tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-07-17 16:19:46 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'master' into gg/llama-kv-cache

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov
2025-02-20 14:26:43 +02:00
parent f95b04a21c 4806498bf1
commit 072280ea6b
34 changed files with 2103 additions and 1218 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
examples/main/main.cpp
View File

@ -4,7 +4,7 @@
#include "log.h"
#include "sampling.h"
#include "llama.h"
#include "chat-template.hpp"
#include "chat.h"
#include <cstdio>
#include <cstring>
@ -158,7 +158,7 @@ int main(int argc, char ** argv) {
}
const llama_vocab * vocab = llama_model_get_vocab(model);
auto chat_templates = common_chat_templates_from_model(model, params.chat_template);
auto chat_templates = common_chat_templates_init(model, params.chat_template);
LOG_INF("%s: llama threadpool init, n_threads = %d\n", __func__, (int) params.cpuparams.n_threads);
@ -201,7 +201,7 @@ int main(int argc, char ** argv) {
}
// auto enable conversation mode if chat template is available
const bool has_chat_template = chat_templates.has_explicit_template && chat_templates.template_default;
const bool has_chat_template = common_chat_templates_was_explicit(chat_templates.get());
if (params.conversation_mode == COMMON_CONVERSATION_MODE_AUTO) {
if (has_chat_template) {
LOG_INF("%s: chat template is available, enabling conversation mode (disable it with -no-cnv)\n", __func__);
@ -219,7 +219,7 @@ int main(int argc, char ** argv) {
// print chat template example in conversation mode
if (params.conversation_mode) {
if (params.enable_chat_template) {
LOG_INF("%s: chat template example:\n%s\n", __func__, common_chat_format_example(*chat_templates.template_default, params.use_jinja).c_str());
LOG_INF("%s: chat template example:\n%s\n", __func__, common_chat_format_example(chat_templates.get(), params.use_jinja).c_str());
} else {
LOG_INF("%s: in-suffix/prefix is specified, chat template will be disabled\n", __func__);
}
@ -264,9 +264,11 @@ int main(int argc, char ** argv) {
std::vector<llama_token> embd_inp;
auto chat_add_and_format = [&chat_msgs, &chat_templates](const std::string & role, const std::string & content) {
common_chat_msg new_msg{role, content, {}};
auto formatted = common_chat_format_single(*chat_templates.template_default, chat_msgs, new_msg, role == "user", g_params->use_jinja);
chat_msgs.push_back({role, content, {}});
common_chat_msg new_msg;
new_msg.role = role;
new_msg.content = content;
auto formatted = common_chat_format_single(chat_templates.get(), chat_msgs, new_msg, role == "user", g_params->use_jinja);
chat_msgs.push_back(new_msg);
LOG_DBG("formatted: '%s'\n", formatted.c_str());
return formatted;
};
@ -755,11 +757,14 @@ int main(int argc, char ** argv) {
// check for reverse prompt using special tokens
llama_token last_token = common_sampler_last(smpl);
if (std::find(antiprompt_token.begin(), antiprompt_token.end(), last_token) != antiprompt_token.end()) {
if (params.interactive) {
is_interacting = true;
for (auto token : antiprompt_token) {
if (token == last_token) {
if (params.interactive) {
is_interacting = true;
}
is_antiprompt = true;
break;
}
is_antiprompt = true;
}
if (is_antiprompt) {

