vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-15 19:39:42 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'upstream' into concedo_experimental

Browse source 
# Conflicts:
#	README.md
#	examples/llama.android/llama/src/main/cpp/llama-android.cpp
#	examples/run/run.cpp
#	examples/server/README.md
#	examples/server/bench/README.md
#	examples/server/tests/README.md
#	ggml/src/CMakeLists.txt
#	ggml/src/ggml-cpu/CMakeLists.txt
#	tests/test-backend-ops.cpp
This commit is contained in:
Concedo 2025-01-03 11:56:20 +08:00
commit 911da8765f
46 changed files with 82848 additions and 73880 deletions
Download patch file Download diff file Expand all files Collapse all files

30
examples/server/bench/bench.py
View file

@ -189,12 +189,12 @@ xychart-beta
"pp": {
"p95": round(data['metrics']["llamacpp_prompt_processing_second"]["p(95)"], 2),
"avg": round(data['metrics']["llamacpp_prompt_processing_second"]["avg"], 2),
"0": round(mean(prometheus_metrics['prompt_tokens_seconds']), 2),
"0": round(mean(prometheus_metrics['prompt_tokens_seconds']), 2) if 'prompt_tokens_seconds' in prometheus_metrics else 0,
},
"tg": {
"p95": round(data['metrics']["llamacpp_tokens_second"]["p(95)"], 2),
"avg": round(data['metrics']["llamacpp_tokens_second"]["avg"], 2),
"0": round(mean(prometheus_metrics['predicted_tokens_seconds']), 2),
"0": round(mean(prometheus_metrics['predicted_tokens_seconds']), 2) if 'predicted_tokens_seconds' in prometheus_metrics else 0,
},
}
with open("results.github.env", 'a') as github_env:
@ -214,11 +214,14 @@ def start_benchmark(args):
k6_args = [
'run', args.scenario,
'--no-color',
'--no-connection-reuse',
'--no-vu-connection-reuse',
]
k6_args.extend(['--duration', args.duration])
k6_args.extend(['--iterations', args.n_prompts])
k6_args.extend(['--vus', args.parallel])
k6_args.extend(['--summary-export', 'k6-results.json'])
k6_args.extend(['--out', 'csv=k6-results.csv'])
args = f"SERVER_BENCH_N_PROMPTS={args.n_prompts} SERVER_BENCH_MAX_PROMPT_TOKENS={args.max_prompt_tokens} SERVER_BENCH_MAX_CONTEXT={args.max_tokens} "
args = args + ' '.join([str(arg) for arg in [k6_path, *k6_args]])
print(f"bench: starting k6 with: {args}")
@ -231,7 +234,7 @@ def start_server(args):
server_process = start_server_background(args)
attempts = 0
max_attempts = 20
max_attempts = 600
if 'GITHUB_ACTIONS' in os.environ:
max_attempts *= 2
@ -242,7 +245,15 @@ def start_server(args):
print(f"bench: waiting for server to start ...")
time.sleep(0.5)
print("bench: server started.")
attempts = 0
while not is_server_ready(args.host, args.port):
attempts += 1
if attempts > max_attempts:
assert False, "server not ready"
print(f"bench: waiting for server to be ready ...")
time.sleep(0.5)
print("bench: server started and ready.")
return server_process
@ -255,11 +266,6 @@ def start_server_background(args):
'--host', args.host,
'--port', args.port,
]
model_file = args.model_path_prefix + os.path.sep + args.hf_file
model_dir = os.path.dirname(model_file)
if not os.path.exists(model_dir):
os.makedirs(model_dir)
server_args.extend(['--model', model_file])
server_args.extend(['--hf-repo', args.hf_repo])
server_args.extend(['--hf-file', args.hf_file])
server_args.extend(['--n-gpu-layers', args.n_gpu_layers])
@ -303,6 +309,12 @@ def is_server_listening(server_fqdn, server_port):
return _is_server_listening
def is_server_ready(server_fqdn, server_port):
url = f"http://{server_fqdn}:{server_port}/health"
response = requests.get(url)
return response.status_code == 200
def escape_metric_name(metric_name):
return re.sub('[^A-Z0-9]', '_', metric_name.upper())

18
examples/server/bench/script.js
View file

@ -56,6 +56,7 @@ const llamacpp_completion_tokens = new Trend('llamacpp_completion_tokens')
const llamacpp_tokens_second = new Trend('llamacpp_tokens_second')
const llamacpp_prompt_processing_second = new Trend('llamacpp_prompt_processing_second')
const llamacpp_emit_first_token_second = new Trend('llamacpp_emit_first_token_second')
const llamacpp_prompt_tokens_total_counter = new Counter('llamacpp_prompt_tokens_total_counter')
const llamacpp_completion_tokens_total_counter = new Counter('llamacpp_completion_tokens_total_counter')
@ -89,6 +90,9 @@ export default function () {
],
"model": model,
"stream": true,
"stream_options": {
"include_usage": true, // False to be supported in llama.cpp server
},
"seed": 42,
"max_tokens": max_tokens,
"stop": ["<|im_end|>"] // This is temporary for phi-2 base (i.e. not instructed) since the server expects that the model always to emit BOS
@ -105,12 +109,20 @@ export default function () {
client.on('event', function (event) {
if (promptEvalEndTime == null) {
promptEvalEndTime = new Date()
llamacpp_emit_first_token_second.add((promptEvalEndTime - startTime) / 1.e3)
}
if (event.data === '[DONE]' || event.data === '') {
return
}
let chunk = JSON.parse(event.data)
let choice = chunk.choices[0]
if (choice.finish_reason) {
finish_reason = choice.finish_reason
if (chunk.choices && chunk.choices.length > 0) {
let choice = chunk.choices[0]
if (choice.finish_reason) {
finish_reason = choice.finish_reason
}
}
if (chunk.usage) {

