Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-19 09:15:18 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge commit 'f7625019c51ca437a5840576d92362cfa710e4a2' into concedo_experimental

Browse Source 
# Conflicts:
#	.github/ISSUE_TEMPLATE/bug.md
#	.github/workflows/build.yml
#	CMakeLists.txt
#	Makefile
#	README.md
#	tests/test-backend-ops.cpp
#	tests/test-opt.cpp
#	tests/test-quantize-fns.cpp
This commit is contained in:
Concedo
2024-02-26 10:44:23 +08:00
parent a6ba735b07 f7625019c5
commit 3ccaf8e09a
37 changed files with 3514 additions and 698 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/server/README.md
+22 -1
View File
@@ -39,9 +39,12 @@ see https://github.com/ggerganov/llama.cpp/issues/1437
- `--mmproj MMPROJ_FILE`: Path to a multimodal projector file for LLaVA.
- `--grp-attn-n`: Set the group attention factor to extend context size through self-extend(default: 1=disabled), used together with group attention width `--grp-attn-w`
- `--grp-attn-w`: Set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n`
- `-n, --n-predict`: Set the maximum tokens to predict (default: -1)
- `-n N, --n-predict N`: Set the maximum tokens to predict (default: -1)
- `--slots-endpoint-disable`: To disable slots state monitoring endpoint. Slots state may contain user data, prompts included.
- `--metrics`: enable prometheus `/metrics` compatible endpoint (default: disabled)
- `--chat-template JINJA_TEMPLATE`: Set custom jinja chat template. This parameter accepts a string, not a file name (default: template taken from model's metadata). We only support [some pre-defined templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
- `--log-disable`: Output logs to stdout only, default: enabled.
- `--log-format FORMAT`: Define the log output to FORMAT: json or text (default: json)
## Build
@@ -98,6 +101,12 @@ curl --request POST \
--data '{"prompt": "Building a website can be done in 10 simple steps:","n_predict": 128}'
```
## Advanced testing
We implemented a [server test framework](./tests/README.md) using human-readable scenario.
*Before submitting an issue, please try to reproduce it with this format.*
## Node JS Test
You need to have [Node.js](https://nodejs.org/en) installed.
@@ -451,6 +460,18 @@ Notice that each `probs` is an array of length `n_probs`.
]
```
- **GET** `/metrics`: [Prometheus](https://prometheus.io/) compatible metrics exporter endpoint if `--metrics` is enabled:
Available metrics:
- `llamacpp:prompt_tokens_total`: Number of prompt tokens processed.
- `llamacpp:tokens_predicted_total`: Number of generation tokens processed.
- `llamacpp:prompt_tokens_seconds`: Average prompt throughput in tokens/s.
- `llamacpp:predicted_tokens_seconds`: Average generation throughput in tokens/s.
- `llamacpp:kv_cache_usage_ratio`: KV-cache usage. 1 means 100 percent usage.
- `llamacpp:kv_cache_tokens`: KV-cache tokens.
- `llamacpp:requests_processing`: Number of request processing.
- `llamacpp:requests_deferred`: Number of request deferred.
## More examples
### Change system prompt on runtime
examples/server/server.cpp
+366 -86
View File
@@ -44,9 +44,11 @@ struct server_params
int32_t read_timeout = 600;
int32_t write_timeout = 600;
bool slots_endpoint = true;
bool metrics_endpoint = false;
};
bool server_verbose = false;
bool server_log_json = true;
static size_t common_part(const std::vector<llama_token> &a, const std::vector<llama_token> &b)
{
@@ -302,12 +304,76 @@ struct llama_client_slot
}
void print_timings() const {
LOG_TEE("\n");
LOG_TEE("%s: prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",
__func__, t_prompt_processing, num_prompt_tokens_processed, t_prompt_processing / num_prompt_tokens_processed, 1e3 / t_prompt_processing * num_prompt_tokens_processed);
LOG_TEE("%s: eval time = %10.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)\n",
__func__, t_token_generation, n_decoded,t_token_generation / n_decoded, 1e3 / t_token_generation * n_decoded);
LOG_TEE("%s: total time = %10.2f ms\n", __func__, t_prompt_processing + t_token_generation);
char buffer[512];
double t_token = t_prompt_processing / num_prompt_tokens_processed;
double n_tokens_second = 1e3 / t_prompt_processing * num_prompt_tokens_processed;
sprintf(buffer, "prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)",
t_prompt_processing, num_prompt_tokens_processed,
t_token, n_tokens_second);
LOG_INFO(buffer, {
{"slot_id", id},
{"task_id", task_id},
{"t_prompt_processing", t_prompt_processing},
{"num_prompt_tokens_processed", num_prompt_tokens_processed},
{"t_token", t_token},
{"n_tokens_second", n_tokens_second},
});
t_token = t_token_generation / n_decoded;
n_tokens_second = 1e3 / t_token_generation * n_decoded;
sprintf(buffer, "generation eval time = %10.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)",
t_token_generation, n_decoded,
t_token, n_tokens_second);
LOG_INFO(buffer, {
{"slot_id", id},
{"task_id", task_id},
{"t_token_generation", t_token_generation},
{"n_decoded", n_decoded},
{"t_token", t_token},
{"n_tokens_second", n_tokens_second},
});
sprintf(buffer, " total time = %10.2f ms", t_prompt_processing + t_token_generation);
LOG_INFO(buffer, {
{"slot_id", id},
{"task_id", task_id},
{"t_prompt_processing", t_prompt_processing},
{"t_token_generation", t_token_generation},
{"t_total", t_prompt_processing + t_token_generation},
});
}
};
struct llama_metrics {
uint64_t n_prompt_tokens_processed_total = 0;
uint64_t n_tokens_predicted_total = 0;
uint64_t n_prompt_tokens_processed = 0;
uint64_t t_prompt_processing = 0;
uint64_t n_tokens_predicted = 0;
uint64_t t_tokens_generation = 0;
void on_prompt_eval(const llama_client_slot &slot) {
n_prompt_tokens_processed_total += slot.num_prompt_tokens_processed;
n_prompt_tokens_processed += slot.num_prompt_tokens_processed;
t_prompt_processing += slot.t_prompt_processing;
}
void on_prediction(const llama_client_slot &slot) {
n_tokens_predicted_total += slot.n_decoded;
n_tokens_predicted += slot.n_decoded;
t_tokens_generation += slot.t_token_generation;
}
void reset_bucket() {
n_prompt_tokens_processed = 0;
t_prompt_processing = 0;
n_tokens_predicted = 0;
t_tokens_generation = 0;
}
};
@@ -345,6 +411,8 @@ struct llama_server_context
llama_server_queue queue_tasks;
llama_server_response queue_results;
llama_metrics metrics;
~llama_server_context()
{
if (ctx)
@@ -364,7 +432,7 @@ struct llama_server_context
params = params_;
if (!params.mmproj.empty()) {
multimodal = true;
LOG_TEE("Multi Modal Mode Enabled");
LOG_INFO("Multi Modal Mode Enabled", {});
clp_ctx = clip_model_load(params.mmproj.c_str(), /*verbosity=*/ 1);
