detokenize(ctx, tokens);
if (check != str) { if (check != str) {
fprintf(stderr, "%s : error: token %d detokenizes to '%s'(%zu) but tokenization of this detokenizes to '%s'(%zu)\n", fprintf(stderr, "%s : error: token %d detokenizes to '%s'(%zu) but tokenization of this detokenizes to '%s'(%zu)\n",
__func__, i, str.c_str(), str.length(), check.c_str(), check.length()); __func__, i, str.c_str(), str.length(), check.c_str(), check.length());
@ -93,8 +93,8 @@ int main(int argc, char ** argv) {
} }
std::string str = unicode_cpt_to_utf8(cp); std::string str = unicode_cpt_to_utf8(cp);
std::vector<llama_token> tokens = llama_tokenize(ctx, str, false, true); std::vector<llama_token> tokens = common_tokenize(ctx, str, false, true);
std::string check = llama_detokenize(ctx, tokens); std::string check = common_detokenize(ctx, tokens);
if (cp != 9601 && str != check) { if (cp != 9601 && str != check) {
fprintf(stderr, "error: codepoint 0x%x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n", fprintf(stderr, "error: codepoint 0x%x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
cp, check.c_str(), check.length(), str.c_str(), str.length()); cp, check.c_str(), check.length(), str.c_str(), str.length());