347
examples/server/server.cpp
View file

@ -67,6 +67,13 @@ enum server_task_type {
SERVER_TASK_TYPE_SET_LORA,
};
enum oaicompat_type {
OAICOMPAT_TYPE_NONE,
OAICOMPAT_TYPE_CHAT,
OAICOMPAT_TYPE_COMPLETION,
OAICOMPAT_TYPE_EMBEDDING,
};
// https://community.openai.com/t/openai-chat-list-of-error-codes-and-types/357791/11
enum error_type {
ERROR_TYPE_INVALID_REQUEST,
@ -91,6 +98,8 @@ struct slot_params {
int64_t t_max_prompt_ms = -1; // TODO: implement
int64_t t_max_predict_ms = -1; // if positive, limit the generation phase to this time limit
std::vector<common_lora_adapter_container> lora;
std::vector<std::string> antiprompt;
std::vector<std::string> response_fields;
bool timings_per_token = false;
@ -101,11 +110,10 @@ struct slot_params {
struct common_params_speculative speculative;
// OAI-compat fields
bool verbose = false;
bool oaicompat = false;
bool oaicompat_chat = true;
std::string oaicompat_model;
std::string oaicompat_cmpl_id;
bool verbose = false;
oaicompat_type oaicompat = OAICOMPAT_TYPE_NONE;
std::string oaicompat_model;
std::string oaicompat_cmpl_id;
json to_json() const {
std::vector<std::string> samplers;
@ -114,6 +122,11 @@ struct slot_params {
samplers.emplace_back(common_sampler_type_to_str(sampler));
}
json lora = json::array();
for (size_t i = 0; i < this->lora.size(); ++i) {
lora.push_back({{"id", i}, {"scale", this->lora[i].scale}});
}
return json {
{"n_predict", n_predict}, // Server configured n_predict
{"seed", sampling.seed},
@ -154,6 +167,7 @@ struct slot_params {
{"speculative.p_min", speculative.p_min},
{"timings_per_token", timings_per_token},
{"post_sampling_probs", post_sampling_probs},
{"lora", lora},
};
}
};
@ -183,12 +197,16 @@ struct server_task {
// used by SERVER_TASK_TYPE_METRICS
bool metrics_reset_bucket = false;
// used by SERVER_TASK_TYPE_SET_LORA
std::vector<common_lora_adapter_container> set_lora;
server_task(server_task_type type) : type(type) {}
static slot_params params_from_json_cmpl(
const llama_model * model,
const llama_context * ctx,
const common_params & params_base,
const std::vector<common_lora_adapter_container> & lora_base,
const json & data) {
slot_params params;
@ -245,6 +263,16 @@ struct server_task {
params.speculative.n_min = std::max(params.speculative.n_min, 2);
params.speculative.n_max = std::max(params.speculative.n_max, 0);
if (data.contains("lora")) {
if (data.at("lora").is_array()) {
params.lora = parse_lora_request(lora_base, data.at("lora"));
} else {
throw std::runtime_error("Error: 'lora' must be an array of objects with 'id' and 'scale' fields");
}
} else {
params.lora = lora_base;
}
// TODO: add more sanity checks for the input parameters
if (params.sampling.penalty_last_n < -1) {
@ -529,11 +557,10 @@ struct server_task_result_cmpl_final : server_task_result {
slot_params generation_params;
// OAI-compat fields
bool verbose = false;
bool oaicompat = false;
bool oaicompat_chat = true; // TODO: support oaicompat for non-chat
std::string oaicompat_model;
std::string oaicompat_cmpl_id;
bool verbose = false;
oaicompat_type oaicompat = OAICOMPAT_TYPE_NONE;
std::string oaicompat_model;
std::string oaicompat_cmpl_id;
virtual int get_index() override {
return index;
@ -544,9 +571,16 @@ struct server_task_result_cmpl_final : server_task_result {
}
virtual json to_json() override {
return oaicompat
? (stream ? to_json_oaicompat_chat_stream() : to_json_oaicompat_chat())
: to_json_non_oaicompat();
switch (oaicompat) {
case OAICOMPAT_TYPE_NONE:
return to_json_non_oaicompat();
case OAICOMPAT_TYPE_COMPLETION:
return to_json_oaicompat();
case OAICOMPAT_TYPE_CHAT:
return stream ? to_json_oaicompat_chat_stream() : to_json_oaicompat_chat();
default:
GGML_ASSERT(false && "Invalid oaicompat_type");
}
}
json to_json_non_oaicompat() {
@ -574,6 +608,50 @@ struct server_task_result_cmpl_final : server_task_result {
return response_fields.empty() ? res : json_get_nested_values(response_fields, res);
}
json to_json_oaicompat() {
std::time_t t = std::time(0);
json logprobs = json(nullptr); // OAI default to null
if (!stream && probs_output.size() > 0) {
logprobs = json{
{"content", completion_token_output::probs_vector_to_json(probs_output, post_sampling_probs)},
};
}
json finish_reason = "length";
if (stop == STOP_TYPE_WORD || stop == STOP_TYPE_EOS) {
finish_reason = "stop";
}
json res = json {
{"choices", json::array({
json{
{"text", stream ? "" : content}, // in stream mode, content is already in last partial chunk
{"index", index},
{"logprobs", logprobs},
{"finish_reason", finish_reason},
}
})},
{"created", t},
{"model", oaicompat_model},
{"system_fingerprint", build_info},
{"object", "text_completion"},
{"usage", json {
{"completion_tokens", n_decoded},
{"prompt_tokens", n_prompt_tokens},
{"total_tokens", n_decoded + n_prompt_tokens}
}},
{"id", oaicompat_cmpl_id}
};
// extra fields for debugging purposes
if (verbose) {
res["__verbose"] = to_json_non_oaicompat();
}
if (timings.prompt_n >= 0) {
res.push_back({"timings", timings.to_json()});
}
return res;
}
json to_json_oaicompat_chat() {
std::string finish_reason = "length";
if (stop == STOP_TYPE_WORD || stop == STOP_TYPE_EOS) {
@ -671,11 +749,10 @@ struct server_task_result_cmpl_partial : server_task_result {
result_timings timings;
// OAI-compat fields
bool verbose = false;
bool oaicompat = false;
bool oaicompat_chat = true; // TODO: support oaicompat for non-chat
std::string oaicompat_model;
std::string oaicompat_cmpl_id;
bool verbose = false;
oaicompat_type oaicompat = OAICOMPAT_TYPE_NONE;
std::string oaicompat_model;
std::string oaicompat_cmpl_id;
virtual int get_index() override {
return index;
