tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-27 12:05:03 +00:00
Code Issues Packages Projects Releases Wiki Activity

server : (refactoring) do not rely on JSON internally (#10643)

Browse Source 
* server : (refactoring) reduce usage of json internally

* move all response types to struct

* wip [no ci]

* many fixes

* add virtual function

* fix index

* minor style fix

* add std::move

* refactor handle_completions_generic

* add virtual functions

* remove server.hpp

* clarify server_sent_event RFC specs

* apply review comments

* fix model_alias and completion_probabilities

* small clean up

* remove virtual for to_json_oai_compat()

* naming oai_compat --> oaicompat

* fix unwanted recursive call

* update docs
This commit is contained in:
Xuan Son Nguyen
2024-12-06 11:14:32 +01:00
committed by GitHub
parent 7736837d62
commit 6c5bc0625f
8 changed files with 985 additions and 697 deletions
Download Patch File Download Diff File Expand all files Collapse all files

249
examples/server/utils.hpp
View File

@ -20,6 +20,7 @@
#include <sstream>
#include <string>
#include <vector>
#include <memory>
#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613"
@ -40,17 +41,6 @@ using json = nlohmann::ordered_json;
#define QUE_ERR(fmt, ...) LOG_ERR("que %12.*s: " fmt, 12, __func__, __VA_ARGS__)
#define QUE_DBG(fmt, ...) LOG_DBG("que %12.*s: " fmt, 12, __func__, __VA_ARGS__)
// https://community.openai.com/t/openai-chat-list-of-error-codes-and-types/357791/11
enum error_type {
ERROR_TYPE_INVALID_REQUEST,
ERROR_TYPE_AUTHENTICATION,
ERROR_TYPE_SERVER,
ERROR_TYPE_NOT_FOUND,
ERROR_TYPE_PERMISSION,
ERROR_TYPE_UNAVAILABLE, // custom error
ERROR_TYPE_NOT_SUPPORTED, // custom error
};
template <typename T>
static T json_value(const json & body, const std::string & key, const T & default_value) {
// Fallback null to default value
@ -485,48 +475,11 @@ static std::string tokens_to_output_formatted_string(const llama_context * ctx,
return out;
}
struct completion_token_output {
llama_token tok;
std::string text_to_send;
struct token_prob {
llama_token tok;
float prob;
};
std::vector<token_prob> probs;
};
// convert a vector of completion_token_output to json
static json probs_vector_to_json(const llama_context * ctx, const std::vector<completion_token_output> & probs) {
json out = json::array();
for (const auto & prob : probs) {
json probs_for_token = json::array();
for (const auto & p : prob.probs) {
const std::string tok_str = tokens_to_output_formatted_string(ctx, p.tok);
probs_for_token.push_back(json {
{"tok_str", tok_str},
{"prob", p.prob},
});
}
const std::string tok_str = tokens_to_output_formatted_string(ctx, prob.tok);
out.push_back(json {
{"content", tok_str},
{"probs", probs_for_token},
});
}
return out;
}
static bool server_sent_event(httplib::DataSink & sink, const char * event, const json & data) {
const std::string str =
std::string(event) + ": " +
data.dump(-1, ' ', false, json::error_handler_t::replace) +
"\n\n"; // note: these newlines are important (not sure why though, if you know, add a comment to explain)
"\n\n"; // required by RFC 8895 - A message is terminated by a blank line (two line terminators in a row).
LOG_DBG("data stream, to_send: %s", str.c_str());
@ -604,164 +557,6 @@ static json oaicompat_completion_params_parse(
return llama_params;
}
static json format_final_response_oaicompat(const json & request, const json & result, const std::string & completion_id, bool streaming = false, bool verbose = false) {
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", completion_id}
};
// extra fields for debugging purposes
if (verbose) {
res["__verbose"] = result;
}
if (result.contains("completion_probabilities")) {
res["completion_probabilities"] = json_value(result, "completion_probabilities", json::array());
}
if (result.contains("timings")) {
res.push_back({"timings", json_value(result, "timings", json::object())});
}
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 json & result, const std::string & completion_id) {
if (!result.contains("model") || !result.contains("oaicompat_token_ctr")) {
return std::vector<json>({result});
}
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", completion_id},
{"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", completion_id},
{"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", completion_id},
{"model", modelname},
{"object", "chat.completion.chunk"}
};
if (result.contains("timings")) {
ret.push_back({"timings", json_value(result, "timings", json::object())});
}
if (!finish_reason.empty()) {
int num_tokens_predicted = json_value(result, "tokens_predicted", 0);
int num_prompt_tokens = json_value(result, "tokens_evaluated", 0);
ret.push_back({"usage", json {
{"completion_tokens", num_tokens_predicted},
{"prompt_tokens", num_prompt_tokens},
{"total_tokens", num_tokens_predicted + num_prompt_tokens}
}});
}
return std::vector<json>({ret});
}
static json format_embeddings_response_oaicompat(const json & request, const json & embeddings) {
json data = json::array();
int i = 0;
@ -853,43 +648,3 @@ static json format_detokenized_response(const std::string & content) {
{"content", content}
};
}
static json format_error_response(const std::string & message, const enum error_type type) {
std::string type_str;
int code = 500;
switch (type) {
case ERROR_TYPE_INVALID_REQUEST:
type_str = "invalid_request_error";
code = 400;
break;
case ERROR_TYPE_AUTHENTICATION:
type_str = "authentication_error";
code = 401;
break;
case ERROR_TYPE_NOT_FOUND:
type_str = "not_found_error";
code = 404;
break;
case ERROR_TYPE_SERVER:
type_str = "server_error";
code = 500;
break;
case ERROR_TYPE_PERMISSION:
type_str = "permission_error";
code = 403;
break;
case ERROR_TYPE_NOT_SUPPORTED:
type_str = "not_supported_error";
code = 501;
break;
case ERROR_TYPE_UNAVAILABLE:
type_str = "unavailable_error";
code = 503;
break;
}
return json {
{"code", code},
{"message", message},
{"type", type_str},
};
}