139
examples/run/run.cpp
View File

@ -24,7 +24,7 @@
#include <string>
#include <vector>
#include "chat-template.hpp"
#include "chat.h"
#include "common.h"
#include "json.hpp"
#include "linenoise.cpp/linenoise.h"
@ -113,6 +113,7 @@ class Opt {
llama_context_params ctx_params;
llama_model_params model_params;
std::string model_;
std::string chat_template_file;
std::string user;
bool use_jinja = false;
int context_size = -1, ngl = -1;
@ -148,6 +149,16 @@ class Opt {
return 0;
}
int handle_option_with_value(int argc, const char ** argv, int & i, std::string & option_value) {
if (i + 1 >= argc) {
return 1;
}
option_value = argv[++i];
return 0;
}
int parse(int argc, const char ** argv) {
bool options_parsing = true;
for (int i = 1, positional_args_i = 0; i < argc; ++i) {
@ -169,6 +180,11 @@ class Opt {
verbose = true;
} else if (options_parsing && strcmp(argv[i], "--jinja") == 0) {
use_jinja = true;
} else if (options_parsing && strcmp(argv[i], "--chat-template-file") == 0){
if (handle_option_with_value(argc, argv, i, chat_template_file) == 1) {
return 1;
}
use_jinja = true;
} else if (options_parsing && parse_flag(argv, i, "-h", "--help")) {
help = true;
return 0;
@ -207,6 +223,11 @@ class Opt {
"Options:\n"
" -c, --context-size <value>\n"
" Context size (default: %d)\n"
" --chat-template-file <path>\n"
" Path to the file containing the chat template to use with the model.\n"
" Only supports jinja templates and implicitly sets the --jinja flag.\n"
" --jinja\n"
" Use jinja templating for the chat template of the model\n"
" -n, -ngl, --ngl <value>\n"
" Number of GPU layers (default: %d)\n"
" --temp <value>\n"
@ -261,13 +282,12 @@ static int get_terminal_width() {
#endif
}
#ifdef LLAMA_USE_CURL
class File {
public:
FILE * file = nullptr;
FILE * open(const std::string & filename, const char * mode) {
file = fopen(filename.c_str(), mode);
file = ggml_fopen(filename.c_str(), mode);
return file;
}
@ -303,6 +323,28 @@ class File {
return 0;
}
std::string read_all(const std::string & filename){
open(filename, "r");
lock();
if (!file) {
printe("Error opening file '%s': %s", filename.c_str(), strerror(errno));
return "";
}
fseek(file, 0, SEEK_END);
size_t size = ftell(file);
fseek(file, 0, SEEK_SET);
std::string out;
out.resize(size);
size_t read_size = fread(&out[0], 1, size, file);
if (read_size != size) {
printe("Error reading file '%s': %s", filename.c_str(), strerror(errno));
return "";
}
return out;
}
~File() {
if (fd >= 0) {
# ifdef _WIN32
@ -327,6 +369,7 @@ class File {
# endif
};
#ifdef LLAMA_USE_CURL
class HttpClient {
public:
int init(const std::string & url, const std::vector<std::string> & headers, const std::string & output_file,
@ -557,7 +600,7 @@ class LlamaData {
llama_model_ptr model;
llama_sampler_ptr sampler;
llama_context_ptr context;
std::vector<llama_chat_message> messages;
std::vector<llama_chat_message> messages; // TODO: switch to common_chat_msg
std::list<std::string> msg_strs;
std::vector<char> fmtted;
@ -834,44 +877,23 @@ static void add_message(const char * role, const std::string & text, LlamaData &
}
// Function to apply the chat template and resize `formatted` if needed
static int apply_chat_template(const common_chat_template & tmpl, LlamaData & llama_data, const bool append, bool use_jinja) {
if (use_jinja) {
json messages = json::array();
for (const auto & msg : llama_data.messages) {
messages.push_back({
{"role", msg.role},
{"content", msg.content},
});
}
try {
minja::chat_template_inputs tmpl_inputs;
tmpl_inputs.messages = messages;
tmpl_inputs.add_generation_prompt = append;
minja::chat_template_options tmpl_opts;
tmpl_opts.use_bos_token = false;
tmpl_opts.use_eos_token = false;
auto result = tmpl.apply(tmpl_inputs, tmpl_opts);
llama_data.fmtted.resize(result.size() + 1);
memcpy(llama_data.fmtted.data(), result.c_str(), result.size() + 1);
return result.size();
} catch (const std::exception & e) {
printe("failed to render the chat template: %s\n", e.what());
return -1;
}
}
int result = llama_chat_apply_template(
tmpl.source().c_str(), llama_data.messages.data(), llama_data.messages.size(), append,
append ? llama_data.fmtted.data() : nullptr, append ? llama_data.fmtted.size() : 0);
if (append && result > static_cast<int>(llama_data.fmtted.size())) {
llama_data.fmtted.resize(result);
result = llama_chat_apply_template(tmpl.source().c_str(), llama_data.messages.data(),
llama_data.messages.size(), append, llama_data.fmtted.data(),
llama_data.fmtted.size());
static int apply_chat_template(const struct common_chat_templates * tmpls, LlamaData & llama_data, const bool append, bool use_jinja) {
common_chat_templates_inputs inputs;
for (const auto & msg : llama_data.messages) {
common_chat_msg cmsg;
cmsg.role = msg.role;
cmsg.content = msg.content;
inputs.messages.push_back(cmsg);
}
inputs.add_generation_prompt = append;
inputs.use_jinja = use_jinja;
return result;
auto chat_params = common_chat_templates_apply(tmpls, inputs);
// TODO: use other params for tool calls.
auto result = chat_params.prompt;
llama_data.fmtted.resize(result.size() + 1);
memcpy(llama_data.fmtted.data(), result.c_str(), result.size() + 1);
return result.size();
}
// Function to tokenize the prompt
@ -1015,8 +1037,8 @@ static int generate_response(LlamaData & llama_data, const std::string & prompt,
}
// Helper function to apply the chat template and handle errors
static int apply_chat_template_with_error_handling(const common_chat_template & tmpl, LlamaData & llama_data, const bool append, int & output_length, bool use_jinja) {
const int new_len = apply_chat_template(tmpl, llama_data, append, use_jinja);
static int apply_chat_template_with_error_handling(const common_chat_templates * tmpls, LlamaData & llama_data, const bool append, int & output_length, bool use_jinja) {
const int new_len = apply_chat_template(tmpls, llama_data, append, use_jinja);
if (new_len < 0) {
printe("failed to apply the chat template\n");
return -1;
@ -1074,12 +1096,33 @@ static int get_user_input(std::string & user_input, const std::string & user) {
return 0;
}
// Reads a chat template file to be used
static std::string read_chat_template_file(const std::string & chat_template_file) {
if(chat_template_file.empty()){
return "";
}
File file;
std::string chat_template = "";
chat_template = file.read_all(chat_template_file);
if(chat_template.empty()){
printe("Error opening chat template file '%s': %s", chat_template_file.c_str(), strerror(errno));
return "";
}
return chat_template;
}
// Main chat loop function
static int chat_loop(LlamaData & llama_data, const std::string & user, bool use_jinja) {
static int chat_loop(LlamaData & llama_data, const std::string & user, const std::string & chat_template_file, bool use_jinja) {
int prev_len = 0;
llama_data.fmtted.resize(llama_n_ctx(llama_data.context.get()));
auto chat_templates = common_chat_templates_from_model(llama_data.model.get(), "");
GGML_ASSERT(chat_templates.template_default);
std::string chat_template = "";
if(!chat_template_file.empty()){
chat_template = read_chat_template_file(chat_template_file);
}
auto chat_templates = common_chat_templates_init(llama_data.model.get(), chat_template.empty() ? nullptr : chat_template);
static const bool stdout_a_terminal = is_stdout_a_terminal();
while (true) {
// Get user input
@ -1090,7 +1133,7 @@ static int chat_loop(LlamaData & llama_data, const std::string & user, bool use_
add_message("user", user.empty() ? user_input : user, llama_data);
int new_len;
if (apply_chat_template_with_error_handling(*chat_templates.template_default, llama_data, true, new_len, use_jinja) < 0) {
if (apply_chat_template_with_error_handling(chat_templates.get(), llama_data, true, new_len, use_jinja) < 0) {
return 1;
}
@ -1105,7 +1148,7 @@ static int chat_loop(LlamaData & llama_data, const std::string & user, bool use_
}
add_message("assistant", response, llama_data);
if (apply_chat_template_with_error_handling(*chat_templates.template_default, llama_data, false, prev_len, use_jinja) < 0) {
if (apply_chat_template_with_error_handling(chat_templates.get(), llama_data, false, prev_len, use_jinja) < 0) {
return 1;
}
}
@ -1165,7 +1208,7 @@ int main(int argc, const char ** argv) {
return 1;
}
if (chat_loop(llama_data, opt.user, opt.use_jinja)) {
if (chat_loop(llama_data, opt.user, opt.chat_template_file, opt.use_jinja)) {
return 1;
}