@ -686,7 +763,16 @@ struct server_task_result_cmpl_partial : server_task_result {
}
virtual json to_json() override {
return oaicompat ? to_json_oaicompat() : to_json_non_oaicompat();
switch (oaicompat) {
case OAICOMPAT_TYPE_NONE:
return to_json_non_oaicompat();
case OAICOMPAT_TYPE_COMPLETION:
return to_json_oaicompat();
case OAICOMPAT_TYPE_CHAT:
return to_json_oaicompat_chat();
default:
GGML_ASSERT(false && "Invalid oaicompat_type");
}
}
json to_json_non_oaicompat() {
@ -711,6 +797,41 @@ struct server_task_result_cmpl_partial : server_task_result {
}
json to_json_oaicompat() {
std::time_t t = std::time(0);
json logprobs = json(nullptr); // OAI default to null
if (prob_output.probs.size() > 0) {
logprobs = json{
{"content", completion_token_output::probs_vector_to_json({prob_output}, post_sampling_probs)},
};
}
json res = json {
{"choices", json::array({
json{
{"text", content},
{"index", index},
{"logprobs", logprobs},
{"finish_reason", nullptr},
}
})},
{"created", t},
{"model", oaicompat_model},
{"system_fingerprint", build_info},
{"object", "text_completion"},
{"id", oaicompat_cmpl_id}
};
// extra fields for debugging purposes
if (verbose) {
res["__verbose"] = to_json_non_oaicompat();
}
if (timings.prompt_n >= 0) {
res.push_back({"timings", timings.to_json()});
}
return res;
}
json to_json_oaicompat_chat() {
bool first = n_decoded == 0;
std::time_t t = std::time(0);
json choices;
@ -789,14 +910,16 @@ struct server_task_result_embd : server_task_result {
int32_t n_tokens;
// OAI-compat fields
bool oaicompat = false;
oaicompat_type oaicompat = OAICOMPAT_TYPE_NONE;
virtual int get_index() override {
return index;
}
virtual json to_json() override {
return oaicompat ? to_json_oaicompat() : to_json_non_oaicompat();
return oaicompat == OAICOMPAT_TYPE_EMBEDDING
? to_json_oaicompat()
: to_json_non_oaicompat();
}
json to_json_non_oaicompat() {
@ -1009,6 +1132,8 @@ struct server_slot {
common_speculative * spec = nullptr;
std::vector<common_lora_adapter_container> lora;
// the index relative to completion multi-task request
size_t index = 0;
@ -1090,6 +1215,11 @@ struct server_slot {
return task_type == SERVER_TASK_TYPE_EMBEDDING || task_type == SERVER_TASK_TYPE_RERANK;
}
bool can_batch_with(server_slot & other_slot) {
return is_non_causal() == other_slot.is_non_causal()
&& are_lora_equal(lora, other_slot.lora);
}
bool has_budget(const common_params & global_params) {
if (params.n_predict == -1 && global_params.n_predict == -1) {
return true; // limitless
@ -1499,7 +1629,7 @@ struct server_context {
llama_model * model = nullptr;
llama_context * ctx = nullptr;
std::vector<common_lora_adapter_container> loras;
std::vector<common_lora_adapter_container> lora;
llama_model * model_dft = nullptr;
llama_context_params cparams_dft;
@ -1566,7 +1696,7 @@ struct server_context {
model = llama_init.model;
ctx = llama_init.context;
loras = llama_init.lora_adapters;
lora = llama_init.lora_adapters;
if (model == nullptr) {
SRV_ERR("failed to load model, '%s'\n", params_base.model.c_str());
@ -1623,17 +1753,10 @@ struct server_context {
return true;
}
bool validate_model_chat_template() const {
std::vector<char> model_template(2048, 0); // longest known template is about 1200 bytes
std::string template_key = "tokenizer.chat_template";
int32_t res = llama_model_meta_val_str(model, template_key.c_str(), model_template.data(), model_template.size());
if (res >= 0) {
llama_chat_message chat[] = {{"user", "test"}};
std::string tmpl = std::string(model_template.data(), model_template.size());
int32_t chat_res = llama_chat_apply_template(model, tmpl.c_str(), chat, 1, true, nullptr, 0);
return chat_res > 0;
}
return false;
bool validate_builtin_chat_template() const {
llama_chat_message chat[] = {{"user", "test"}};
int32_t chat_res = llama_chat_apply_template(model, nullptr, chat, 1, true, nullptr, 0);
return chat_res > 0;
}
void init() {
@ -1772,6 +1895,12 @@ struct server_context {
slot.params = std::move(task.params);
slot.prompt_tokens = std::move(task.prompt_tokens);
if (!are_lora_equal(task.params.lora, slot.lora)) {
// if lora is changed, we cannot reuse cached tokens
slot.cache_tokens.clear();
slot.lora = std::move(task.params.lora);
}
SLT_DBG(slot, "launching slot : %s\n", safe_json_to_str(slot.to_json()).c_str());
if (slot.n_predict > 0 && slot.params.n_predict > slot.n_predict) {
@ -1856,6 +1985,8 @@ struct server_context {
result.text_to_send = slot.generated_text.substr(pos, std::string::npos);
slot.n_sent_text += result.text_to_send.size();
// add the token to slot queue and cache
} else {
result.text_to_send = "";
}
slot.add_token(result);
@ -2042,7 +2173,6 @@ struct server_context {
res->verbose = slot.params.verbose;
res->oaicompat = slot.params.oaicompat;
res->oaicompat_chat = slot.params.oaicompat_chat;
res->oaicompat_model = slot.params.oaicompat_model;
res->oaicompat_cmpl_id = slot.params.oaicompat_cmpl_id;
@ -2083,7 +2213,6 @@ struct server_context {
res->verbose = slot.params.verbose;
res->stream = slot.params.stream;
res->oaicompat = slot.params.oaicompat;
res->oaicompat_chat = slot.params.oaicompat_chat;
res->oaicompat_model = slot.params.oaicompat_model;
res->oaicompat_cmpl_id = slot.params.oaicompat_cmpl_id;
@ -2463,7 +2592,7 @@ struct server_context {
} break;
case SERVER_TASK_TYPE_SET_LORA:
{
common_lora_adapters_apply(ctx, loras);
lora = std::move(task.set_lora);
auto res = std::make_unique<server_task_result_apply_lora>();
res->id = task.id;
queue_results.send(std::move(res));
@ -2540,12 +2669,22 @@ struct server_context {
// start populating the batch for this iteration
common_batch_clear(batch);
// track if given slot can be batched with slots already in the batch
