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

server: add comments

Browse source
This commit is contained in:
ngxson 2024-01-21 14:58:26 +01:00
parent 12829b2e64
commit 906afe7810
3 changed files with 266 additions and 255 deletions
Download patch file Download diff file Expand all files Collapse all files

194
examples/server/oai.hpp
View file

@ -10,5 +10,199 @@
#include "json.hpp"
#include "utils.hpp"
#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613"
using json = nlohmann::json;
inline static json oaicompat_completion_params_parse(
const json &body /* openai api json semantics */)
{
json llama_params;
llama_params["__oaicompat"] = true;
// Map OpenAI parameters to llama.cpp parameters
//
// For parameters that are defined by the OpenAI documentation (e.g.
// temperature), we explicitly specify OpenAI's intended default; we
// need to do that because sometimes OpenAI disagrees with llama.cpp
//
// https://platform.openai.com/docs/api-reference/chat/create
llama_sampling_params default_sparams;
llama_params["model"] = json_value(body, "model", std::string("unknown"));
llama_params["prompt"] = format_chatml(body["messages"]); // OpenAI 'messages' to llama.cpp 'prompt'
llama_params["cache_prompt"] = json_value(body, "cache_prompt", false);
llama_params["temperature"] = json_value(body, "temperature", 0.0);
llama_params["top_k"] = json_value(body, "top_k", default_sparams.top_k);
llama_params["top_p"] = json_value(body, "top_p", 1.0);
llama_params["n_predict"] = json_value(body, "max_tokens", -1);
llama_params["logit_bias"] = json_value(body, "logit_bias",json::object());
llama_params["frequency_penalty"] = json_value(body, "frequency_penalty", 0.0);
llama_params["presence_penalty"] = json_value(body, "presence_penalty", 0.0);
llama_params["seed"] = json_value(body, "seed", LLAMA_DEFAULT_SEED);
llama_params["stream"] = json_value(body, "stream", false);
llama_params["mirostat"] = json_value(body, "mirostat", default_sparams.mirostat);
llama_params["mirostat_tau"] = json_value(body, "mirostat_tau", default_sparams.mirostat_tau);
llama_params["mirostat_eta"] = json_value(body, "mirostat_eta", default_sparams.mirostat_eta);
llama_params["penalize_nl"] = json_value(body, "penalize_nl", default_sparams.penalize_nl);
llama_params["typical_p"] = json_value(body, "typical_p", default_sparams.typical_p);
llama_params["repeat_last_n"] = json_value(body, "repeat_last_n", default_sparams.penalty_last_n);
llama_params["ignore_eos"] = json_value(body, "ignore_eos", false);
llama_params["tfs_z"] = json_value(body, "tfs_z", default_sparams.tfs_z);
if (body.count("grammar") != 0) {
llama_params["grammar"] = json_value(body, "grammar", json::object());
}
// Handle 'stop' field
if (body.contains("stop") && body["stop"].is_string()) {
llama_params["stop"] = json::array({body["stop"].get<std::string>()});
} else {
llama_params["stop"] = json_value(body, "stop", json::array());
}
// Ensure there is ChatML-specific end sequence among stop words
llama_params["stop"].push_back("<|im_end|>");
return llama_params;
}
inline static json format_final_response_oaicompat(const json &request, const task_result &response, bool streaming = false)
{
json result = response.result_json;
bool stopped_word = result.count("stopped_word") != 0;
bool stopped_eos = json_value(result, "stopped_eos", false);
int num_tokens_predicted = json_value(result, "tokens_predicted", 0);
int num_prompt_tokens = json_value(result, "tokens_evaluated", 0);
std::string content = json_value(result, "content", std::string(""));
std::string finish_reason = "length";
if (stopped_word || stopped_eos) {
finish_reason = "stop";
}
json choices =
streaming ? json::array({json{{"finish_reason", finish_reason},
{"index", 0},
{"delta", json::object()}}})
: json::array({json{{"finish_reason", finish_reason},
{"index", 0},
{"message", json{{"content", content},
{"role", "assistant"}}}}});
std::time_t t = std::time(0);
json res =
json{{"choices", choices},
{"created", t},
{"model",
json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
{"object", streaming ? "chat.completion.chunk" : "chat.completion"},
{"usage",
json{{"completion_tokens", num_tokens_predicted},
{"prompt_tokens", num_prompt_tokens},
{"total_tokens", num_tokens_predicted + num_prompt_tokens}}},
{"id", gen_chatcmplid()}};
if (server_verbose) {
res["__verbose"] = result;
}
if (result.contains("completion_probabilities")) {
res["completion_probabilities"] = json_value(result, "completion_probabilities", json::array());
}
return res;
}
// return value is vector as there is one case where we might need to generate two responses