if(clp_ctx == nullptr) {
LOG_ERROR("unable to load clip model", {{"model", params.mmproj}});
@@ -417,7 +485,7 @@ struct llama_server_context
const int32_t n_ctx_slot = n_ctx / params.n_parallel;
LOG_TEE("Available slots:\n");
LOG_INFO("initializing slots", {{"n_slots", params.n_parallel}});
for (int i = 0; i < params.n_parallel; i++)
{
llama_client_slot slot;
@@ -426,7 +494,10 @@ struct llama_server_context
slot.n_ctx = n_ctx_slot;
slot.n_predict = params.n_predict;
LOG_TEE(" -> Slot %i - max context: %i\n", slot.id, n_ctx_slot);
LOG_INFO("new slot", {
{"slot_id", slot.id},
{"n_ctx_slot", slot.n_ctx}
});
const int ga_n = params.grp_attn_n;
const int ga_w = params.grp_attn_w;
@@ -436,7 +507,12 @@ struct llama_server_context
GGML_ASSERT(ga_w % ga_n == 0 && "ga_w must be a multiple of ga_n"); // NOLINT
//GGML_ASSERT(n_ctx_train % ga_w == 0 && "n_ctx_train must be a multiple of ga_w"); // NOLINT
//GGML_ASSERT(n_ctx >= n_ctx_train * ga_n && "n_ctx must be at least n_ctx_train * ga_n"); // NOLINT
LOG_TEE(" -> Slot %i - self-extend: ga_n = %d, ga_w = %d\n", slot.id, ga_n, ga_w);
LOG_INFO("slot self-extend", {
{"slot_id", slot.id},
{"ga_n", ga_n},
{"ga_w", ga_w}
});
}
slot.ga_i = 0;
@@ -730,10 +806,16 @@ struct llama_server_context
img_sl.img_data = clip_image_u8_init();
if (!clip_image_load_from_bytes(image_buffer.data(), image_buffer.size(), img_sl.img_data))
{
LOG_TEE("slot %i - failed to load image [id: %i]\n", slot->id, img_sl.id);
LOG_ERROR("failed to load image", {
{"slot_id", slot->id},
{"img_sl_id", img_sl.id}
});
return false;
}
LOG_TEE("slot %i - loaded image\n", slot->id);
LOG_VERBOSE("image loaded", {
{"slot_id", slot->id},
{"img_sl_id", img_sl.id}
});
img_sl.request_encode_image = true;
slot->images.push_back(img_sl);
}
@@ -793,7 +875,10 @@ struct llama_server_context
all_slots_are_idle = false;
LOG_TEE("slot %i is processing [task id: %i]\n", slot->id, slot->task_id);
LOG_INFO("slot is processing task", {
{"slot_id", slot->id},
{"task_id", slot->task_id},
});
return true;
}
@@ -818,10 +903,24 @@ struct llama_server_context
llama_batch_add(batch, system_tokens[i], i, { 0 }, false);
}
if (llama_decode(ctx, batch) != 0)
for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += params.n_batch)
{
LOG_TEE("%s: llama_decode() failed\n", __func__);
return;
const int32_t n_tokens = std::min(params.n_batch, (int32_t) (batch.n_tokens - i));
llama_batch batch_view = {
n_tokens,
batch.token + i,
nullptr,
batch.pos + i,
batch.n_seq_id + i,
batch.seq_id + i,
batch.logits + i,
0, 0, 0, // unused
};
if (llama_decode(ctx, batch_view) != 0)
{
LOG_TEE("%s: llama_decode() failed\n", __func__);
return;
}
}
// assign the system KV cache to all parallel sequences
@@ -1356,7 +1455,7 @@ struct llama_server_context
if (slot == nullptr)
{
// if no slot is available, we defer this task for processing later
LOG_VERBOSE("no slot is available", {});
LOG_VERBOSE("no slot is available", {{"task_id", task.id}});
queue_tasks.defer(task);
break;
}
@@ -1405,17 +1504,12 @@ struct llama_server_context
case TASK_TYPE_NEXT_RESPONSE: {
// do nothing
} break;
case TASK_TYPE_SLOTS_DATA: {
case TASK_TYPE_METRICS: {
json slots_data = json::array();
int n_idle_slots = 0;
int n_processing_slots = 0;
for (llama_client_slot &slot: slots) {
if (slot.available()) {
n_idle_slots++;
} else {
n_processing_slots++;
}
json slot_data = get_formated_generation(slot);
slot_data["id"] = slot.id;
slot_data["task_id"] = slot.task_id;
@@ -1430,19 +1524,48 @@ struct llama_server_context
{"stopped_limit", slot.stopped_limit},
{"stopping_word", slot.stopping_word},
};
if (slot_data["state"] == IDLE) {
n_idle_slots++;
} else {
n_processing_slots++;
}
slots_data.push_back(slot_data);
}
LOG_TEE("task %i - slots data: idle=%i processing=%i\n", task.id, n_idle_slots, n_processing_slots);
LOG_INFO("slot data", {
{"task_id", task.id},
{"n_idle_slots", n_idle_slots},
{"n_processing_slots", n_processing_slots}
});
LOG_VERBOSE("slot data", {
{"task_id", task.id},
{"n_idle_slots", n_idle_slots},
{"n_processing_slots", n_processing_slots},
{"slots", slots_data}
});
task_result res;
res.id = task.id;
res.multitask_id = task.multitask_id;
res.stop = true;
res.error = false;
res.result_json = {
{ "idle", n_idle_slots },
{ "processing", n_processing_slots },
{ "slots", slots_data }
{ "idle", n_idle_slots },
{ "processing", n_processing_slots },
{ "deferred", queue_tasks.queue_tasks_deferred.size() },
{ "n_prompt_tokens_processed_total", metrics.n_prompt_tokens_processed_total},
{ "n_tokens_predicted_total", metrics.n_tokens_predicted_total},
{ "n_prompt_tokens_processed", metrics.n_prompt_tokens_processed},
{ "t_prompt_processing", metrics.t_prompt_processing},
{ "n_tokens_predicted", metrics.n_tokens_predicted},
{ "t_tokens_generation", metrics.t_tokens_generation},
{ "kv_cache_tokens_count", llama_get_kv_cache_token_count(ctx)},
{ "kv_cache_used_cells", llama_get_kv_cache_used_cells(ctx)},
{ "slots", slots_data },
};
metrics.reset_bucket();
queue_results.send(res);
} break;
}
@@ -1470,7 +1593,7 @@ struct llama_server_context
bool update_slots() {
if (system_need_update)
{
LOG_TEE("updating system prompt\n");
LOG_INFO("updating system prompt", {});
update_system_prompt();
}
@@ -1480,12 +1603,13 @@ struct llama_server_context
{
if (system_prompt.empty() && clean_kv_cache)
{
LOG_TEE("all slots are idle and system prompt is empty, clear the KV cache\n");
LOG_INFO("all slots are idle and system prompt is empty, clear the KV cache", {});
kv_cache_clear();
}
return true;
}
LOG_VERBOSE("posting NEXT_RESPONSE", {});
task_server task;
task.type = TASK_TYPE_NEXT_RESPONSE;
task.target_id = -1;
@@ -1499,10 +1623,20 @@ struct llama_server_context
{
// Shift context
const int n_keep = slot.params.n_keep + add_bos_token;
const int n_left = system_tokens.size() + slot.n_past - n_keep;
const int n_left = (int) system_tokens.size() + slot.n_past - n_keep;
const int n_discard = n_left / 2;
LOG_TEE("slot %d: context shift - n_keep = %d, n_left = %d, n_discard = %d\n", slot.id, n_keep, n_left, n_discard);
LOG_INFO("slot context shift", {
{"slot_id", slot.id},
{"task_id", slot.task_id},
{"n_keep", n_keep},
{"n_left", n_left},
{"n_discard", n_discard},
{"n_ctx", n_ctx},
{"n_past", slot.n_past},
{"n_system_tokens", system_tokens.size()},
{"n_cache_tokens", slot.cache_tokens.size()}
});
llama_kv_cache_seq_rm (ctx, slot.id, n_keep , n_keep + n_discard);
llama_kv_cache_seq_shift(ctx, slot.id, n_keep + n_discard, system_tokens.size() + slot.n_past, -n_discard);
@@ -1516,17 +1650,12 @@ struct llama_server_context
slot.n_past -= n_discard;