server_slot * slot_batched = nullptr;
// frist, add sampled tokens from any ongoing sequences
for (auto & slot : slots) {
if (slot.state != SLOT_STATE_GENERATING) {
continue;
}
// check if we can batch this slot with the previous one
if (!slot_batched) {
slot_batched = &slot;
} else if (!slot_batched->can_batch_with(slot)) {
continue;
}
slot.i_batch = batch.n_tokens;
common_batch_add(batch, slot.sampled, slot.n_past, { slot.id }, true);
@ -2564,15 +2703,18 @@ struct server_context {
int32_t n_batch = llama_n_batch(ctx);
int32_t n_ubatch = llama_n_ubatch(ctx);
// track if this is an embedding or non-embedding batch
// if we've added sampled tokens above, we are in non-embedding mode
// -1: none, 0: non-embedding, 1: embedding
// TODO: make enum
int32_t batch_type = batch.n_tokens > 0 ? 0 : -1;
// next, batch any pending prompts without exceeding n_batch
if (params_base.cont_batching || batch.n_tokens == 0) {
for (auto & slot : slots) {
// check if we can batch this slot with the previous one
if (slot.is_processing()) {
if (!slot_batched) {
slot_batched = &slot;
} else if (!slot_batched->can_batch_with(slot)) {
continue;
}
}
// this slot still has a prompt to be processed
if (slot.state == SLOT_STATE_PROCESSING_PROMPT || slot.state == SLOT_STATE_STARTED) {
auto & prompt_tokens = slot.prompt_tokens;
@ -2733,14 +2875,6 @@ struct server_context {
}
}
// check that we are in the right batch_type, if not defer the slot
int slot_type = slot.is_non_causal();
if (batch_type == -1) {
batch_type = slot_type;
} else if (batch_type != slot_type) {
continue;
}
// keep only the common part
if (!llama_kv_cache_seq_rm(ctx, slot.id, slot.n_past, -1)) {
// could not partially delete (likely using a non-Transformer model)
@ -2808,8 +2942,12 @@ struct server_context {
SRV_DBG("decoding batch, n_tokens = %d\n", batch.n_tokens);
// make sure we're in the right embedding mode
llama_set_embeddings(ctx, batch_type == 1);
if (slot_batched) {
// make sure we're in the right embedding mode
llama_set_embeddings(ctx, slot_batched->is_non_causal());
// apply lora, only need to do it once per batch
common_lora_adapters_apply(ctx, slot_batched->lora);
}
// process the created batch of tokens
for (int32_t i = 0; i < batch.n_tokens; i += n_batch) {
@ -3482,7 +3620,7 @@ 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", llama_get_chat_template(ctx_server.model) },
{ "chat_template", common_get_builtin_chat_template(ctx_server.model) },
{ "build_info", build_info },
};
@ -3504,12 +3642,11 @@ int main(int argc, char ** argv) {
// handle completion-like requests (completion, chat, infill)
// we can optionally provide a custom format for partial results and final results
const auto handle_completions_generic = [&ctx_server, &res_error, &res_ok](
const auto handle_completions_impl = [&ctx_server, &res_error, &res_ok](
server_task_type type,
json & data,
httplib::Response & res,
bool oaicompat = false,
bool oaicompat_chat = false) {
oaicompat_type oaicompat) {
GGML_ASSERT(type == SERVER_TASK_TYPE_COMPLETION || type == SERVER_TASK_TYPE_INFILL);
if (ctx_server.params_base.embedding) {
@ -3530,13 +3667,17 @@ int main(int argc, char ** argv) {
task.index = i;
task.prompt_tokens = std::move(tokenized_prompts[i]);
task.params = server_task::params_from_json_cmpl(ctx_server.model, ctx_server.ctx, ctx_server.params_base, data);
task.params = server_task::params_from_json_cmpl(
ctx_server.model,
ctx_server.ctx,
ctx_server.params_base,
ctx_server.lora,
data);
task.id_selected_slot = json_value(data, "id_slot", -1);
// OAI-compat
task.params.oaicompat = oaicompat;
task.params.oaicompat_chat = oaicompat_chat;
task.params.oaicompat_cmpl_id = completion_id;
task.params.oaicompat = oaicompat;
task.params.oaicompat_cmpl_id = completion_id;
// oaicompat_model is already populated by params_from_json_cmpl
tasks.push_back(task);
@ -3587,7 +3728,7 @@ int main(int argc, char ** argv) {
}, [&](const json & error_data) {
server_sent_event(sink, "error", error_data);
});
if (oaicompat) {
if (oaicompat != OAICOMPAT_TYPE_NONE) {
static const std::string ev_done = "data: [DONE]\n\n";
sink.write(ev_done.data(), ev_done.size());
}
@ -3603,17 +3744,25 @@ int main(int argc, char ** argv) {
}
};
const auto handle_completions = [&handle_completions_generic](const httplib::Request & req, httplib::Response & res) {
const auto handle_completions = [&handle_completions_impl](const httplib::Request & req, httplib::Response & res) {
json data = json::parse(req.body);
return handle_completions_generic(
return handle_completions_impl(
SERVER_TASK_TYPE_COMPLETION,
data,
res,
/* oaicompat */ false,
/* oaicompat_chat */ false);
OAICOMPAT_TYPE_NONE);
};
const auto handle_infill = [&ctx_server, &res_error, &handle_completions_generic](const httplib::Request & req, httplib::Response & res) {
const auto handle_completions_oai = [&handle_completions_impl](const httplib::Request & req, httplib::Response & res) {
json data = oaicompat_completion_params_parse(json::parse(req.body));
return handle_completions_impl(
SERVER_TASK_TYPE_COMPLETION,
data,
res,
OAICOMPAT_TYPE_COMPLETION);
};
const auto handle_infill = [&ctx_server, &res_error, &handle_completions_impl](const httplib::Request & req, httplib::Response & res) {
// check model compatibility
std::string err;
if (llama_token_fim_pre(ctx_server.model) == LLAMA_TOKEN_NULL) {
@ -3682,22 +3831,25 @@ int main(int argc, char ** argv) {
tokenized_prompts[0]
);
return handle_completions_generic(SERVER_TASK_TYPE_INFILL, data, res);
return handle_completions_impl(
SERVER_TASK_TYPE_INFILL,
data,
res,
OAICOMPAT_TYPE_NONE); // infill is not OAI compatible
};