inline static std::vector<json> format_partial_response_oaicompat(const task_result &response) {
json result = response.result_json;
if (!result.contains("model") || !result.contains("oaicompat_token_ctr")) {
return std::vector<json>({response.result_json});
}
bool first = json_value(result, "oaicompat_token_ctr", 0) == 0;
std::string modelname = json_value(result, "model", std::string(DEFAULT_OAICOMPAT_MODEL));
bool stopped_word = json_value(result, "stopped_word", false);
bool stopped_eos = json_value(result, "stopped_eos", false);
bool stopped_limit = json_value(result, "stopped_limit", false);
std::string content = json_value(result, "content", std::string(""));
std::string finish_reason;
if (stopped_word || stopped_eos) {
finish_reason = "stop";
}
if (stopped_limit) {
finish_reason = "length";
}
std::time_t t = std::time(0);
json choices;
if (!finish_reason.empty()) {
choices = json::array({json{{"finish_reason", finish_reason},
{"index", 0},
{"delta", json::object()}}});
} else {
if (first) {
if (content.empty()) {
choices = json::array({json{{"finish_reason", nullptr},
{"index", 0},
{"delta", json{{"role", "assistant"}}}}});
} else {
// We have to send this as two updates to conform to openai behavior
json initial_ret = json{{"choices", json::array({json{
{"finish_reason", nullptr},
{"index", 0},
{"delta", json{
{"role", "assistant"}
}}}})},
{"created", t},
{"id", gen_chatcmplid()},
{"model", modelname},
{"object", "chat.completion.chunk"}};
json second_ret = json{
{"choices", json::array({json{{"finish_reason", nullptr},
{"index", 0},
{"delta", json{
{"content", content}}}
}})},
{"created", t},
{"id", gen_chatcmplid()},
{"model", modelname},
{"object", "chat.completion.chunk"}};
return std::vector<json>({initial_ret, second_ret});
}
} else {
// Some idiosyncrasy in task processing logic makes several trailing calls
// with empty content, we ignore these at the calee site.
if (content.empty()) {
return std::vector<json>({json::object()});
}
choices = json::array({json{
{"finish_reason", nullptr},
{"index", 0},
{"delta",
json{
{"content", content},
}},
}});
}
}
json ret = json{{"choices", choices},
{"created", t},
{"id", gen_chatcmplid()},
{"model", modelname},
{"object", "chat.completion.chunk"}};
return std::vector<json>({ret});
}

247
examples/server/server.cpp
View file

@ -2,6 +2,7 @@
#include "llama.h"
#include "grammar-parser.h"
#include "utils.hpp"
#include "oai.hpp"
#include "../llava/clip.h"
@ -29,8 +30,6 @@
#include <condition_variable>
#include <atomic>
#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613"
using json = nlohmann::json;
struct server_params
@ -45,9 +44,6 @@ struct server_params
bool server_verbose = false;
json oaicompat_completion_params_parse(const json &body);
std::string format_chatml(std::vector<json> messages);
static size_t common_part(const std::vector<llama_token> &a, const std::vector<llama_token> &b)
{
size_t i;
@ -143,15 +139,6 @@ static json probs_vector_to_json(const llama_context *ctx, const std::vector<com
return out;
}
template <typename T>
static T json_value(const json &body, const std::string &key, const T &default_value)
{
// Fallback null to default value
return body.contains(key) && !body.at(key).is_null()
? body.value(key, default_value)
: default_value;
}
struct llama_client_slot
{
int id;
@ -2264,239 +2251,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
}
}
static std::string random_string()
{
static const std::string str("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz");
std::random_device rd;
std::mt19937 generator(rd());
std::string result(32, ' ');
for (int i = 0; i < 32; ++i) {
result[i] = str[generator() % str.size()];
}
return result;
}
static std::string gen_chatcmplid()
{
std::stringstream chatcmplid;
chatcmplid << "chatcmpl-" << random_string();
return chatcmplid.str();
}
std::string format_chatml(std::vector<json> messages)
{
std::ostringstream chatml_msgs;
for (auto it = messages.begin(); it != messages.end(); ++it) {
chatml_msgs << "<|im_start|>"
<< json_value(*it, "role", std::string("user")) << '\n';
chatml_msgs << json_value(*it, "content", std::string(""))
<< "<|im_end|>\n";
}
chatml_msgs << "<|im_start|>assistant" << '\n';
return chatml_msgs.str();
}
/* llama.cpp completion api semantics */
json oaicompat_completion_params_parse(
const json &body /* openai api json semantics */)
{
json llama_params;
llama_params["__oaicompat"] = true;
// Map OpenAI parameters to llama.cpp parameters
//
// For parameters that are defined by the OpenAI documentation (e.g.