slot.truncated = true;
LOG_VERBOSE("context shift", {
{ "n_ctx", n_ctx },
{ "n_keep", n_keep },
{ "n_left", n_left },
});
}
}
}
// decode any currently ongoing sequences
LOG_VERBOSE("decoding ongoing sequences", {});
for (auto & slot : slots)
{
// release the slot
@@ -1536,7 +1665,15 @@ struct llama_server_context
slot.command = NONE;
slot.t_last_used = ggml_time_us();
LOG_TEE("slot %d released (%d tokens in cache)\n", slot.id, (int) slot.cache_tokens.size());
LOG_INFO("slot released", {
{"slot_id", slot.id},
{"task_id", slot.task_id},
{"n_ctx", n_ctx},
{"n_past", slot.n_past},
{"n_system_tokens", system_tokens.size()},
{"n_cache_tokens", slot.cache_tokens.size()},
{"truncated", slot.truncated}
});
queue_tasks.notify_slot_changed();
continue;
@@ -1663,6 +1800,14 @@ struct llama_server_context
}
slot.n_past = common_part(slot.cache_tokens, prompt_tokens);
// the last token of the cache is not in the KV cache until the next call to llama_decode
// (it was sampled, pushed into the "cache_tokens", but not yet put in the context)
if (slot.n_past > 0 && slot.n_past == (int32_t) slot.cache_tokens.size())
{
slot.n_past -= 1;
}
slot.num_prompt_tokens_processed = slot.num_prompt_tokens - slot.n_past;
if (slot.ga_n != 1)
@@ -1684,7 +1829,12 @@ struct llama_server_context
slot.ga_i = ga_i;
}
LOG_TEE("slot %d : in cache: %i tokens | to process: %i tokens\n", slot.id, slot.n_past, slot.num_prompt_tokens_processed);
LOG_INFO("slot progression", {
{ "slot_id", slot.id },
{ "task_id", slot.task_id },
{ "n_past", slot.n_past },
{ "num_prompt_tokens_processed", slot.num_prompt_tokens_processed }
});
}
slot.cache_tokens = prompt_tokens;
@@ -1692,7 +1842,10 @@ struct llama_server_context
if (slot.n_past == slot.num_prompt_tokens && slot.n_past > 0)
{
// we have to evaluate at least 1 token to generate logits.
LOG_TEE("slot %d : we have to evaluate at least 1 token to generate logits\n", slot.id);
LOG_INFO("we have to evaluate at least 1 token to generate logits", {
{ "slot_id", slot.id },
{ "task_id", slot.task_id }
});
slot.n_past--;
if (slot.ga_i > 0)
{
@@ -1700,9 +1853,13 @@ struct llama_server_context
}
}
LOG_TEE("slot %d : kv cache rm - [%d, end)\n", slot.id, (int) system_tokens.size() + slot.n_past);
llama_kv_cache_seq_rm(ctx, slot.id, system_tokens.size() + slot.n_past, -1);
int p0 = (int) system_tokens.size() + slot.n_past;
LOG_INFO("kv cache rm [p0, end)", {
{ "slot_id", slot.id },
{ "task_id", slot.task_id },
{ "p0", p0 }
});
llama_kv_cache_seq_rm(ctx, slot.id, p0, -1);
LOG_VERBOSE("prompt ingested", {
{"n_past", slot.n_past},
@@ -1737,7 +1894,13 @@ struct llama_server_context
if (has_images && !ingest_images(slot, n_batch))
{
LOG_TEE("failed processing images\n");
LOG_ERROR("failed processing images", {
"slot_id", slot.id,
"task_id", slot.task_id,
});
// FIXME @phymbert: to be properly tested
// early returning without changing the slot state will block the slot for ever
// no one at the moment is checking the return value
return false;
}
@@ -1837,7 +2000,7 @@ struct llama_server_context
send_embedding(slot);
slot.release();
slot.i_batch = -1;
return true;
continue;
}
completion_token_output result;
@@ -1850,6 +2013,7 @@ struct llama_server_context
{
slot.t_start_genereration = ggml_time_us();
slot.t_prompt_processing = (slot.t_start_genereration - slot.t_start_process_prompt) / 1e3;
metrics.on_prompt_eval(slot);
}
llama_token_data_array cur_p = { slot.ctx_sampling->cur.data(), slot.ctx_sampling->cur.size(), false };
@@ -1872,11 +2036,14 @@ struct llama_server_context
slot.release();
slot.print_timings();
send_final_response(slot);
metrics.on_prediction(slot);
}
slot.i_batch = -1;
}
}
LOG_VERBOSE("slots updated", {});
return true;
}
@@ -1954,8 +2121,10 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
printf(" -ctv TYPE, --cache-type-v TYPE\n");
printf(" KV cache data type for V (default: f16)\n");
printf(" --mmproj MMPROJ_FILE path to a multimodal projector file for LLaVA.\n");
printf(" --log-format log output format: json or text (default: json)\n");
printf(" --log-disable disables logging to a file.\n");
printf(" --slots-endpoint-disable disables slots monitoring endpoint.\n");
printf(" --metrics enable prometheus compatible metrics endpoint (default: %s).\n", sparams.metrics_endpoint ? "enabled" : "disabled");
printf("\n");
printf(" -n, --n-predict maximum tokens to predict (default: %d)\n", params.n_predict);
printf(" --override-kv KEY=TYPE:VALUE\n");
@@ -2087,9 +2256,9 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
break;
}
std::string value(argv[i]);
/**/ if (value == "none") { params.rope_scaling_type = LLAMA_ROPE_SCALING_NONE; }
else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_LINEAR; }
else if (value == "yarn") { params.rope_scaling_type = LLAMA_ROPE_SCALING_YARN; }
/**/ if (value == "none") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_NONE; }
else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_LINEAR; }
else if (value == "yarn") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_YARN; }
else { invalid_param = true; break; }
}
else if (arg == "--rope-freq-base")
@@ -2213,15 +2382,15 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
std::string arg_next = argv[i];
if (arg_next == "none")
{
params.split_mode = LLAMA_SPLIT_NONE;
params.split_mode = LLAMA_SPLIT_MODE_NONE;
}
else if (arg_next == "layer")
{
params.split_mode = LLAMA_SPLIT_LAYER;
params.split_mode = LLAMA_SPLIT_MODE_LAYER;
}
else if (arg_next == "row")
{
params.split_mode = LLAMA_SPLIT_ROW;
params.split_mode = LLAMA_SPLIT_MODE_ROW;
}
else {
invalid_param = true;
@@ -2406,6 +2575,27 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
}
params.mmproj = argv[i];
}
else if (arg == "--log-format")
{
if (++i >= argc)
{
invalid_param = true;
break;
}
if (std::strcmp(argv[i], "json") == 0)
{
server_log_json = true;
}
else if (std::strcmp(argv[i], "text") == 0)
{
server_log_json = false;
}
else
{
invalid_param = true;
break;
}
}
else if (arg == "--log-disable")
{
log_set_target(stdout);
@@ -2415,6 +2605,10 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
{
sparams.slots_endpoint = false;
}
else if (arg == "--metrics")
{
sparams.metrics_endpoint = true;
}
else if (arg == "--chat-template")
{
if (++i >= argc)
@@ -2448,15 +2642,15 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
sep++;
if (strncmp(sep, "int:", 4) == 0) {
sep += 4;
kvo.tag = LLAMA_KV_OVERRIDE_INT;
kvo.tag = LLAMA_KV_OVERRIDE_TYPE_INT;
kvo.int_value = std::atol(sep);
} else if (strncmp(sep, "float:", 6) == 0) {
sep += 6;
kvo.tag = LLAMA_KV_OVERRIDE_FLOAT;
kvo.tag = LLAMA_KV_OVERRIDE_TYPE_FLOAT;
