root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-26 11:24:35 +00:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

server : various fixes (#10704)

Browse Source 
* server : various fixes

ggml-ci

* server : show curent seed in slot_params

ggml-ci

* fix /slots endpoint

* Update examples/server/server.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* server : reflect endpoint response changes in the readme

ggml-ci

---------

Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
Co-authored-by: Xuan Son Nguyen <thichthat@gmail.com>
This commit is contained in:
Georgi Gerganov 2024-12-07 18:02:05 +02:00 committed by GitHub
parent 19d8762ab6
commit ce4a7b8493
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
4 changed files with 178 additions and 97 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
examples/server/CMakeLists.txt
View File

@ -34,14 +34,6 @@ endforeach()
add_executable(${TARGET} ${TARGET_SRCS}) add_executable(${TARGET} ${TARGET_SRCS})
install(TARGETS ${TARGET} RUNTIME) install(TARGETS ${TARGET} RUNTIME)
# clean up generated files in pre-build step
foreach(asset ${PUBLIC_ASSETS})
set(output "${CMAKE_CURRENT_BINARY_DIR}/${asset}.hpp")
add_custom_command(TARGET ${TARGET} PRE_BUILD
COMMAND "${CMAKE_COMMAND}" -E remove -f "${output}"
)
endforeach()
target_link_libraries(${TARGET} PRIVATE common ${CMAKE_THREAD_LIBS_INIT}) target_link_libraries(${TARGET} PRIVATE common ${CMAKE_THREAD_LIBS_INIT})
if (LLAMA_SERVER_SSL) if (LLAMA_SERVER_SSL)

177
examples/server/README.md
View File

@ -618,9 +618,76 @@ This endpoint is public (no API key check). By default, it is read-only. To make
```json ```json
{ {
"default_generation_settings": { ... }, "default_generation_settings": {
"id": 0,
"id_task": -1,
"n_ctx": 1024,
"speculative": false,
"is_processing": false,
"params": {
"n_predict": -1,
"seed": 4294967295,
"temperature": 0.800000011920929,
"dynatemp_range": 0.0,
"dynatemp_exponent": 1.0,
"top_k": 40,
"top_p": 0.949999988079071,
"min_p": 0.05000000074505806,
"xtc_probability": 0.0,
"xtc_threshold": 0.10000000149011612,
"typical_p": 1.0,
"repeat_last_n": 64,
"repeat_penalty": 1.0,
"presence_penalty": 0.0,
"frequency_penalty": 0.0,
"dry_multiplier": 0.0,
"dry_base": 1.75,
"dry_allowed_length": 2,
"dry_penalty_last_n": -1,
"dry_sequence_breakers": [
"\n",
":",
"\"",
"*"
],
"mirostat": 0,
"mirostat_tau": 5.0,
"mirostat_eta": 0.10000000149011612,
"penalize_nl": false,
"stop": [],
"max_tokens": -1,
"n_keep": 0,
"n_discard": 0,
"ignore_eos": false,
"stream": true,
"n_probs": 0,
"min_keep": 0,