// temperature), we explicitly specify OpenAI's intended default; we
// need to do that because sometimes OpenAI disagrees with llama.cpp
//
// https://platform.openai.com/docs/api-reference/chat/create
llama_sampling_params default_sparams;
llama_params["model"] = json_value(body, "model", std::string("unknown"));
llama_params["prompt"] = format_chatml(body["messages"]); // OpenAI 'messages' to llama.cpp 'prompt'
llama_params["cache_prompt"] = json_value(body, "cache_prompt", false);
llama_params["temperature"] = json_value(body, "temperature", 0.0);
llama_params["top_k"] = json_value(body, "top_k", default_sparams.top_k);
llama_params["top_p"] = json_value(body, "top_p", 1.0);
llama_params["n_predict"] = json_value(body, "max_tokens", -1);
llama_params["logit_bias"] = json_value(body, "logit_bias",json::object());
llama_params["frequency_penalty"] = json_value(body, "frequency_penalty", 0.0);
llama_params["presence_penalty"] = json_value(body, "presence_penalty", 0.0);
llama_params["seed"] = json_value(body, "seed", LLAMA_DEFAULT_SEED);
llama_params["stream"] = json_value(body, "stream", false);
llama_params["mirostat"] = json_value(body, "mirostat", default_sparams.mirostat);
llama_params["mirostat_tau"] = json_value(body, "mirostat_tau", default_sparams.mirostat_tau);
llama_params["mirostat_eta"] = json_value(body, "mirostat_eta", default_sparams.mirostat_eta);
llama_params["penalize_nl"] = json_value(body, "penalize_nl", default_sparams.penalize_nl);
llama_params["typical_p"] = json_value(body, "typical_p", default_sparams.typical_p);
llama_params["repeat_last_n"] = json_value(body, "repeat_last_n", default_sparams.penalty_last_n);
llama_params["ignore_eos"] = json_value(body, "ignore_eos", false);
llama_params["tfs_z"] = json_value(body, "tfs_z", default_sparams.tfs_z);
if (body.count("grammar") != 0) {
llama_params["grammar"] = json_value(body, "grammar", json::object());
}
// Handle 'stop' field
if (body.contains("stop") && body["stop"].is_string()) {
llama_params["stop"] = json::array({body["stop"].get<std::string>()});
} else {
llama_params["stop"] = json_value(body, "stop", json::array());
}
// Ensure there is ChatML-specific end sequence among stop words
llama_params["stop"].push_back("<|im_end|>");
return llama_params;
}
static json format_final_response_oaicompat(const json &request, const task_result &response, bool streaming = false)
{
json result = response.result_json;
bool stopped_word = result.count("stopped_word") != 0;
bool stopped_eos = json_value(result, "stopped_eos", false);
int num_tokens_predicted = json_value(result, "tokens_predicted", 0);
int num_prompt_tokens = json_value(result, "tokens_evaluated", 0);
std::string content = json_value(result, "content", std::string(""));
std::string finish_reason = "length";
if (stopped_word || stopped_eos) {
finish_reason = "stop";
}
json choices =