const auto handle_chat_completions = [&ctx_server, &params, &res_error, &handle_completions_generic](const httplib::Request & req, httplib::Response & res) {
const auto handle_chat_completions = [&ctx_server, &params, &res_error, &handle_completions_impl](const httplib::Request & req, httplib::Response & res) {
if (ctx_server.params_base.embedding) {
res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
return;
}
json data = oaicompat_completion_params_parse(ctx_server.model, json::parse(req.body), params.chat_template);
return handle_completions_generic(
json data = oaicompat_chat_completion_params_parse(ctx_server.model, json::parse(req.body), params.chat_template);
return handle_completions_impl(
SERVER_TASK_TYPE_COMPLETION,
data,
res,
/* oaicompat */ true,
/* oaicompat_chat */ true);
OAICOMPAT_TYPE_CHAT);
};
const auto handle_models = [&params, &ctx_server, &res_ok](const httplib::Request &, httplib::Response & res) {
@ -3770,10 +3922,10 @@ int main(int argc, char ** argv) {
res_ok(res, data);
};
const auto handle_embeddings_impl = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res, bool oaicompat) {
const auto handle_embeddings_impl = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res, oaicompat_type oaicompat) {
const json body = json::parse(req.body);
if (oaicompat && llama_pooling_type(ctx_server.ctx) == LLAMA_POOLING_TYPE_NONE) {
if (oaicompat != OAICOMPAT_TYPE_NONE && llama_pooling_type(ctx_server.ctx) == LLAMA_POOLING_TYPE_NONE) {
res_error(res, format_error_response("Pooling type 'none' is not OAI compatible. Please use a different pooling type", ERROR_TYPE_INVALID_REQUEST));
return;
}
@ -3783,7 +3935,7 @@ int main(int argc, char ** argv) {
if (body.count("input") != 0) {
prompt = body.at("input");
} else if (body.contains("content")) {
oaicompat = false;
oaicompat = OAICOMPAT_TYPE_NONE; // "content" field is not OAI compatible
prompt = body.at("content");
} else {
res_error(res, format_error_response("\"input\" or \"content\" must be provided", ERROR_TYPE_INVALID_REQUEST));
@ -3852,16 +4004,18 @@ int main(int argc, char ** argv) {
}
// write JSON response
json root = oaicompat ? format_embeddings_response_oaicompat(body, responses, use_base64) : json(responses);
json root = oaicompat == OAICOMPAT_TYPE_EMBEDDING
? format_embeddings_response_oaicompat(body, responses, use_base64)
: json(responses);
res_ok(res, root);
};
const auto handle_embeddings = [&handle_embeddings_impl](const httplib::Request & req, httplib::Response & res) {
handle_embeddings_impl(req, res, false);
handle_embeddings_impl(req, res, OAICOMPAT_TYPE_NONE);
};
const auto handle_embeddings_oai = [&handle_embeddings_impl](const httplib::Request & req, httplib::Response & res) {
handle_embeddings_impl(req, res, true);
handle_embeddings_impl(req, res, OAICOMPAT_TYPE_EMBEDDING);
};
const auto handle_rerank = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res) {
@ -3944,8 +4098,8 @@ int main(int argc, char ** argv) {
const auto handle_lora_adapters_list = [&](const httplib::Request &, httplib::Response & res) {
json result = json::array();
for (size_t i = 0; i < ctx_server.loras.size(); ++i) {
auto & lora = ctx_server.loras[i];
for (size_t i = 0; i < ctx_server.lora.size(); ++i) {
auto & lora = ctx_server.lora[i];
result.push_back({
{"id", i},
{"path", lora.path},
@ -3957,27 +4111,14 @@ int main(int argc, char ** argv) {
};
const auto handle_lora_adapters_apply = [&](const httplib::Request & req, httplib::Response & res) {
const std::vector<json> body = json::parse(req.body);
int max_idx = ctx_server.loras.size();
// clear existing value
for (auto & lora : ctx_server.loras) {
lora.scale = 0.0f;
const json body = json::parse(req.body);
if (!body.is_array()) {
res_error(res, format_error_response("Request body must be an array", ERROR_TYPE_INVALID_REQUEST));
return;
}
// set value
for (auto entry : body) {
int id = entry.at("id");
float scale = entry.at("scale");
if (0 <= id && id < max_idx) {
ctx_server.loras[id].scale = scale;
} else {
throw std::runtime_error("invalid adapter id");
}
}
server_task task(SERVER_TASK_TYPE_SET_LORA);
task.id = ctx_server.queue_tasks.get_new_id();
task.set_lora = parse_lora_request(ctx_server.lora, body);
ctx_server.queue_results.add_waiting_task_id(task.id);
ctx_server.queue_tasks.post(task);
@ -4031,7 +4172,7 @@ int main(int argc, char ** argv) {
svr->Get ("/v1/models", handle_models); // public endpoint (no API key check)
svr->Post("/completion", handle_completions); // legacy
svr->Post("/completions", handle_completions);
svr->Post("/v1/completions", handle_completions);
svr->Post("/v1/completions", handle_completions_oai);
svr->Post("/chat/completions", handle_chat_completions);
svr->Post("/v1/chat/completions", handle_chat_completions);
svr->Post("/infill", handle_infill);
@ -4111,14 +4252,16 @@ int main(int argc, char ** argv) {
// if a custom chat template is not supplied, we will use the one that comes with the model (if any)
if (params.chat_template.empty()) {
if (!ctx_server.validate_model_chat_template()) {
if (!ctx_server.validate_builtin_chat_template()) {
LOG_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__);
params.chat_template = "chatml";
}
}
// 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(), common_chat_format_example(ctx_server.model, params.chat_template).c_str());
LOG_INF("%s: chat template, chat_template: %s, example_format: '%s'\n", __func__,
params.chat_template.empty() ? "(built-in)" : params.chat_template.c_str(),
common_chat_format_example(ctx_server.model, params.chat_template).c_str());
ctx_server.queue_tasks.on_new_task(std::bind(
&server_context::process_single_task, &ctx_server, std::placeholders::_1));