BIN
examples/server/public/index.html.gz
View File

Binary file not shown.

80
examples/server/server.cpp
View File

@ -274,7 +274,7 @@ struct server_task {
params.speculative.p_min = json_value(data, "speculative.p_min", defaults.speculative.p_min);
params.speculative.n_min = std::min(params.speculative.n_max, params.speculative.n_min);
params.speculative.n_min = std::max(params.speculative.n_min, 2);
params.speculative.n_min = std::max(params.speculative.n_min, 0);
params.speculative.n_max = std::max(params.speculative.n_max, 0);
// Use OpenAI API logprobs only if n_probs wasn't provided
@ -329,9 +329,6 @@ struct server_task {
}
// process "json_schema" and "grammar"
if (data.contains("json_schema") && !data.at("json_schema").is_null() && data.contains("grammar") && !data.at("grammar").is_null()) {
throw std::runtime_error("Either \"json_schema\" or \"grammar\" can be specified, but not both");
}
if (data.contains("json_schema") && !data.contains("grammar")) {
try {
auto schema = json_value(data, "json_schema", json::object());
@ -1807,7 +1804,7 @@ struct server_context {
// Necessary similarity of prompt for slot selection
float slot_prompt_similarity = 0.0f;
common_chat_templates chat_templates;
common_chat_templates_ptr chat_templates;
~server_context() {
// Clear any sampling context
@ -1891,45 +1888,17 @@ struct server_context {
llama_init_dft.context.reset();
}
if (params_base.chat_template.empty() && !validate_builtin_chat_template(params.use_jinja)) {
chat_templates = common_chat_templates_init(model, params_base.chat_template);
try {
common_chat_format_example(chat_templates.get(), params.use_jinja);
} catch (const std::exception & e) {
SRV_WRN("%s: The chat template that comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses\n", __func__);
chat_templates = common_chat_templates_from_model(model, "chatml");
} else {
chat_templates = common_chat_templates_from_model(model, params_base.chat_template);
chat_templates = common_chat_templates_init(model, "chatml");
}
GGML_ASSERT(chat_templates.template_default.get() != nullptr);
return true;
}
bool validate_builtin_chat_template(bool use_jinja) const {
llama_chat_message chat[] = {{"user", "test"}};
if (use_jinja) {
auto templates = common_chat_templates_from_model(model, "");
common_chat_inputs inputs;
inputs.messages = json::array({{
{"role", "user"},
{"content", "test"},
}});
GGML_ASSERT(templates.template_default);
try {
common_chat_params_init(*templates.template_default, inputs);
if (templates.template_tool_use) {
common_chat_params_init(*templates.template_tool_use, inputs);
}
return true;
} catch (const std::exception & e) {
SRV_ERR("failed to apply template: %s\n", e.what());
return false;
}
} else {
const char * tmpl = llama_model_chat_template(model, /* name */ nullptr);
const int32_t chat_res = llama_chat_apply_template(tmpl, chat, 1, true, nullptr, 0);
return chat_res > 0;
}
}
void init() {
const int32_t n_ctx_slot = n_ctx / params_base.n_parallel;
@ -3822,13 +3791,15 @@ int main(int argc, char ** argv) {
{ "default_generation_settings", ctx_server.default_generation_settings_for_props },
{ "total_slots", ctx_server.params_base.n_parallel },
{ "model_path", ctx_server.params_base.model },
{ "chat_template", ctx_server.chat_templates.template_default->source() },
{ "bos_token", ctx_server.chat_templates.template_default->bos_token() },
{ "eos_token", ctx_server.chat_templates.template_default->eos_token() },
{ "chat_template", common_chat_templates_source(ctx_server.chat_templates.get()) },
{ "bos_token", common_token_to_piece(ctx_server.ctx, llama_vocab_bos(ctx_server.vocab), /* special= */ true)},
{ "eos_token", common_token_to_piece(ctx_server.ctx, llama_vocab_eos(ctx_server.vocab), /* special= */ true)},
{ "build_info", build_info },
};
if (ctx_server.params_base.use_jinja && ctx_server.chat_templates.template_tool_use) {
data["chat_template_tool_use"] = ctx_server.chat_templates.template_tool_use->source();
if (ctx_server.params_base.use_jinja) {
if (auto tool_use_src = common_chat_templates_source(ctx_server.chat_templates.get(), "tool_use")) {
data["chat_template_tool_use"] = tool_use_src;
}
}
res_ok(res, data);
@ -4063,7 +4034,7 @@ int main(int argc, char ** argv) {
}
auto body = json::parse(req.body);
json data = oaicompat_completion_params_parse(body, params.use_jinja, params.reasoning_format, ctx_server.chat_templates);
json data = oaicompat_completion_params_parse(body, params.use_jinja, params.reasoning_format, ctx_server.chat_templates.get());
return handle_completions_impl(
SERVER_TASK_TYPE_COMPLETION,
@ -4076,7 +4047,7 @@ int main(int argc, char ** argv) {
// same with handle_chat_completions, but without inference part
const auto handle_apply_template = [&ctx_server, &params, &res_ok](const httplib::Request & req, httplib::Response & res) {
auto body = json::parse(req.body);
json data = oaicompat_completion_params_parse(body, params.use_jinja, params.reasoning_format, ctx_server.chat_templates);
json data = oaicompat_completion_params_parse(body, params.use_jinja, params.reasoning_format, ctx_server.chat_templates.get());
res_ok(res, {{ "prompt", std::move(data.at("prompt")) }});
};
@ -4263,6 +4234,11 @@ int main(int argc, char ** argv) {
// return;
//}
// if true, use TEI API format, otherwise use Jina API format
// Jina: https://jina.ai/reranker/
// TEI: https://huggingface.github.io/text-embeddings-inference/#/Text%20Embeddings%20Inference/rerank
bool is_tei_format = body.contains("texts");
json query;
if (body.count("query") == 1) {
query = body.at("query");
@ -4275,7 +4251,8 @@ int main(int argc, char ** argv) {
return;
}
std::vector<std::string> documents = json_value(body, "documents", std::vector<std::string>());
std::vector<std::string> documents = json_value(body, "documents",
json_value(body, "texts", std::vector<std::string>()));
if (documents.empty()) {
res_error(res, format_error_response("\"documents\" must be a non-empty string array", ERROR_TYPE_INVALID_REQUEST));
return;
@ -4320,7 +4297,12 @@ int main(int argc, char ** argv) {
}
// write JSON response
json root = format_response_rerank(body, responses);
json root = format_response_rerank(
body,
responses,
is_tei_format,
documents);
res_ok(res, root);
};
@ -4482,8 +4464,8 @@ int main(int argc, char ** argv) {
// print sample chat example to make it clear which template is used
LOG_INF("%s: chat template, chat_template: %s, example_format: '%s'\n", __func__,
ctx_server.chat_templates.template_default->source().c_str(),
common_chat_format_example(*ctx_server.chat_templates.template_default, ctx_server.params_base.use_jinja).c_str());
common_chat_templates_source(ctx_server.chat_templates.get()),
common_chat_format_example(ctx_server.chat_templates.get(), ctx_server.params_base.use_jinja).c_str());
ctx_server.queue_tasks.on_new_task([&ctx_server](const server_task & task) {
ctx_server.process_single_task(task);