kvo.float_value = std::atof(sep);
} else if (strncmp(sep, "bool:", 5) == 0) {
sep += 5;
kvo.tag = LLAMA_KV_OVERRIDE_BOOL;
kvo.tag = LLAMA_KV_OVERRIDE_TYPE_BOOL;
if (std::strcmp(sep, "true") == 0) {
kvo.bool_value = true;
} else if (std::strcmp(sep, "false") == 0) {
@@ -2515,32 +2709,40 @@ static json format_partial_response(
static json format_tokenizer_response(const std::vector<llama_token> &tokens)
{
return json{
{"tokens", tokens}};
return json {
{"tokens", tokens}
};
}
static json format_detokenized_response(std::string content)
{
return json{
{"content", content}};
return json {
{"content", content}
};
}
static void log_server_request(const httplib::Request &req, const httplib::Response &res)
{
// skip GH copilot requests when using default port
if (req.path == "/v1/health" || req.path == "/v1/completions")
{
return;
}
LOG_INFO("request", {
{"remote_addr", req.remote_addr},
{"remote_port", req.remote_port},
{"status", res.status},
{"method", req.method},
{"path", req.path},
{"params", req.params},
});
{"remote_addr", req.remote_addr},
{"remote_port", req.remote_port},
{"status", res.status},
{"method", req.method},
{"path", req.path},
{"params", req.params},
});
LOG_VERBOSE("request", {
{"request", req.body},
{"response", res.body},
});
{"request", req.body},
{"response", res.body},
});
}
struct token_translator
@@ -2622,7 +2824,7 @@ int main(int argc, char **argv)
// request slots data using task queue
task_server task;
task.id = llama.queue_tasks.get_new_id();
task.type = TASK_TYPE_SLOTS_DATA;
task.type = TASK_TYPE_METRICS;
task.target_id = -1;
llama.queue_results.add_waiting_task_id(task.id);
@@ -2669,7 +2871,7 @@ int main(int argc, char **argv)
// request slots data using task queue
task_server task;
task.id = llama.queue_tasks.get_new_id();
task.type = TASK_TYPE_SLOTS_DATA;
task.type = TASK_TYPE_METRICS;
task.target_id = -1;
llama.queue_results.add_waiting_task_id(task.id);
@@ -2684,6 +2886,87 @@ int main(int argc, char **argv)
});
}
if (sparams.metrics_endpoint) {
svr.Get("/metrics", [&](const httplib::Request&, httplib::Response& res) {
// request slots data using task queue
task_server task;
task.id = llama.queue_tasks.get_new_id();
task.type = TASK_TYPE_METRICS;
task.target_id = -1;
llama.queue_results.add_waiting_task_id(task.id);
llama.queue_tasks.post(task);
// get the result
task_result result = llama.queue_results.recv(task.id);
llama.queue_results.remove_waiting_task_id(task.id);
json data = result.result_json;
uint64_t n_prompt_tokens_processed = data["n_prompt_tokens_processed"];
uint64_t t_prompt_processing = data["t_prompt_processing"];
uint64_t n_tokens_predicted = data["n_tokens_predicted"];
uint64_t t_tokens_generation = data["t_tokens_generation"];
int32_t kv_cache_used_cells = data["kv_cache_used_cells"];
// metrics definition: https://prometheus.io/docs/practices/naming/#metric-names
json all_metrics_def = json {
{"counter", {{
{"name", "prompt_tokens_total"},
{"help", "Number of prompt tokens processed."},
{"value", data["n_prompt_tokens_processed_total"]}
}, {
{"name", "tokens_predicted_total"},
{"help", "Number of generation tokens processed."},
{"value", data["n_tokens_predicted_total"]}
}}},
{"gauge", {{
{"name", "prompt_tokens_seconds"},
{"help", "Average prompt throughput in tokens/s."},
{"value", n_prompt_tokens_processed ? 1e3 / t_prompt_processing * n_prompt_tokens_processed : 0}
},{
{"name", "predicted_tokens_seconds"},
{"help", "Average generation throughput in tokens/s."},
{"value", n_tokens_predicted ? 1e3 / t_tokens_generation * n_tokens_predicted : 0}
},{
{"name", "kv_cache_usage_ratio"},
{"help", "KV-cache usage. 1 means 100 percent usage."},
{"value", 1. * kv_cache_used_cells / params.n_ctx}
},{
{"name", "kv_cache_tokens"},
{"help", "KV-cache tokens."},
{"value", data["kv_cache_tokens_count"]}
},{
{"name", "requests_processing"},
{"help", "Number of request processing."},
{"value", data["processing"]}
},{
{"name", "requests_deferred"},
{"help", "Number of request deferred."},
{"value", data["deferred"]}
}}}
};
std::stringstream prometheus;
for (const auto& el : all_metrics_def.items()) {
const auto& type = el.key();
const auto& metrics_def = el.value();
for (const auto& metric_def : metrics_def) {
std::string name = metric_def["name"];
std::string help = metric_def["help"];
prometheus << "# HELP llamacpp:" << name << " " << help << "\n"
<< "# TYPE llamacpp:" << name << " " << type << "\n"
<< "llamacpp:" << name << " " << metric_def["value"] << "\n";
}
}
res.set_content(prometheus.str(), "text/plain; version=0.0.4");
res.status = 200; // HTTP OK
});
}
svr.set_logger(log_server_request);
svr.set_exception_handler([](const httplib::Request &, httplib::Response &res, std::exception_ptr ep)
@@ -2736,9 +3019,6 @@ int main(int argc, char **argv)
// Set the base directory for serving static files
svr.set_base_dir(sparams.public_path);
// to make it ctrl+clickable:
LOG_TEE("\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port);
std::unordered_map<std::string, std::string> log_data;
log_data["hostname"] = sparams.hostname;
log_data["port"] = std::to_string(sparams.port);
@@ -2749,19 +3029,6 @@ int main(int argc, char **argv)
log_data["api_key"] = "api_key: " + std::to_string(sparams.api_keys.size()) + " keys loaded";
}
LOG_INFO("HTTP server listening", log_data);
// run the HTTP server in a thread - see comment below
std::thread t([&]()
{
if (!svr.listen_after_bind())
{
state.store(SERVER_STATE_ERROR);
return 1;
}
return 0;
});
// load the model
if (!llama.load_model(params))
{
@@ -3229,6 +3496,19 @@ int main(int argc, char **argv)
}*/
//);
LOG_INFO("HTTP server listening", log_data);
// run the HTTP server in a thread - see comment below
std::thread t([&]()
{
if (!svr.listen_after_bind())
{
state.store(SERVER_STATE_ERROR);
return 1;
}
return 0;
});
llama.queue_tasks.on_new_task(std::bind(
&llama_server_context::process_single_task, &llama, std::placeholders::_1));
llama.queue_tasks.on_finish_multitask(std::bind(
examples/server/tests/README.md
+47
View File
@@ -0,0 +1,47 @@
# Server tests
Python based server tests scenario using [BDD](https://en.wikipedia.org/wiki/Behavior-driven_development) and [behave](https://behave.readthedocs.io/en/latest/):
* [issues.feature](./features/issues.feature) Pending issues scenario
* [parallel.feature](./features/parallel.feature) Scenario involving multi slots and concurrent requests
* [security.feature](./features/security.feature) Security, CORS and API Key
* [server.feature](./features/server.feature) Server base scenario: completion, embedding, tokenization, etc...