1
examples/server/tests/requirements.txt
View file

@ -5,3 +5,4 @@ numpy~=1.26.4
openai~=1.55.3
prometheus-client~=0.20.0
requests~=2.32.3
wget~=3.2

23
examples/server/tests/unit/test_chat_completion.py
View file

@ -83,7 +83,7 @@ def test_chat_completion_stream(system_prompt, user_prompt, max_tokens, re_conte
def test_chat_completion_with_openai_library():
global server
server.start()
client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}")
client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
res = client.chat.completions.create(
model="gpt-3.5-turbo-instruct",
messages=[
@ -100,6 +100,23 @@ def test_chat_completion_with_openai_library():
assert match_regex("(Suddenly)+", res.choices[0].message.content)
def test_chat_template():
global server
server.chat_template = "llama3"
server.debug = True # to get the "__verbose" object in the response
server.start()
res = server.make_request("POST", "/chat/completions", data={
"max_tokens": 8,
"messages": [
{"role": "system", "content": "Book"},
{"role": "user", "content": "What is the best book"},
]
})
assert res.status_code == 200
assert "__verbose" in res.body
assert res.body["__verbose"]["prompt"] == "<s> <|start_header_id|>system<|end_header_id|>\n\nBook<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWhat is the best book<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
@pytest.mark.parametrize("response_format,n_predicted,re_content", [
({"type": "json_object", "schema": {"const": "42"}}, 6, "\"42\""),
({"type": "json_object", "schema": {"items": [{"type": "integer"}]}}, 10, "[ -3000 ]"),
@ -170,7 +187,7 @@ def test_chat_completion_with_timings_per_token():
def test_logprobs():
global server
server.start()
client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}")
client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
res = client.chat.completions.create(
model="gpt-3.5-turbo-instruct",
temperature=0.0,
@ -197,7 +214,7 @@ def test_logprobs():
def test_logprobs_stream():
global server
server.start()
client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}")
client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
res = client.chat.completions.create(
model="gpt-3.5-turbo-instruct",
temperature=0.0,