2
examples/server/tests/README.md
View File

@ -48,7 +48,7 @@ DEBUG=1 ./tests.sh -s -v -x
To run all the tests in a file:
```shell
./tests.sh unit/test_chat_completion.py.py -v -x
./tests.sh unit/test_chat_completion.py -v -x
```
To run a single test:

45
examples/server/tests/unit/test_chat_completion.py
View File

@ -21,6 +21,8 @@ def create_server():
(None, "Book", "What is the best book", 8, "^ blue", 23, 8, "length", True, "This is not a chat template, it is"),
("codellama70b", "You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 64, "length", False, None),
("codellama70b", "You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 64, "length", True, None),
(None, "Book", [{"type": "text", "text": "What is"}, {"type": "text", "text": "the best book"}], 8, "Whillicter", 79, 8, "length", False, None),
(None, "Book", [{"type": "text", "text": "What is"}, {"type": "text", "text": "the best book"}], 8, "Whillicter", 79, 8, "length", True, None),
]
)
def test_chat_completion(model, system_prompt, user_prompt, max_tokens, re_content, n_prompt, n_predicted, finish_reason, jinja, chat_template):
@ -44,7 +46,7 @@ def test_chat_completion(model, system_prompt, user_prompt, max_tokens, re_conte
assert res.body["usage"]["completion_tokens"] == n_predicted
choice = res.body["choices"][0]
assert "assistant" == choice["message"]["role"]
assert match_regex(re_content, choice["message"]["content"])
assert match_regex(re_content, choice["message"]["content"]), f'Expected {re_content}, got {choice["message"]["content"]}'
assert choice["finish_reason"] == finish_reason
@ -169,6 +171,47 @@ def test_completion_with_response_format(response_format: dict, n_predicted: int
assert "error" in res.body
@pytest.mark.parametrize("jinja,json_schema,n_predicted,re_content", [
(False, {"const": "42"}, 6, "\"42\""),
(True, {"const": "42"}, 6, "\"42\""),
])
def test_completion_with_json_schema(jinja: bool, json_schema: dict, n_predicted: int, re_content: str):
global server
server.jinja = jinja
server.start()
res = server.make_request("POST", "/chat/completions", data={
"max_tokens": n_predicted,
"messages": [
{"role": "system", "content": "You are a coding assistant."},
{"role": "user", "content": "Write an example"},
],
"json_schema": json_schema,
})
assert res.status_code == 200, f'Expected 200, got {res.status_code}'
choice = res.body["choices"][0]
assert match_regex(re_content, choice["message"]["content"]), f'Expected {re_content}, got {choice["message"]["content"]}'
@pytest.mark.parametrize("jinja,grammar,n_predicted,re_content", [
(False, 'root ::= "a"{5,5}', 6, "a{5,5}"),
(True, 'root ::= "a"{5,5}', 6, "a{5,5}"),
])
def test_completion_with_grammar(jinja: bool, grammar: str, n_predicted: int, re_content: str):
global server
server.jinja = jinja
server.start()
res = server.make_request("POST", "/chat/completions", data={
"max_tokens": n_predicted,
"messages": [
{"role": "user", "content": "Does not matter what I say, does it?"},
],
"grammar": grammar,
})
assert res.status_code == 200, res.body
choice = res.body["choices"][0]
assert match_regex(re_content, choice["message"]["content"]), choice["message"]["content"]
@pytest.mark.parametrize("messages", [
None,
"string",