Tests target GitHub workflows job runners with 4 vCPU.
Requests are using [aiohttp](https://docs.aiohttp.org/en/stable/client_reference.html), [asyncio](https://docs.python.org/fr/3/library/asyncio.html) based http client.
Note: If the host architecture inference speed is faster than GitHub runners one, parallel scenario may randomly fail. To mitigate it, you can increase values in `n_predict`, `kv_size`.
### Install dependencies
`pip install -r requirements.txt`
### Run tests
1. Build the server
```shell
cd ../../..
mkdir build
cd build
cmake ../
cmake --build . --target server
```
2. download required models:
1. `../../../scripts/hf.sh --repo ggml-org/models --file tinyllamas/stories260K.gguf`
3. Start the test: `./tests.sh`
It's possible to override some scenario steps values with environment variables:
- `PORT` -> `context.server_port` to set the listening port of the server during scenario, default: `8080`
- `LLAMA_SERVER_BIN_PATH` -> to change the server binary path, default: `../../../build/bin/server`
- `DEBUG` -> "ON" to enable steps and server verbose mode `--verbose`
- `SERVER_LOG_FORMAT_JSON` -> if set switch server logs to json format
### Run @bug, @wip or @wrong_usage annotated scenario
Feature or Scenario must be annotated with `@llama.cpp` to be included in the default scope.
- `@bug` annotation aims to link a scenario with a GitHub issue.
- `@wrong_usage` are meant to show user issue that are actually an expected behavior
- `@wip` to focus on a scenario working in progress
To run a scenario annotated with `@bug`, start:
`DEBUG=ON ./tests.sh --no-skipped --tags bug`
After changing logic in `steps.py`, ensure that `@bug` and `@wrong_usage` scenario are updated.
examples/server/tests/features/environment.py
+69
View File
@@ -0,0 +1,69 @@
import os
import socket
import subprocess
import time
from contextlib import closing
from signal import SIGKILL
def before_scenario(context, scenario):
print(f"\x1b[33;42mStarting new scenario: {scenario.name}!\x1b[0m")
port = 8080
if 'PORT' in os.environ:
port = int(os.environ['PORT'])
if is_server_listening("localhost", port):
assert False, "Server already started"
def after_scenario(context, scenario):
if context.server_process is None:
return
if scenario.status == "failed":
if 'GITHUB_ACTIONS' in os.environ:
print(f"\x1b[33;101mSCENARIO FAILED: {scenario.name} server logs:\x1b[0m\n\n")
if os.path.isfile('llama.log'):
with closing(open('llama.log', 'r')) as f:
for line in f:
print(line)
if not is_server_listening(context.server_fqdn, context.server_port):
print("\x1b[33;101mERROR: Server stopped listening\x1b[0m")
if not pid_exists(context.server_process.pid):
assert False, f"Server not running pid={context.server_process.pid} ..."
print(f"stopping server pid={context.server_process.pid} ...")
context.server_process.kill()
# Wait few for socket to free up
time.sleep(0.05)
attempts = 0
while is_server_listening(context.server_fqdn, context.server_port):
print(f"stopping server pid={context.server_process.pid} ...")
os.kill(context.server_process.pid, SIGKILL)
time.sleep(0.1)
attempts += 1
if attempts > 5:
print(f"Server dangling exits, killing all {context.server_path} ...")
process = subprocess.run(['killall', '-9', context.server_path],
stderr=subprocess.PIPE,
universal_newlines=True)
print(process)
def is_server_listening(server_fqdn, server_port):
with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:
result = sock.connect_ex((server_fqdn, server_port))
return result == 0
def pid_exists(pid):
"""Check whether pid exists in the current process table."""
import errno
if pid < 0:
return False
try:
os.kill(pid, 0)
except OSError as e:
return e.errno == errno.EPERM
else:
return True
examples/server/tests/features/issues.feature
+4
View File
@@ -0,0 +1,4 @@
# List of ongoing issues
@bug
Feature: Issues
# No confirmed issue at the moment
examples/server/tests/features/parallel.feature
+123
View File
@@ -0,0 +1,123 @@
@llama.cpp
Feature: Parallel
Background: Server startup
Given a server listening on localhost:8080
And a model file stories260K.gguf
And a model alias tinyllama-2
And 42 as server seed
And 64 KV cache size
And 2 slots
And embeddings extraction
And continuous batching
Then the server is starting
Then the server is healthy
Scenario Outline: Multi users completion
Given a prompt:
"""
Write a very long story about AI.
"""
And a prompt:
"""
Write another very long music lyrics.
"""
And <n_predict> max tokens to predict
Given concurrent completion requests
Then the server is busy
Then the server is idle
And all slots are idle
Then all prompts are predicted with <n_predict> tokens
Examples:
| n_predict |
| 128 |
Scenario Outline: Multi users OAI completions compatibility
Given a system prompt You are a writer.
And a model tinyllama-2
Given a prompt:
"""
Write a very long book.
"""
And a prompt:
"""
Write another a poem.
"""
And <n_predict> max tokens to predict
And streaming is <streaming>
Given concurrent OAI completions requests
Then the server is busy
Then the server is idle
Then all prompts are predicted with <n_predict> tokens
Examples:
| streaming | n_predict |
| disabled | 128 |
| enabled | 64 |
Scenario: Multi users with total number of tokens to predict exceeds the KV Cache size #3969
Given a prompt:
"""
Write a very long story about AI.
"""
And a prompt:
"""
Write another very long music lyrics.
"""
And a prompt:
"""
Write a very long poem.