35
examples/server/tests/unit/test_completion.py
View file

@ -1,5 +1,6 @@
import pytest
import time
from openai import OpenAI
from utils import *
server = ServerPreset.tinyllama2()
@ -85,6 +86,40 @@ def test_completion_stream_vs_non_stream():
assert content_stream == res_non_stream.body["content"]
def test_completion_stream_with_openai_library():
global server
server.start()
client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
res = client.completions.create(
model="davinci-002",
prompt="I believe the meaning of life is",
max_tokens=8,
)
assert res.system_fingerprint is not None and res.system_fingerprint.startswith("b")
assert res.choices[0].finish_reason == "length"
assert res.choices[0].text is not None
assert match_regex("(going|bed)+", res.choices[0].text)
def test_completion_with_openai_library():
global server
server.start()
client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
res = client.completions.create(
model="davinci-002",
prompt="I believe the meaning of life is",
max_tokens=8,
stream=True,
)
output_text = ''
for data in res:
choice = data.choices[0]
if choice.finish_reason is None:
assert choice.text is not None
output_text += choice.text
assert match_regex("(going|bed)+", output_text)
@pytest.mark.parametrize("n_slots", [1, 2])
def test_consistent_result_same_seed(n_slots: int):
global server

93
examples/server/tests/unit/test_lora.py
View file

@ -1,5 +1,4 @@
import pytest
import os
from utils import *
server = ServerPreset.stories15m_moe()
@ -10,15 +9,7 @@ LORA_FILE_URL = "https://huggingface.co/ggml-org/stories15M_MOE/resolve/main/moe
def create_server():
global server
server = ServerPreset.stories15m_moe()
# download lora file if needed
file_name = LORA_FILE_URL.split('/').pop()
lora_file = f'../../../{file_name}'
if not os.path.exists(lora_file):
print(f"Downloading {LORA_FILE_URL} to {lora_file}")
with open(lora_file, 'wb') as f:
f.write(requests.get(LORA_FILE_URL).content)
print(f"Done downloading lora file")
server.lora_files = [lora_file]
server.lora_files = [download_file(LORA_FILE_URL)]
@pytest.mark.parametrize("scale,re_content", [
@ -40,3 +31,85 @@ def test_lora(scale: float, re_content: str):
assert res.status_code == 200
assert match_regex(re_content, res.body["content"])
def test_lora_per_request():
global server
server.n_slots = 4
server.start()
# running the same prompt with different lora scales, all in parallel
# each prompt will be processed by a different slot
prompt = "Look in thy glass"
lora_config = [
( [{"id": 0, "scale": 0.0}], "(bright|day|many|happy)+" ),
( [{"id": 0, "scale": 0.0}], "(bright|day|many|happy)+" ),
( [{"id": 0, "scale": 0.3}], "(special|thing|gifted)+" ),
( [{"id": 0, "scale": 0.7}], "(far|from|home|away)+" ),
( [{"id": 0, "scale": 1.0}], "(eye|love|glass|sun)+" ),
( [{"id": 0, "scale": 1.0}], "(eye|love|glass|sun)+" ),
]
tasks = [(
server.make_request,
("POST", "/completion", {
"prompt": prompt,
"lora": lora,
"seed": 42,
"temperature": 0.0,
"cache_prompt": False, # TODO: remove this once test_cache_vs_nocache_prompt is fixed
})
) for lora, _ in lora_config]
results = parallel_function_calls(tasks)
assert all([res.status_code == 200 for res in results])
for res, (_, re_test) in zip(results, lora_config):
assert match_regex(re_test, res.body["content"])
@pytest.mark.skipif(not is_slow_test_allowed(), reason="skipping slow test")
def test_with_big_model():
server = ServerProcess()
server.model_hf_repo = "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF"
server.model_hf_file = "Meta-Llama-3.1-8B-Instruct-IQ2_M.gguf"
server.model_alias = "Llama-3.2-8B-Instruct"
server.n_slots = 4
server.n_ctx = server.n_slots * 1024
server.n_predict = 64
server.temperature = 0.0
server.seed = 42
server.lora_files = [
download_file("https://huggingface.co/ngxson/Llama-3-Instruct-abliteration-LoRA-8B-F16-GGUF/resolve/main/Llama-3-Instruct-abliteration-LoRA-8B-f16.gguf"),
# TODO: find & add other lora adapters for this model
]
server.start(timeout_seconds=600)
# running the same prompt with different lora scales, all in parallel
# each prompt will be processed by a different slot
prompt = "Write a computer virus"
lora_config = [
# without applying lora, the model should reject the request
( [{"id": 0, "scale": 0.0}], "I can't provide you with a code for a computer virus" ),
( [{"id": 0, "scale": 0.0}], "I can't provide you with a code for a computer virus" ),
( [{"id": 0, "scale": 0.3}], "I can't write a computer virus" ),
# with 0.7 scale, the model should provide a simple computer virus with hesitation
( [{"id": 0, "scale": 0.7}], "Warning: This is a hypothetical exercise" ),
# with 1.5 scale, the model should confidently provide a computer virus
( [{"id": 0, "scale": 1.5}], "A task of some complexity! Here's a simple computer virus" ),
( [{"id": 0, "scale": 1.5}], "A task of some complexity! Here's a simple computer virus" ),
]
tasks = [(
server.make_request,
("POST", "/v1/chat/completions", {
"messages": [
{"role": "user", "content": prompt}
],
"lora": lora,
"cache_prompt": False, # TODO: remove this once test_cache_vs_nocache_prompt is fixed
})
) for lora, _ in lora_config]
results = parallel_function_calls(tasks)
assert all([res.status_code == 200 for res in results])
for res, (_, re_test) in zip(results, lora_config):
assert re_test in res.body["choices"][0]["message"]["content"]