38
examples/server/tests/unit/test_rerank.py
View File

@ -10,17 +10,20 @@ def create_server():
server = ServerPreset.jina_reranker_tiny()
TEST_DOCUMENTS = [
"A machine is a physical system that uses power to apply forces and control movement to perform an action. The term is commonly applied to artificial devices, such as those employing engines or motors, but also to natural biological macromolecules, such as molecular machines.",
"Learning is the process of acquiring new understanding, knowledge, behaviors, skills, values, attitudes, and preferences. The ability to learn is possessed by humans, non-human animals, and some machines; there is also evidence for some kind of learning in certain plants.",
"Machine learning is a field of study in artificial intelligence concerned with the development and study of statistical algorithms that can learn from data and generalize to unseen data, and thus perform tasks without explicit instructions.",
"Paris, capitale de la France, est une grande ville européenne et un centre mondial de l'art, de la mode, de la gastronomie et de la culture. Son paysage urbain du XIXe siècle est traversé par de larges boulevards et la Seine."
]
def test_rerank():
global server
server.start()
res = server.make_request("POST", "/rerank", data={
"query": "Machine learning is",
"documents": [
"A machine is a physical system that uses power to apply forces and control movement to perform an action. The term is commonly applied to artificial devices, such as those employing engines or motors, but also to natural biological macromolecules, such as molecular machines.",
"Learning is the process of acquiring new understanding, knowledge, behaviors, skills, values, attitudes, and preferences. The ability to learn is possessed by humans, non-human animals, and some machines; there is also evidence for some kind of learning in certain plants.",
"Machine learning is a field of study in artificial intelligence concerned with the development and study of statistical algorithms that can learn from data and generalize to unseen data, and thus perform tasks without explicit instructions.",
"Paris, capitale de la France, est une grande ville européenne et un centre mondial de l'art, de la mode, de la gastronomie et de la culture. Son paysage urbain du XIXe siècle est traversé par de larges boulevards et la Seine."
]
"documents": TEST_DOCUMENTS,
})
assert res.status_code == 200
assert len(res.body["results"]) == 4
@ -38,6 +41,29 @@ def test_rerank():
assert least_relevant["index"] == 3
def test_rerank_tei_format():
global server
server.start()
res = server.make_request("POST", "/rerank", data={
"query": "Machine learning is",
"texts": TEST_DOCUMENTS,
})
assert res.status_code == 200
assert len(res.body) == 4
most_relevant = res.body[0]
least_relevant = res.body[0]
for doc in res.body:
if doc["score"] > most_relevant["score"]:
most_relevant = doc
if doc["score"] < least_relevant["score"]:
least_relevant = doc
assert most_relevant["score"] > least_relevant["score"]
assert most_relevant["index"] == 2
assert least_relevant["index"] == 3
@pytest.mark.parametrize("documents", [
[],
None,

10
examples/server/tests/unit/test_tool_call.py
View File

@ -356,12 +356,12 @@ def test_weather(hf_repo: str, template_override: str | Tuple[str, str | None] |
(None, 128, "bartowski/functionary-small-v3.2-GGUF:Q8_0", ("meetkai/functionary-medium-v3.2", None)),
(None, 128, "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", None),
(None, 128, "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", None),
("^> 0.56$", 128, "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", "chatml"),
(None, 128, "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", "chatml"),
(None, 128, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
# TODO: fix these (wrong results, either didn't respect decimal instruction or got wrong value)
("^The y-coordinate [\\s\\S]*?\\*\\*0.5\\*\\*", 8192, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
("[\\s\\S]*?\\*\\*0\\.5\\*\\*", 8192, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
("[\\s\\S]*?\\*\\*\\s*0.5($|\\*\\*)", 8192, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
# ("[\\s\\S]*?\\*\\*\\s*0.5($|\\*\\*)", 8192, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
])
def test_calc_result(result_override: str | None, n_predict: int, hf_repo: str, template_override: str | Tuple[str, str | None] | None):
global server
@ -401,7 +401,7 @@ def test_calc_result(result_override: str | None, n_predict: int, hf_repo: str,
{
"role": "tool",
"name": "calculate",
"content": 0.55644242476,
"content": "0.55644242476",
"tool_call_id": "call_6789"
}
],
@ -444,7 +444,7 @@ def test_calc_result(result_override: str | None, n_predict: int, hf_repo: str,
(128, None, "^The sum of 102 and 7 is 109.*", None, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
(1024, 'deepseek', "To find the sum of.*", "I need to calculate the sum of 102 and 7.*", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
(1024, 'none', "<think>\n?I need[\\s\\S]*?</think>\n?To find.*", None, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
(1024, 'none', "^I need[\\s\\S]*?</think>\n?To find.*", None, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
(1024, 'deepseek', "To find the sum of.*", "First, I [\\s\\S]*", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
])