"""
And a prompt:
"""
Write a very long joke.
"""
And 128 max tokens to predict
Given concurrent completion requests
Then the server is busy
Then the server is idle
Then all prompts are predicted
Scenario: Multi users embeddings
Given a prompt:
"""
Write a very long story about AI.
"""
And a prompt:
"""
Write another very long music lyrics.
"""
And a prompt:
"""
Write a very long poem.
"""
And a prompt:
"""
Write a very long joke.
"""
Given concurrent embedding requests
Then the server is busy
Then the server is idle
Then all embeddings are generated
Scenario: Multi users OAI compatibility embeddings
Given a prompt:
"""
In which country Paris is located ?
"""
And a prompt:
"""
Is Madrid the capital of Spain ?
"""
And a prompt:
"""
What is the biggest US city ?
"""
And a prompt:
"""
What is the capital of Bulgaria ?
"""
And a model tinyllama-2
Given concurrent OAI embedding requests
Then the server is busy
Then the server is idle
Then all embeddings are generated
examples/server/tests/features/security.feature
+50
View File
@@ -0,0 +1,50 @@
@llama.cpp
Feature: Security
Background: Server startup with an api key defined
Given a server listening on localhost:8080
And a model file stories260K.gguf
And a server api key llama.cpp
Then the server is starting
Then the server is healthy
Scenario Outline: Completion with some user api key
Given a prompt test
And a user api key <api_key>
And 4 max tokens to predict
And a completion request with <api_error> api error
Examples: Prompts
| api_key | api_error |
| llama.cpp | no |
| llama.cpp | no |
| hackeme | raised |
| | raised |
Scenario Outline: OAI Compatibility
Given a system prompt test
And a user prompt test
And a model test
And 2 max tokens to predict
And streaming is disabled
And a user api key <api_key>
Given an OAI compatible chat completions request with <api_error> api error
Examples: Prompts
| api_key | api_error |
| llama.cpp | no |
| llama.cpp | no |
| hackme | raised |
Scenario Outline: CORS Options
When an OPTIONS request is sent from <origin>
Then CORS header <cors_header> is set to <cors_header_value>
Examples: Headers
| origin | cors_header | cors_header_value |
| localhost | Access-Control-Allow-Origin | localhost |
| web.mydomain.fr | Access-Control-Allow-Origin | web.mydomain.fr |
| origin | Access-Control-Allow-Credentials | true |
| web.mydomain.fr | Access-Control-Allow-Methods | POST |
| web.mydomain.fr | Access-Control-Allow-Headers | * |
examples/server/tests/features/server.feature
+84
View File
@@ -0,0 +1,84 @@
@llama.cpp
Feature: llama.cpp server
Background: Server startup
Given a server listening on localhost:8080
And a model file stories260K.gguf
And a model alias tinyllama-2
And 42 as server seed
# KV Cache corresponds to the total amount of tokens
# that can be stored across all independent sequences: #4130
# see --ctx-size and #5568
And 32 KV cache size
And 1 slots
And embeddings extraction
And 32 server max tokens to predict
And prometheus compatible metrics exposed
Then the server is starting
Then the server is healthy
Scenario: Health
Then the server is ready
And all slots are idle
Scenario Outline: Completion
Given a prompt <prompt>
And <n_predict> max tokens to predict
And a completion request with no api error
Then <n_predicted> tokens are predicted matching <re_content>
And prometheus metrics are exposed
Examples: Prompts
| prompt | n_predict | re_content | n_predicted |
| I believe the meaning of life is | 8 | (read<or>going)+ | 8 |
| Write a joke about AI | 64 | (park<or>friends<or>scared<or>always)+ | 32 |
Scenario Outline: OAI Compatibility
Given a model <model>
And a system prompt <system_prompt>
And a user prompt <user_prompt>
And <max_tokens> max tokens to predict
And streaming is <enable_streaming>
Given an OAI compatible chat completions request with no api error
Then <n_predicted> tokens are predicted matching <re_content>
Examples: Prompts
| model | system_prompt | user_prompt | max_tokens | re_content | n_predicted | enable_streaming |
| llama-2 | Book | What is the best book | 8 | (Mom<or>what)+ | 8 | disabled |
| codellama70b | You are a coding assistant. | Write the fibonacci function in c++. | 64 | (thanks<or>happy<or>bird)+ | 32 | enabled |
Scenario: Embedding
When embeddings are computed for:
"""
What is the capital of Bulgaria ?
"""
Then embeddings are generated
Scenario: OAI Embeddings compatibility
Given a model tinyllama-2
When an OAI compatible embeddings computation request for:
"""
What is the capital of Spain ?
"""
Then embeddings are generated
Scenario: OAI Embeddings compatibility with multiple inputs
Given a model tinyllama-2
Given a prompt:
"""
In which country Paris is located ?
"""
And a prompt:
"""
Is Madrid the capital of Spain ?
"""
When an OAI compatible embeddings computation request for multiple inputs
Then embeddings are generated
Scenario: Tokenize / Detokenize
When tokenizing:
"""
What is the capital of France ?
"""
Then tokens can be detokenize
examples/server/tests/features/steps/steps.py
+801
View File
@@ -0,0 +1,801 @@
import asyncio
import collections
import json
import os
import re
import socket
import subprocess
import time
from contextlib import closing
from re import RegexFlag
import aiohttp
import openai
from behave import step
from behave.api.async_step import async_run_until_complete
from prometheus_client import parser
@step(u"a server listening on {server_fqdn}:{server_port}")
def step_server_config(context, server_fqdn, server_port):
context.server_fqdn = server_fqdn
context.server_port = int(server_port)
if 'PORT' in os.environ:
context.server_port = int(os.environ['PORT'])
print(f"$PORT set, overriding server port with to {context.server_port}")
context.base_url = f'http://{context.server_fqdn}:{context.server_port}'
context.debug = 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON'
context.model_alias = None
context.n_ctx = None
context.n_predict = None
context.n_server_predict = None
context.n_slots = None
context.server_api_key = None
context.server_continuous_batching = False
context.server_embeddings = False
context.server_metrics = False
context.server_process = None
context.server_seed = None
context.user_api_key = None
context.tasks_result = []
context.concurrent_tasks = []
context.prompts = []
@step(u'a model file {model_file}')
def step_model_file(context, model_file):
context.model_file = model_file
@step(u'a model alias {model_alias}')
def step_model_alias(context, model_alias):
context.model_alias = model_alias
@step(u'{seed} as server seed')
def step_seed(context, seed):
context.server_seed = int(seed)
@step(u'{n_ctx} KV cache size')
def step_n_ctx(context, n_ctx):
context.n_ctx = int(n_ctx)
@step(u'{n_slots} slots')
def step_n_slots(context, n_slots):
context.n_slots = int(n_slots)
@step(u'{n_predict} server max tokens to predict')
def step_server_n_predict(context, n_predict):
context.n_server_predict = int(n_predict)
@step(u'continuous batching')
def step_server_continuous_batching(context):
context.server_continuous_batching = True
@step(u'embeddings extraction')
def step_server_embeddings(context):
context.server_embeddings = True
@step(u'prometheus compatible metrics exposed')
def step_server_metrics(context):
context.server_metrics = True
@step(u"the server is starting")
def step_start_server(context):
start_server_background(context)
attempts = 0
while True:
with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:
result = sock.connect_ex((context.server_fqdn, context.server_port))
if result == 0:
print("\x1b[33;46mserver started!\x1b[0m")
return
attempts += 1
if attempts > 20:
assert False, "server not started"
print(f"waiting for server to start, connect error code = {result}...")