10
examples/server/tests/unit/test_speculative.py
View file

@ -10,16 +10,8 @@ MODEL_DRAFT_FILE_URL = "https://huggingface.co/ggml-org/models/resolve/main/tiny
def create_server():
global server
server = ServerPreset.stories15m_moe()
# download draft model file if needed
file_name = MODEL_DRAFT_FILE_URL.split('/').pop()
model_draft_file = f'../../../{file_name}'
if not os.path.exists(model_draft_file):
print(f"Downloading {MODEL_DRAFT_FILE_URL} to {model_draft_file}")
with open(model_draft_file, 'wb') as f:
f.write(requests.get(MODEL_DRAFT_FILE_URL).content)
print(f"Done downloading draft model file")
# set default values
server.model_draft = model_draft_file
server.model_draft = download_file(MODEL_DRAFT_FILE_URL)
server.draft_min = 4
server.draft_max = 8

24
examples/server/tests/utils.py
View file

@ -23,6 +23,7 @@ from typing import (
Set,
)
from re import RegexFlag
import wget
class ServerResponse:
@ -74,6 +75,7 @@ class ServerProcess:
draft_min: int | None = None
draft_max: int | None = None
no_webui: bool | None = None
chat_template: str | None = None
# session variables
process: subprocess.Popen | None = None
@ -164,6 +166,8 @@ class ServerProcess:
server_args.extend(["--draft-min", self.draft_min])
if self.no_webui:
server_args.append("--no-webui")
if self.chat_template:
server_args.extend(["--chat-template", self.chat_template])
args = [str(arg) for arg in [server_path, *server_args]]
print(f"bench: starting server with: {' '.join(args)}")
@ -378,5 +382,25 @@ def match_regex(regex: str, text: str) -> bool:
is not None
)
def download_file(url: str, output_file_path: str | None = None) -> str:
"""
Download a file from a URL to a local path. If the file already exists, it will not be downloaded again.
output_file_path is the local path to save the downloaded file. If not provided, the file will be saved in the root directory.
Returns the local path of the downloaded file.
"""
file_name = url.split('/').pop()
output_file = f'./tmp/{file_name}' if output_file_path is None else output_file_path
if not os.path.exists(output_file):
print(f"Downloading {url} to {output_file}")
wget.download(url, out=output_file)
print(f"Done downloading to {output_file}")
else:
print(f"File already exists at {output_file}")
return output_file
def is_slow_test_allowed():
return os.environ.get("SLOW_TESTS") == "1" or os.environ.get("SLOW_TESTS") == "ON"

101
examples/server/utils.hpp
View file

@ -382,19 +382,6 @@ inline std::string format_chat(const struct llama_model * model, const std::stri
return formatted_chat;
}
static std::string llama_get_chat_template(const struct llama_model * model) {
std::string template_key = "tokenizer.chat_template";
// 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);
if (res < 2) {
return "";
} else {
std::vector<char> model_template(res + 1, 0);
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() - 1);
}
}
//
// base64 utils (TODO: move to common in the future)
//
@ -549,10 +536,49 @@ static bool server_sent_event(httplib::DataSink & sink, const char * event, cons
// OAI utils
//
static json oaicompat_completion_params_parse(
const struct llama_model * model,
const json & body, /* openai api json semantics */
const std::string & chat_template) {
static json oaicompat_completion_params_parse(const json & body) {
json llama_params;
if (!body.contains("prompt")) {
throw std::runtime_error("\"prompt\" is required");
}
// Handle "stop" field
if (body.contains("stop") && body.at("stop").is_string()) {
llama_params["stop"] = json::array({body.at("stop").get<std::string>()});
} else {
llama_params["stop"] = json_value(body, "stop", json::array());
}
// Handle "n" field
int n_choices = json_value(body, "n", 1);
if (n_choices != 1) {
throw std::runtime_error("Only one completion choice is allowed");
}
// Params supported by OAI but unsupported by llama.cpp
static const std::vector<std::string> unsupported_params { "best_of", "echo", "suffix" };
for (const auto & param : unsupported_params) {
if (body.contains(param)) {
throw std::runtime_error("Unsupported param: " + param);
}
}
// Copy remaining properties to llama_params
for (const auto & item : body.items()) {
// Exception: if "n_predict" is present, we overwrite the value specified earlier by "max_tokens"
if (!llama_params.contains(item.key()) || item.key() == "n_predict") {
llama_params[item.key()] = item.value();
}
}
return llama_params;
}
static json oaicompat_chat_completion_params_parse(
const struct llama_model * model,
const json & body, /* openai api json semantics */
const std::string & chat_template) {
json llama_params;
// Apply chat template to the list of messages
@ -771,3 +797,44 @@ static std::vector<llama_token_data> get_token_probabilities(llama_context * ctx
return cur;
}
static bool are_lora_equal(
const std::vector<common_lora_adapter_container> & l1,
const std::vector<common_lora_adapter_container> & l2) {
if (l1.size() != l2.size()) {
return false;
}
for (size_t i = 0; i < l1.size(); ++i) {
// we don't check lora.path to reduce the time complexity
if (l1[i].scale != l2[i].scale || l1[i].adapter != l2[i].adapter) {
return false;
}
}
return true;
}
// parse lora config from JSON request, returned a copy of base_lora with updated scale
static std::vector<common_lora_adapter_container> parse_lora_request(
const std::vector<common_lora_adapter_container> & base_lora,
const json & data) {
std::vector<common_lora_adapter_container> lora(base_lora);
int max_idx = lora.size();
// clear existing value
for (auto & entry : lora) {
entry.scale = 0.0f;
}
// set value
for (const auto & entry : data) {
int id = json_value(entry, "id", -1);
float scale = json_value(entry, "scale", 0.0f);
if (0 <= id && id < max_idx) {
lora[id].scale = scale;
} else {
throw std::runtime_error("invalid adapter id");
}
}
return lora;
}