188
examples/server/utils.hpp
View File

@ -12,9 +12,7 @@
// Change JSON_ASSERT from assert() to GGML_ASSERT:
#define JSON_ASSERT GGML_ASSERT
#include "json.hpp"
#include "minja.hpp"
#include "chat.hpp"
#include "chat-template.hpp"
#include "chat.h"
#include <random>
#include <sstream>
@ -347,41 +345,6 @@ static llama_tokens format_infill(
return embd_inp;
}
// Format given chat. If tmpl is empty, we take the template from model metadata
inline std::string format_chat(const common_chat_template & tmpl, const std::vector<json> & messages) {
std::vector<common_chat_msg> chat;
for (size_t i = 0; i < messages.size(); ++i) {
const auto & curr_msg = messages[i];
std::string role = json_value(curr_msg, "role", std::string(""));
std::string content;
if (curr_msg.contains("content")) {
if (curr_msg["content"].is_string()) {
content = curr_msg["content"].get<std::string>();
} else if (curr_msg["content"].is_array()) {
for (const auto & part : curr_msg["content"]) {
if (part.contains("text")) {
content += "\n" + part["text"].get<std::string>();
}
}
} else {
throw std::runtime_error("Invalid 'content' type (ref: https://github.com/ggml-org/llama.cpp/issues/8367)");
}
} else {
throw std::runtime_error("Missing 'content' (ref: https://github.com/ggml-org/llama.cpp/issues/8367)");
}
chat.push_back({role, content, /* tool_calls= */ {}});
}
const auto formatted_chat = common_chat_apply_template(tmpl, chat, true, /* use_jinja= */ false);
LOG_DBG("formatted_chat: '%s'\n", formatted_chat.c_str());
return formatted_chat;
}
//
// base64 utils (TODO: move to common in the future)
//
@ -579,12 +542,9 @@ static json oaicompat_completion_params_parse(
const json & body, /* openai api json semantics */
bool use_jinja,
common_reasoning_format reasoning_format,
const common_chat_templates & chat_templates)
const struct common_chat_templates * tmpls)
{
json llama_params;
const auto & tmpl = body.contains("tools") && chat_templates.template_tool_use
? *chat_templates.template_tool_use
: *chat_templates.template_default;
auto tools = json_value(body, "tools", json());
auto stream = json_value(body, "stream", false);
@ -610,62 +570,58 @@ static json oaicompat_completion_params_parse(
llama_params["stop"] = json_value(body, "stop", json::array());
}
auto json_schema = json_value(body, "json_schema", json());
auto grammar = json_value(body, "grammar", std::string());
if (!json_schema.is_null() && !grammar.empty()) {
throw std::runtime_error("Cannot use both json_schema and grammar");
}
// Handle "response_format" field
if (body.contains("response_format")) {
json response_format = json_value(body, "response_format", json::object());
std::string response_type = json_value(response_format, "type", std::string());
if (response_type == "json_object") {
llama_params["json_schema"] = json_value(response_format, "schema", json::object());
json_schema = json_value(response_format, "schema", json::object());
} else if (response_type == "json_schema") {
json json_schema = json_value(response_format, "json_schema", json::object());
llama_params["json_schema"] = json_value(json_schema, "schema", json::object());
json_schema = json_value(json_schema, "schema", json::object());
} else if (!response_type.empty() && response_type != "text") {
throw std::runtime_error("response_format type must be one of \"text\" or \"json_object\", but got: " + response_type);
}
}
// Apply chat template to the list of messages
if (use_jinja) {
auto tool_choice = json_value(body, "tool_choice", std::string("auto"));
if (tool_choice != "none" && tool_choice != "auto" && tool_choice != "required") {
throw std::runtime_error("Invalid tool_choice: " + tool_choice);
}
if (tool_choice != "none" && llama_params.contains("grammar")) {
throw std::runtime_error("Cannot use custom grammar constraints with tools.");
}
common_chat_inputs inputs;
inputs.extract_reasoning = reasoning_format != COMMON_REASONING_FORMAT_NONE;
inputs.messages = body.at("messages");
inputs.tools = tools;
inputs.tool_choice = tool_choice;
inputs.parallel_tool_calls = json_value(body, "parallel_tool_calls", false);
if (inputs.parallel_tool_calls && !tmpl.original_caps().supports_parallel_tool_calls) {
LOG_DBG("Disabling parallel_tool_calls because the template does not support it\n");
inputs.parallel_tool_calls = false;
}
inputs.stream = stream;
// TODO: support mixing schema w/ tools beyond generic format.
inputs.json_schema = json_value(llama_params, "json_schema", json());
auto chat_params = common_chat_params_init(tmpl, inputs);
common_chat_templates_inputs inputs;
inputs.messages = common_chat_msgs_parse_oaicompat(body.at("messages"));
inputs.tools = common_chat_tools_parse_oaicompat(tools);
inputs.tool_choice = common_chat_tool_choice_parse_oaicompat(json_value(body, "tool_choice", std::string("auto")));
inputs.json_schema = json_schema.is_null() ? "" : json_schema.dump();
inputs.grammar = grammar;
inputs.add_generation_prompt = true;
inputs.use_jinja = use_jinja;
inputs.parallel_tool_calls = json_value(body, "parallel_tool_calls", false);
inputs.extract_reasoning = reasoning_format != COMMON_REASONING_FORMAT_NONE;
if (!inputs.tools.empty() && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE && body.contains("grammar")) {
throw std::runtime_error("Cannot use custom grammar constraints with tools.");
}
llama_params["chat_format"] = static_cast<int>(chat_params.format);
llama_params["prompt"] = chat_params.prompt;
llama_params["grammar"] = chat_params.grammar;
llama_params["grammar_lazy"] = chat_params.grammar_lazy;
auto grammar_triggers = json::array();
for (const auto & trigger : chat_params.grammar_triggers) {
grammar_triggers.push_back({
{"word", trigger.word},
{"at_start", trigger.at_start},
});
}
llama_params["grammar_triggers"] = grammar_triggers;
llama_params["preserved_tokens"] = chat_params.preserved_tokens;
for (const auto & stop : chat_params.additional_stops) {
llama_params["stop"].push_back(stop);
}
} else {
llama_params["prompt"] = format_chat(tmpl, body.at("messages"));
// Apply chat template to the list of messages
auto chat_params = common_chat_templates_apply(tmpls, inputs);
llama_params["chat_format"] = static_cast<int>(chat_params.format);
llama_params["prompt"] = chat_params.prompt;
llama_params["grammar"] = chat_params.grammar;
llama_params["grammar_lazy"] = chat_params.grammar_lazy;
auto grammar_triggers = json::array();
for (const auto & trigger : chat_params.grammar_triggers) {
grammar_triggers.push_back({
{"word", trigger.word},
{"at_start", trigger.at_start},
});
}
llama_params["grammar_triggers"] = grammar_triggers;
llama_params["preserved_tokens"] = chat_params.preserved_tokens;
for (const auto & stop : chat_params.additional_stops) {
llama_params["stop"].push_back(stop);
}
// Handle "n" field
@ -737,29 +693,51 @@ static json format_embeddings_response_oaicompat(const json & request, const jso
return res;
}
static json format_response_rerank(const json & request, const json & ranks) {
json data = json::array();
int32_t n_tokens = 0;
int i = 0;
for (const auto & rank : ranks) {
data.push_back(json{
{"index", i++},
{"relevance_score", json_value(rank, "score", 0.0)},
});
static json format_response_rerank(
const json & request,
const json & ranks,
bool is_tei_format,
std::vector<std::string> & texts) {
json res;
if (is_tei_format) {
// TEI response format
res = json::array();
bool return_text = json_value(request, "return_text", false);
for (const auto & rank : ranks) {
int index = json_value(rank, "index", 0);
json elem = json{
{"index", index},
{"score", json_value(rank, "score", 0.0)},
};
if (return_text) {
elem["text"] = std::move(texts[index]);
}
res.push_back(elem);
}
} else {
// Jina response format
json results = json::array();
int32_t n_tokens = 0;
for (const auto & rank : ranks) {
results.push_back(json{
{"index", json_value(rank, "index", 0)},
{"relevance_score", json_value(rank, "score", 0.0)},
});
n_tokens += json_value(rank, "tokens_evaluated", 0);
n_tokens += json_value(rank, "tokens_evaluated", 0);
}
res = json{
{"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
{"object", "list"},
{"usage", json{
{"prompt_tokens", n_tokens},
{"total_tokens", n_tokens}
}},
{"results", results}
};
}
json res = json {
{"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
{"object", "list"},
{"usage", json {
{"prompt_tokens", n_tokens},
{"total_tokens", n_tokens}
}},
{"results", data}
};
return res;
}