streaming ? json::array({json{{"finish_reason", finish_reason},
{"index", 0},
{"delta", json::object()}}})
: json::array({json{{"finish_reason", finish_reason},
{"index", 0},
{"message", json{{"content", content},
{"role", "assistant"}}}}});
std::time_t t = std::time(0);
json res =
json{{"choices", choices},
{"created", t},
{"model",
json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
{"object", streaming ? "chat.completion.chunk" : "chat.completion"},
{"usage",
json{{"completion_tokens", num_tokens_predicted},
{"prompt_tokens", num_prompt_tokens},
{"total_tokens", num_tokens_predicted + num_prompt_tokens}}},
{"id", gen_chatcmplid()}};
if (server_verbose) {
res["__verbose"] = result;
}
if (result.contains("completion_probabilities")) {
res["completion_probabilities"] = json_value(result, "completion_probabilities", json::array());
}
return res;
}
// return value is vector as there is one case where we might need to generate two responses
static std::vector<json> format_partial_response_oaicompat(const task_result &response) {
json result = response.result_json;
if (!result.contains("model") || !result.contains("oaicompat_token_ctr")) {
return std::vector<json>({response.result_json});
}
bool first = json_value(result, "oaicompat_token_ctr", 0) == 0;
std::string modelname = json_value(result, "model", std::string(DEFAULT_OAICOMPAT_MODEL));
bool stopped_word = json_value(result, "stopped_word", false);
bool stopped_eos = json_value(result, "stopped_eos", false);
bool stopped_limit = json_value(result, "stopped_limit", false);
std::string content = json_value(result, "content", std::string(""));
std::string finish_reason;
if (stopped_word || stopped_eos) {
finish_reason = "stop";
}
if (stopped_limit) {
finish_reason = "length";
}
std::time_t t = std::time(0);
json choices;
if (!finish_reason.empty()) {
choices = json::array({json{{"finish_reason", finish_reason},
{"index", 0},
{"delta", json::object()}}});
} else {
if (first) {
if (content.empty()) {
choices = json::array({json{{"finish_reason", nullptr},
{"index", 0},
{"delta", json{{"role", "assistant"}}}}});
} else {
// We have to send this as two updates to conform to openai behavior
json initial_ret = json{{"choices", json::array({json{
{"finish_reason", nullptr},
{"index", 0},
{"delta", json{
{"role", "assistant"}
}}}})},
{"created", t},
{"id", gen_chatcmplid()},
{"model", modelname},
{"object", "chat.completion.chunk"}};
json second_ret = json{
{"choices", json::array({json{{"finish_reason", nullptr},
{"index", 0},
{"delta", json{
{"content", content}}}
}})},
{"created", t},
{"id", gen_chatcmplid()},
{"model", modelname},
{"object", "chat.completion.chunk"}};
return std::vector<json>({initial_ret, second_ret});
}
} else {
// Some idiosyncrasy in task processing logic makes several trailing calls
// with empty content, we ignore these at the calee site.