"grammar": "",
"samplers": [
"dry",
"top_k",
"typ_p",
"top_p",
"min_p",
"xtc",
"temperature"
],
"speculative.n_max": 16,
"speculative.n_min": 5,
"speculative.p_min": 0.8999999761581421,
"timings_per_token": false
},
"prompt": "",
"next_token": {
"has_next_token": true,
"has_new_line": false,
"n_remain": -1,
"n_decoded": 0,
"stopping_word": ""
}
},
"total_slots": 1, "total_slots": 1,
"chat_template": "" "chat_template": "..."
} }
``` ```
@ -740,54 +807,72 @@ Example:
```json ```json
[ [
{ {
"dynatemp_exponent": 1.0,
"dynatemp_range": 0.0,
"frequency_penalty": 0.0,
"grammar": "",
"id": 0, "id": 0,
"ignore_eos": false, "id_task": -1,
"n_ctx": 1024,
"speculative": false,
"is_processing": false, "is_processing": false,
"logit_bias": [], "params": {
"min_p": 0.05000000074505806, "n_predict": -1,
"mirostat": 0, "seed": 4294967295,
"mirostat_eta": 0.10000000149011612, "temperature": 0.800000011920929,
"mirostat_tau": 5.0, "dynatemp_range": 0.0,
"model": "llama-2-7b-32k-instruct.Q2_K.gguf", "dynatemp_exponent": 1.0,
"n_ctx": 2048,
"n_keep": 0,
"n_predict": 100000,
"n_probs": 0,
"next_token": {
"has_next_token": true,
"n_remain": -1,
"n_decoded": 0,
"stopped_eos": false,
"stopped_limit": false,
"stopped_word": false,
"stopping_word": ""
},
"penalize_nl": true,
"presence_penalty": 0.0,
"prompt": "Say hello to llama.cpp",
"repeat_last_n": 64,
"repeat_penalty": 1.100000023841858,
"samplers": [
"top_k",
"typical_p",
"top_p",
"min_p",
"temperature"
],
"seed": 42,
"stop": [
"\n"
],
"stream": false,
"task_id": 0,
"temperature": 0.0,
"top_k": 40, "top_k": 40,
"top_p": 0.949999988079071, "top_p": 0.949999988079071,
"typical_p": 1.0 "min_p": 0.05000000074505806,
"xtc_probability": 0.0,
"xtc_threshold": 0.10000000149011612,
"typical_p": 1.0,
"repeat_last_n": 64,
"repeat_penalty": 1.0,
"presence_penalty": 0.0,
"frequency_penalty": 0.0,
"dry_multiplier": 0.0,
"dry_base": 1.75,
"dry_allowed_length": 2,
"dry_penalty_last_n": -1,
"dry_sequence_breakers": [
"\n",
":",
"\"",
"*"
],
"mirostat": 0,
"mirostat_tau": 5.0,
"mirostat_eta": 0.10000000149011612,
"penalize_nl": false,
"stop": [],
"max_tokens": -1,
"n_keep": 0,
"n_discard": 0,
"ignore_eos": false,
"stream": true,
"n_probs": 0,
"min_keep": 0,
"grammar": "",
"samplers": [
"dry",
"top_k",
"typ_p",
"top_p",
"min_p",
"xtc",
"temperature"
],
"speculative.n_max": 16,
"speculative.n_min": 5,
"speculative.p_min": 0.8999999761581421,
"timings_per_token": false
},
"prompt": "",
"next_token": {
"has_next_token": true,
"has_new_line": false,
"n_remain": -1,
"n_decoded": 0,
"stopping_word": ""
}
} }
] ]
``` ```