29
examples/server/webui/src/components/ChatMessage.tsx
View File

@ -159,6 +159,35 @@ export default function ChatMessage({
</div>
</details>
)}
{msg.extra && msg.extra.length > 0 && (
<details
className={classNames({
'collapse collapse-arrow mb-4 bg-base-200': true,
'bg-opacity-10': msg.role !== 'assistant',
})}
>
<summary className="collapse-title">
Extra content
</summary>
<div className="collapse-content">
{msg.extra.map(
(extra, i) =>
extra.type === 'textFile' ? (
<div key={extra.name}>
<b>{extra.name}</b>
<pre>{extra.content}</pre>
</div>
) : extra.type === 'context' ? (
<div key={i}>
<pre>{extra.content}</pre>
</div>
) : null // TODO: support other extra types
)}
</div>
</details>
)}
<MarkdownDisplay
content={content}
isGenerating={isPending}

26
examples/server/webui/src/components/ChatScreen.tsx
View File

@ -1,10 +1,11 @@
import { useEffect, useMemo, useState } from 'react';
import { useEffect, useMemo, useRef, useState } from 'react';
import { CallbackGeneratedChunk, useAppContext } from '../utils/app.context';
import ChatMessage from './ChatMessage';
import { CanvasType, Message, PendingMessage } from '../utils/types';
import { classNames, throttle } from '../utils/misc';
import CanvasPyInterpreter from './CanvasPyInterpreter';
import StorageUtils from '../utils/storage';
import { useVSCodeContext } from '../utils/llama-vscode';
/**
* A message display is a message node with additional information for rendering.
@ -81,6 +82,14 @@ export default function ChatScreen() {
replaceMessageAndGenerate,
} = useAppContext();
const [inputMsg, setInputMsg] = useState('');
const inputRef = useRef<HTMLTextAreaElement>(null);
const { extraContext, clearExtraContext } = useVSCodeContext(
inputRef,
setInputMsg
);
// TODO: improve this when we have "upload file" feature
const currExtra: Message['extra'] = extraContext ? [extraContext] : undefined;
// keep track of leaf node for rendering
const [currNodeId, setCurrNodeId] = useState<number>(-1);
@ -115,10 +124,20 @@ export default function ChatScreen() {
setCurrNodeId(-1);
// get the last message node
const lastMsgNodeId = messages.at(-1)?.msg.id ?? null;
if (!(await sendMessage(currConvId, lastMsgNodeId, inputMsg, onChunk))) {
if (
!(await sendMessage(
currConvId,
lastMsgNodeId,
inputMsg,
currExtra,
onChunk
))
) {
// restore the input message if failed
setInputMsg(lastInpMsg);
}
// OK
clearExtraContext();
};
const handleEditMessage = async (msg: Message, content: string) => {
@ -129,6 +148,7 @@ export default function ChatScreen() {
viewingChat.conv.id,
msg.parent,
content,
msg.extra,
onChunk
);
setCurrNodeId(-1);
@ -143,6 +163,7 @@ export default function ChatScreen() {
viewingChat.conv.id,
msg.parent,
null,
msg.extra,
onChunk
);
setCurrNodeId(-1);
@ -203,6 +224,7 @@ export default function ChatScreen() {
<textarea
className="textarea textarea-bordered w-full"
placeholder="Type a message (Shift+Enter to add a new line)"
ref={inputRef}
value={inputMsg}
onChange={(e) => setInputMsg(e.target.value)}
onKeyDown={(e) => {