if (content.empty()) {
return std::vector<json>({json::object()});
}
choices = json::array({json{
{"finish_reason", nullptr},
{"index", 0},
{"delta",
json{
{"content", content},
}},
}});
}
}
json ret = json{{"choices", choices},
{"created", t},
{"id", gen_chatcmplid()},
{"model", modelname},
{"object", "chat.completion.chunk"}};
return std::vector<json>({ret});
}
static json format_partial_response(
llama_server_context &llama, llama_client_slot *slot, const std::string &content, const std::vector<completion_token_output> &probs
) {

80
examples/server/utils.hpp
View file

@ -154,6 +154,35 @@ static inline void server_log(const char *level, const char *function, int line,
fflush(stdout);
}
//
// server utils
//
template <typename T>
static T json_value(const json &body, const std::string &key, const T &default_value)
{
// Fallback null to default value
return body.contains(key) && !body.at(key).is_null()
? body.value(key, default_value)
: default_value;
}
inline std::string format_chatml(std::vector<json> messages)
{
std::ostringstream chatml_msgs;
for (auto it = messages.begin(); it != messages.end(); ++it) {
chatml_msgs << "<|im_start|>"
<< json_value(*it, "role", std::string("user")) << '\n';
chatml_msgs << json_value(*it, "content", std::string(""))
<< "<|im_end|>\n";
}
chatml_msgs << "<|im_start|>assistant" << '\n';
return chatml_msgs.str();
}
//
// work queue utils
//
@ -168,6 +197,7 @@ struct llama_server_queue {
std::function<void(T)> callback_new_task;
std::function<void(void)> callback_all_task_finished;
// Add a new task to the end of the queue
int post(T task) {
std::unique_lock<std::mutex> lock(mutex_tasks);
task.id = id++;
@ -176,24 +206,29 @@ struct llama_server_queue {
return task.id;
}
// Add a new task, but defer until the next loop
void defer(T task) {
std::unique_lock<std::mutex> lock(mutex_tasks);
queue_tasks_deferred.push_back(std::move(task));
}
// Get the next task id
int get_next_id() {
std::unique_lock<std::mutex> lock(mutex_tasks);
return id++;
}
// Register function to process a new task
void on_new_task(std::function<void(T)> callback) {
callback_new_task = callback;
}
// Register the function to be called when the batch of tasks is finished
void on_all_tasks_finished(std::function<void(void)> callback) {
callback_all_task_finished = callback;
}
// Start the main loop. This call is blocking
void start_loop() {
while (true) {
// new task arrived
@ -215,13 +250,8 @@ struct llama_server_queue {
// move deferred tasks back to main loop
{
std::unique_lock<std::mutex> lock(mutex_tasks);
//queue_tasks.insert(
// queue_tasks.end(),
// std::make_move_iterator(queue_tasks_deferred.begin()),
// std::make_move_iterator(queue_tasks_deferred.end())
//);
for (auto & task : queue_tasks_deferred) {
queue_tasks.push_back(task);
queue_tasks.push_back(std::move(task));
}
queue_tasks_deferred.clear();
lock.unlock();
@ -245,12 +275,14 @@ struct llama_server_queue {
struct llama_server_response_event {
typedef std::function<void(int, int, task_result&)> callback_multitask_t;
std::vector<task_result> queue_results;
callback_multitask_t callback_update_multitask;
// for keeping track of all tasks waiting for the result
std::mutex mutex_task_ids;
std::set<int> waiting_task_ids;
// the main result queue
std::vector<task_result> queue_results;
std::mutex mutex_results;
std::condition_variable condition_results;
callback_multitask_t callback_update_multitask;
void add_waiting_task_id(int task_id) {
std::unique_lock<std::mutex> lock(mutex_task_ids);
@ -262,6 +294,7 @@ struct llama_server_response_event {
waiting_task_ids.erase(task_id);
}
// This function blocks the thread until there is a response for this task_id
task_result recv(int task_id) {
while (true)
{
@ -286,16 +319,18 @@ struct llama_server_response_event {
// should never reach here
}
// Register the function to update multitask
void on_multitask_update(callback_multitask_t callback) {
callback_update_multitask = callback;
}
// Send a new result to a waiting task_id
void send(task_result result) {
std::unique_lock<std::mutex> lock(mutex_results);
std::unique_lock<std::mutex> lock1(mutex_task_ids);
LOG_VERBOSE("send new result", {});
for (auto& task_id : waiting_task_ids) {
LOG_TEE("waiting task id %i \n", task_id);
// LOG_TEE("waiting task id %i \n", task_id);
// for now, tasks that have associated parent multitasks just get erased once multitask picks up the result
if (result.multitask_id == task_id)
{
@ -387,4 +422,31 @@ static inline std::vector<uint8_t> base64_decode(const std::string & encoded_str
}
return ret;
}
//
// random string / id
//
static std::string random_string()
{
static const std::string str("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz");
std::random_device rd;
std::mt19937 generator(rd());
std::string result(32, ' ');
for (int i = 0; i < 32; ++i) {
result[i] = str[generator() % str.size()];
}
return result;
}
static std::string gen_chatcmplid()
{
std::stringstream chatcmplid;
chatcmplid << "chatcmpl-" << random_string();
return chatcmplid.str();
}