68
examples/server/server.cpp
View File

@ -122,11 +122,6 @@ struct slot_params {
struct common_params_sampling sampling; struct common_params_sampling sampling;
struct common_params_speculative speculative; struct common_params_speculative speculative;
// params only used in to_json()
int32_t n_ctx;
uint32_t seed_cur;
bool can_speculative;
// OAI-compat fields // OAI-compat fields
bool verbose = false; bool verbose = false;
bool oaicompat = false; bool oaicompat = false;
@ -134,7 +129,7 @@ struct slot_params {
std::string oaicompat_model; std::string oaicompat_model;
std::string oaicompat_cmpl_id; std::string oaicompat_cmpl_id;
json to_json() { json to_json() const {
std::vector<std::string> samplers; std::vector<std::string> samplers;
samplers.reserve(sampling.samplers.size()); samplers.reserve(sampling.samplers.size());
for (const auto & sampler : sampling.samplers) { for (const auto & sampler : sampling.samplers) {
@ -142,8 +137,8 @@ struct slot_params {
} }
return json { return json {
{"n_ctx", n_ctx},
{"n_predict", n_predict}, // Server configured n_predict {"n_predict", n_predict}, // Server configured n_predict
{"seed", sampling.seed},
{"temperature", sampling.temp}, {"temperature", sampling.temp},
{"dynatemp_range", sampling.dynatemp_range}, {"dynatemp_range", sampling.dynatemp_range},
{"dynatemp_exponent", sampling.dynatemp_exponent}, {"dynatemp_exponent", sampling.dynatemp_exponent},
@ -177,7 +172,6 @@ struct slot_params {
{"min_keep", sampling.min_keep}, {"min_keep", sampling.min_keep},
{"grammar", sampling.grammar}, {"grammar", sampling.grammar},
{"samplers", samplers}, {"samplers", samplers},
{"speculative", can_speculative},
{"speculative.n_max", speculative.n_max}, {"speculative.n_max", speculative.n_max},
{"speculative.n_min", speculative.n_min}, {"speculative.n_min", speculative.n_min},
{"speculative.p_min", speculative.p_min}, {"speculative.p_min", speculative.p_min},
@ -483,12 +477,6 @@ struct server_task_result_cmpl_partial : server_task_result {
return std::vector<json>({initial_ret, second_ret}); return std::vector<json>({initial_ret, second_ret});
} }
} else { } 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{ choices = json::array({json{
{"finish_reason", nullptr}, {"finish_reason", nullptr},
{"index", 0}, {"index", 0},
@ -722,6 +710,7 @@ struct server_slot {
llama_batch batch_spec = {}; llama_batch batch_spec = {};
llama_context * ctx = nullptr;
llama_context * ctx_dft = nullptr; llama_context * ctx_dft = nullptr;
common_speculative * spec = nullptr; common_speculative * spec = nullptr;
@ -906,6 +895,27 @@ struct server_slot {
t_token_generation, n_decoded, t_gen, n_gen_second, t_token_generation, n_decoded, t_gen, n_gen_second,
t_prompt_processing + t_token_generation, n_prompt_tokens_processed + n_decoded); t_prompt_processing + t_token_generation, n_prompt_tokens_processed + n_decoded);
} }
json to_json() const {
return json {
{"id", id},
{"id_task", id_task},
{"n_ctx", n_ctx},
{"speculative", can_speculate()},
{"is_processing", is_processing()},
{"params", params.to_json()},
{"prompt", common_detokenize(ctx, prompt_tokens)},
{"next_token",
{
{"has_next_token", has_next_token},
{"has_new_line", has_new_line},
{"n_remain", n_remaining},
{"n_decoded", n_decoded},
{"stopping_word", stopping_word},
}
},
};
}
}; };
struct server_metrics { struct server_metrics {
@ -1338,6 +1348,7 @@ struct server_context {
server_slot slot; server_slot slot;
slot.id = i; slot.id = i;
slot.ctx = ctx;
slot.n_ctx = n_ctx_slot; slot.n_ctx = n_ctx_slot;
slot.n_predict = params_base.n_predict; slot.n_predict = params_base.n_predict;
@ -1370,8 +1381,7 @@ struct server_context {
slots.push_back(slot); slots.push_back(slot);
} }
default_generation_settings_for_props = slots[0].params.to_json(); default_generation_settings_for_props = slots[0].to_json();
default_generation_settings_for_props["seed"] = -1;
// the update_slots() logic will always submit a maximum of n_batch or n_parallel tokens // the update_slots() logic will always submit a maximum of n_batch or n_parallel tokens
// note that n_batch can be > n_ctx (e.g. for non-causal attention models such as BERT where the KV cache is not used) // note that n_batch can be > n_ctx (e.g. for non-causal attention models such as BERT where the KV cache is not used)
@ -1848,8 +1858,9 @@ struct server_context {
queue_results.send(std::move(res)); queue_results.send(std::move(res));
} }
void send_partial_response(server_slot & slot, completion_token_output tkn) { void send_partial_response(server_slot & slot, const completion_token_output & tkn) {