6
examples/server/webui/src/utils/app.context.tsx
View File

@ -25,6 +25,7 @@ interface AppContextValue {
convId: string | null,
leafNodeId: Message['id'] | null,
content: string,
extra: Message['extra'],
onChunk: CallbackGeneratedChunk
) => Promise<boolean>;
stopGenerating: (convId: string) => void;
@ -32,6 +33,7 @@ interface AppContextValue {
convId: string,
parentNodeId: Message['id'], // the parent node of the message to be replaced
content: string | null,
extra: Message['extra'],
onChunk: CallbackGeneratedChunk
) => Promise<void>;
@ -274,6 +276,7 @@ export const AppContextProvider = ({
convId: string | null,
leafNodeId: Message['id'] | null,
content: string,
extra: Message['extra'],
onChunk: CallbackGeneratedChunk
): Promise<boolean> => {
if (isGenerating(convId ?? '') || content.trim().length === 0) return false;
@ -298,6 +301,7 @@ export const AppContextProvider = ({
convId,
role: 'user',
content,
extra,
parent: leafNodeId,
children: [],
},
@ -324,6 +328,7 @@ export const AppContextProvider = ({
convId: string,
parentNodeId: Message['id'], // the parent node of the message to be replaced
content: string | null,
extra: Message['extra'],
onChunk: CallbackGeneratedChunk
) => {
if (isGenerating(convId)) return;
@ -339,6 +344,7 @@ export const AppContextProvider = ({
convId,
role: 'user',
content,
extra,
parent: parentNodeId,
children: [],
},

62
examples/server/webui/src/utils/llama-vscode.ts Normal file
View File

@ -0,0 +1,62 @@
import { useEffect, useState } from 'react';
import { MessageExtraContext } from './types';
// Extra context when using llama.cpp WebUI from llama-vscode, inside an iframe
// Ref: https://github.com/ggml-org/llama.cpp/pull/11940
interface SetTextEvData {
text: string;
context: string;
}
/**
* To test it:
* window.postMessage({ command: 'setText', text: 'Spot the syntax error', context: 'def test()\n return 123' }, '*');
*/
export const useVSCodeContext = (
inputRef: React.RefObject<HTMLTextAreaElement>,
setInputMsg: (text: string) => void
) => {
const [extraContext, setExtraContext] = useState<MessageExtraContext | null>(
null
);
// Accept setText message from a parent window and set inputMsg and extraContext
useEffect(() => {
const handleMessage = (event: MessageEvent) => {
if (event.data?.command === 'setText') {
const data: SetTextEvData = event.data;
setInputMsg(data?.text);
if (data?.context && data.context.length > 0) {
setExtraContext({
type: 'context',
content: data.context,
});
}
inputRef.current?.focus();
}
};
window.addEventListener('message', handleMessage);
return () => window.removeEventListener('message', handleMessage);
}, []);
// Add a keydown listener that sends the "escapePressed" message to the parent window
useEffect(() => {
const handleKeyDown = (event: KeyboardEvent) => {
if (event.key === 'Escape') {
window.parent.postMessage({ command: 'escapePressed' }, '*');
}
};
window.addEventListener('keydown', handleKeyDown);
return () => window.removeEventListener('keydown', handleKeyDown);
}, []);
return {
extraContext,
// call once the user message is sent, to clear the extra context
clearExtraContext: () => setExtraContext(null),
};
};

13
examples/server/webui/src/utils/misc.ts
View File

@ -53,12 +53,23 @@ export const copyStr = (textToCopy: string) => {
/**
* filter out redundant fields upon sending to API
* also format extra into text
*/
export function normalizeMsgsForAPI(messages: Readonly<Message[]>) {
return messages.map((msg) => {
let newContent = '';
for (const extra of msg.extra ?? []) {
if (extra.type === 'context') {
newContent += `${extra.content}\n\n`;
}
}
newContent += msg.content;
return {
role: msg.role,
content: msg.content,
content: newContent,
};
}) as APIMessage[];
}

14
examples/server/webui/src/utils/types.ts
View File

@ -42,11 +42,25 @@ export interface Message {
role: 'user' | 'assistant' | 'system';
content: string;
timings?: TimingReport;
extra?: MessageExtra[];
// node based system for branching
parent: Message['id'];
children: Message['id'][];
}
type MessageExtra = MessageExtraTextFile | MessageExtraContext; // TODO: will add more in the future
export interface MessageExtraTextFile {
type: 'textFile';
name: string;
content: string;
}
export interface MessageExtraContext {
type: 'context';
content: string;
}
export type APIMessage = Pick<Message, 'role' | 'content'>;
export interface Conversation {