time.sleep(0.1)
@step(u"the server is {expecting_status}")
@async_run_until_complete
async def step_wait_for_the_server_to_be_started(context, expecting_status):
match expecting_status:
case 'healthy':
await wait_for_health_status(context, context.base_url, 200, 'ok')
case 'ready' | 'idle':
await wait_for_health_status(context, context.base_url, 200, 'ok',
params={'fail_on_no_slot': 0, 'include_slots': 0},
slots_idle=context.n_slots,
slots_processing=0,
expected_slots=[{'id': slot_id, 'state': 0}
for slot_id in range(context.n_slots)])
case 'busy':
await wait_for_health_status(context, context.base_url, 503,
'no slot available',
params={'fail_on_no_slot': 0, 'include_slots': 0},
slots_idle=0,
slots_processing=context.n_slots,
expected_slots=[{'id': slot_id, 'state': 1}
for slot_id in range(context.n_slots)])
case _:
assert False, "unknown status"
@step(u'all slots are {expected_slot_status_string}')
@async_run_until_complete
async def step_all_slots_status(context, expected_slot_status_string):
match expected_slot_status_string:
case 'idle':
expected_slot_status = 0
case 'busy':
expected_slot_status = 1
case _:
assert False, "unknown status"
expected_slots = [{'id': slot_id, 'state': expected_slot_status}
for slot_id in range(context.n_slots)]
await request_slots_status(context, expected_slots)
@step(u'a completion request with {api_error} api error')
@async_run_until_complete
async def step_request_completion(context, api_error):
expect_api_error = api_error == 'raised'
completion = await request_completion(context.prompts.pop(),
context.base_url,
debug=context.debug,
n_predict=context.n_predict,
server_seed=context.server_seed,
expect_api_error=expect_api_error,
user_api_key=context.user_api_key)
context.tasks_result.append(completion)
if context.debug:
print(f"Completion response: {completion}")
if expect_api_error:
assert completion == 401, f"completion must be an 401 status code: {completion}"
@step(u'{predicted_n} tokens are predicted matching {re_content}')
def step_n_tokens_predicted_with_content(context, predicted_n, re_content):
assert_n_tokens_predicted(context.tasks_result.pop(), int(predicted_n), re_content)
@step(u'{predicted_n} tokens are predicted')
def step_n_tokens_predicted(context, predicted_n):
assert_n_tokens_predicted(context.tasks_result.pop(), int(predicted_n))
@step(u'a user prompt {user_prompt}')
def step_user_prompt(context, user_prompt):
context.prompts.append(user_prompt)
@step(u'a system prompt {system_prompt}')
def step_system_prompt(context, system_prompt):
context.system_prompt = system_prompt
@step(u'a model {model}')
def step_model(context, model):
context.model = model
@step(u'{max_tokens} max tokens to predict')
def step_max_tokens(context, max_tokens):
context.n_predict = int(max_tokens)
@step(u'streaming is {enable_streaming}')
def step_streaming(context, enable_streaming):
context.enable_streaming = enable_streaming == 'enabled'
@step(u'a user api key {user_api_key}')
def step_user_api_key(context, user_api_key):
context.user_api_key = user_api_key
@step(u'no user api key')
def step_no_user_api_key(context):
context.user_api_key = None
@step(u'a user api key ')
def step_no_user_api_key_space(context):
context.user_api_key = None
@step(u'a server api key {server_api_key}')
def step_server_api_key(context, server_api_key):
context.server_api_key = server_api_key
@step(u'an OAI compatible chat completions request with {api_error} api error')
@async_run_until_complete
async def step_oai_chat_completions(context, api_error):
if context.debug:
print(f"Submitting OAI compatible completions request...")
expect_api_error = api_error == 'raised'
completion = await oai_chat_completions(context.prompts.pop(),
context.system_prompt,
context.base_url,
False,
model=context.model if hasattr(context, 'model') else None,
n_predict=context.n_predict
if hasattr(context, 'n_predict') else None,
enable_streaming=context.enable_streaming
if hasattr(context, 'enable_streaming') else None,
server_seed=context.server_seed
if hasattr(context, 'server_seed') else None,
user_api_key=context.user_api_key
if hasattr(context, 'user_api_key') else None,
expect_api_error=expect_api_error)
context.tasks_result.append(completion)
if context.debug:
print(f"Completion response: {completion}")
if expect_api_error:
assert completion == 401, f"completion must be an 401 status code: {completion}"
if context.debug:
print(f"Completion response: {completion}")
@step(u'a prompt')
def step_a_prompt(context):
context.prompts.append(context.text)
@step(u'a prompt {prompt}')
def step_a_prompt_prompt(context, prompt):
context.prompts.append(prompt)
@step(u'concurrent completion requests')
@async_run_until_complete()
async def step_concurrent_completion_requests(context):
await concurrent_requests(context,
request_completion,
# prompt is inserted automatically
context.base_url,
debug=context.debug,
n_predict=context.n_predict if hasattr(context, 'n_predict') else None,
server_seed=context.server_seed if hasattr(context, 'server_seed') else None,
user_api_key=context.user_api_key if hasattr(context,
'user_api_key') else None)
@step(u'concurrent OAI completions requests')
@async_run_until_complete
async def step_oai_chat_completions(context):
await concurrent_requests(context, oai_chat_completions,
# user_prompt is inserted automatically
context.system_prompt,
context.base_url,
True, # async_client
model=context.model
if hasattr(context, 'model') else None,
n_predict=context.n_predict
if hasattr(context, 'n_predict') else None,
enable_streaming=context.enable_streaming
if hasattr(context, 'enable_streaming') else None,
server_seed=context.server_seed
if hasattr(context, 'server_seed') else None,
user_api_key=context.user_api_key
if hasattr(context, 'user_api_key') else None)
@step(u'all prompts are predicted')
@async_run_until_complete
async def step_all_prompts_are_predicted(context):
await all_prompts_are_predicted(context)
@step(u'all prompts are predicted with {n_predict} tokens')
@async_run_until_complete
async def step_all_prompts_are_predicted_with_n_tokens(context, n_predict):
expected_predicted_n = int(n_predict)
await all_prompts_are_predicted(context, expected_predicted_n)
async def all_prompts_are_predicted(context, expected_predicted_n=None):
n_completions = await gather_tasks_results(context)
assert n_completions > 0
for i in range(n_completions):
assert_n_tokens_predicted(context.tasks_result.pop(), expected_predicted_n=expected_predicted_n)
assert len(context.concurrent_tasks) == 0, f"{len(context.concurrent_tasks)} pending requests"
@step(u'embeddings are computed for')
@async_run_until_complete
async def step_compute_embedding(context):
context.embeddings = await request_embedding(context.text, base_url=context.base_url)
@step(u'embeddings are generated')
def step_assert_embeddings(context):
if len(context.prompts) == 0:
assert_embeddings(context.embeddings)
else:
assert len(context.embeddings) == len(context.prompts), (f"unexpected response:\n"
f"context.prompts={context.prompts}\n"
f"context.embeddings={context.embeddings}")
for embedding in context.embeddings:
context.prompts.pop()
assert_embeddings(embedding)