auto res = std::make_unique<server_task_result_cmpl_partial>(); auto res = std::make_unique<server_task_result_cmpl_partial>();
res->id = slot.id_task; res->id = slot.id_task;
res->index = slot.index; res->index = slot.index;
res->content = tkn.text_to_send; res->content = tkn.text_to_send;
@ -1869,6 +1880,7 @@ struct server_context {
// populate res.probs_output // populate res.probs_output
if (slot.params.sampling.n_probs > 0) { if (slot.params.sampling.n_probs > 0) {
const llama_tokens to_send_toks = common_tokenize(ctx, tkn.text_to_send, false); const llama_tokens 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());
@ -1891,7 +1903,8 @@ struct server_context {
void send_final_response(server_slot & slot) { void send_final_response(server_slot & slot) {
if (slot.params.stream) { if (slot.params.stream) {
// if in stream mode, send the last partial response // if in stream mode, send the last partial response
return send_partial_response(slot, {0, "", {}}); send_partial_response(slot, {0, "", {}});
return;
} }
auto res = std::make_unique<server_task_result_cmpl_final>(); auto res = std::make_unique<server_task_result_cmpl_final>();
@ -2012,6 +2025,7 @@ struct server_context {
std::vector<server_task> tasks; std::vector<server_task> tasks;
auto create_task = [&](json & task_data, llama_tokens & prompt_tokens) { auto create_task = [&](json & task_data, llama_tokens & prompt_tokens) {
SRV_DBG("create task, n_tokens = %d\n", (int) prompt_tokens.size()); SRV_DBG("create task, n_tokens = %d\n", (int) prompt_tokens.size());
server_task task; server_task task;
task.id = queue_tasks.get_new_id(); task.id = queue_tasks.get_new_id();
task.inf_type = inf_type; task.inf_type = inf_type;
@ -2205,18 +2219,7 @@ struct server_context {
int n_processing_slots = 0; int n_processing_slots = 0;
for (server_slot & slot : slots) { for (server_slot & slot : slots) {
json slot_data = slot.params.to_json(); json slot_data = slot.to_json();
slot_data["id"] = slot.id;
slot_data["id_task"] = slot.id_task;
slot_data["is_processing"] = slot.is_processing();
slot_data["prompt"] = common_detokenize(ctx, slot.prompt_tokens);
slot_data["next_token"] = {
{"has_next_token", slot.has_next_token},
{"has_new_line", slot.has_new_line},
{"n_remain", slot.n_remaining},
{"n_decoded", slot.n_decoded},
{"stopping_word", slot.stopping_word},
};
if (slot.is_processing()) { if (slot.is_processing()) {
n_processing_slots++; n_processing_slots++;
@ -2230,6 +2233,7 @@ struct server_context {
auto res = std::make_unique<server_task_result_metrics>(); auto res = std::make_unique<server_task_result_metrics>();
res->id = task.id; res->id = task.id;
res->slots_data = std::move(slots_data);
res->n_idle_slots = n_idle_slots; res->n_idle_slots = n_idle_slots;
res->n_processing_slots = n_processing_slots; res->n_processing_slots = n_processing_slots;
res->n_tasks_deferred = queue_tasks.queue_tasks_deferred.size(); res->n_tasks_deferred = queue_tasks.queue_tasks_deferred.size();
@ -3003,11 +3007,11 @@ int main(int argc, char ** argv) {
res.status = 200; res.status = 200;
}; };
svr->set_exception_handler([&res_error](const httplib::Request &, httplib::Response & res, std::exception_ptr ep) { svr->set_exception_handler([&res_error](const httplib::Request &, httplib::Response & res, const std::exception_ptr & ep) {
std::string message; std::string message;
try { try {
std::rethrow_exception(ep); std::rethrow_exception(ep);
} catch (std::exception & e) { } catch (const std::exception & e) {
message = e.what(); message = e.what();
} catch (...) { } catch (...) {
message = "Unknown Exception"; message = "Unknown Exception";

4
examples/server/utils.hpp
View File

@ -327,12 +327,12 @@ static std::string llama_get_chat_template(const struct llama_model * model) {
std::string template_key = "tokenizer.chat_template"; std::string template_key = "tokenizer.chat_template";
// call with NULL buffer to get the total size of the string // call with NULL buffer to get the total size of the string
int32_t res = llama_model_meta_val_str(model, template_key.c_str(), NULL, 0); int32_t res = llama_model_meta_val_str(model, template_key.c_str(), NULL, 0);
if (res < 0) { if (res < 2) {
return ""; return "";
} else { } else {
std::vector<char> model_template(res, 0); std::vector<char> model_template(res, 0);
llama_model_meta_val_str(model, template_key.c_str(), model_template.data(), model_template.size()); llama_model_meta_val_str(model, template_key.c_str(), model_template.data(), model_template.size());
return std::string(model_template.data(), model_template.size()); return std::string(model_template.data(), model_template.size() - 1);
} }
} }