@step(u'an OAI compatible embeddings computation request for')
@async_run_until_complete
async def step_oai_compute_embeddings(context):
context.embeddings = await request_oai_embeddings(context.text,
base_url=context.base_url,
user_api_key=context.user_api_key,
model=context.model)
@step(u'an OAI compatible embeddings computation request for multiple inputs')
@async_run_until_complete
async def step_oai_compute_embeddings_multiple_inputs(context):
context.embeddings = await request_oai_embeddings(context.prompts,
base_url=context.base_url,
user_api_key=context.user_api_key,
model=context.model)
@step(u'concurrent embedding requests')
@async_run_until_complete()
async def step_concurrent_embedding_requests(context):
await concurrent_requests(context,
request_embedding,
# prompt is inserted automatically
base_url=context.base_url)
@step(u'concurrent OAI embedding requests')
@async_run_until_complete()
async def step_concurrent_oai_embedding_requests(context):
await concurrent_requests(context,
request_oai_embeddings,
# prompt is inserted automatically
base_url=context.base_url,
async_client=True,
model=context.model)
@step(u'all embeddings are generated')
@async_run_until_complete()
async def all_embeddings_are_generated(context):
n_embedding_requests = await gather_tasks_results(context)
assert n_embedding_requests > 0
for i in range(n_embedding_requests):
assert_embeddings(context.tasks_result.pop())
@step(u'tokenizing')
@async_run_until_complete
async def step_tokenize(context):
context.tokenized_text = context.text
async with aiohttp.ClientSession() as session:
async with session.post(f'{context.base_url}/tokenize',
json={
"content": context.tokenized_text,
}) as response:
assert response.status == 200
tokenize_json = await response.json()
context.tokens = tokenize_json['tokens']
@step(u'tokens can be detokenize')
@async_run_until_complete
async def step_detokenize(context):
assert len(context.tokens) > 0
async with aiohttp.ClientSession() as session:
async with session.post(f'{context.base_url}/detokenize',
json={
"tokens": context.tokens,
}) as response:
assert response.status == 200
detokenize_json = await response.json()
# SPM tokenizer adds a whitespace prefix: https://github.com/google/sentencepiece/issues/15
assert context.tokenized_text == detokenize_json['content'].strip()
@step(u'an OPTIONS request is sent from {origin}')
@async_run_until_complete
async def step_options_request(context, origin):
async with aiohttp.ClientSession() as session:
async with session.options(f'{context.base_url}/v1/chat/completions',
headers={"Origin": origin}) as response:
assert response.status == 200
context.options_response = response
@step(u'CORS header {cors_header} is set to {cors_header_value}')
def step_check_options_header_value(context, cors_header, cors_header_value):
assert context.options_response.headers[cors_header] == cors_header_value
@step(u'prometheus metrics are exposed')
@async_run_until_complete
async def step_prometheus_metrics_exported(context):
async with aiohttp.ClientSession() as session:
async with await session.get(f'{context.base_url}/metrics') as metrics_response:
assert metrics_response.status == 200
assert metrics_response.headers['Content-Type'] == "text/plain; version=0.0.4"
metrics_raw = await metrics_response.text()
metric_exported = False
for metric in parser.text_string_to_metric_families(metrics_raw):
match metric.name:
case "llamacpp:kv_cache_usage_ratio":
assert len(metric.samples) > 0
metric_exported = True
assert metric_exported, "No metrics exported"
async def concurrent_requests(context, f_completion, *args, **kwargs):
n_prompts = len(context.prompts)
if context.debug:
print(f"starting {n_prompts} concurrent completion requests...")
assert n_prompts > 0
for prompt_no in range(n_prompts):
shifted_args = [context.prompts.pop(), *args]
context.concurrent_tasks.append(asyncio.create_task(f_completion(*shifted_args, **kwargs)))
await asyncio.sleep(0.1)
async def request_completion(prompt,
base_url,
debug=False,
n_predict=None,
server_seed=None,
expect_api_error=None,
user_api_key=None):
if debug:
print(f"Sending completion request: {prompt}")
origin = "my.super.domain"
headers = {
'Origin': origin
}
if user_api_key is not None:
if debug:
print(f"Set user_api_key: {user_api_key}")
headers['Authorization'] = f'Bearer {user_api_key}'
async with aiohttp.ClientSession() as session:
async with session.post(f'{base_url}/completion',
json={
"prompt": prompt,
"n_predict": int(n_predict) if n_predict is not None else -1,
"seed": server_seed if server_seed is not None else 42
},
headers=headers) as response:
if expect_api_error is None or not expect_api_error:
assert response.status == 200
assert response.headers['Access-Control-Allow-Origin'] == origin
return await response.json()
else:
return response.status
async def oai_chat_completions(user_prompt,
system_prompt,
base_url,
async_client,
debug=False,
model=None,
n_predict=None,
enable_streaming=None,
server_seed=None,
user_api_key=None,
expect_api_error=None):
if debug:
print(f"Sending OAI Chat completions request: {user_prompt}")
# openai client always expects an api key
user_api_key = user_api_key if user_api_key is not None else 'nope'
seed = server_seed if server_seed is not None else 42
enable_streaming = enable_streaming if enable_streaming is not None else False
payload = {
"messages": [
{
"role": "system",
"content": system_prompt,
},
{
"role": "user",
"content": user_prompt,
}
],
"model": model,
"max_tokens": n_predict,
"stream": enable_streaming,
"seed": seed
}
completion_response = {
'content': '',
'timings': {
'predicted_n': 0
}
}
if async_client:
origin = 'llama.cpp'
headers = {'Authorization': f'Bearer {user_api_key}', 'Origin': origin}
async with aiohttp.ClientSession() as session:
async with session.post(f'{base_url}/v1/chat/completions',
json=payload,
headers=headers) as response:
if enable_streaming:
assert response.status == 200
assert response.headers['Access-Control-Allow-Origin'] == origin
assert response.headers['Content-Type'] == "text/event-stream"
event_received = True
while event_received:
event_received = False
async for line_in_bytes in response.content:
line = line_in_bytes.decode('utf8')
line = line.rstrip('\n').rstrip('\r')
if line == '':
continue
event_data = line.split(': ', 1)
assert event_data[0] == 'data', f'Bad event code received: ```{event_data}```'
chunk_raw = event_data[1]
chunk = json.loads(chunk_raw)
assert len(chunk['choices']) == 1, f"no choices provided, line ```{line}```"
delta = chunk['choices'][0]['delta']
if 'content' in delta:
completion_response['content'] += delta['content']
completion_response['timings']['predicted_n'] += 1
else:
if expect_api_error is None or not expect_api_error:
assert response.status == 200
assert response.headers['Access-Control-Allow-Origin'] == origin
assert response.headers['Content-Type'] == "application/json; charset=utf-8"
chat_completion_raw = await response.json()
completion_response = {
'content': chat_completion_raw['choices'][0]['message'],
'timings': {
'predicted_n': chat_completion_raw['usage']['completion_tokens']
}
}
else:
return response.status
else:
try:
openai.api_key = user_api_key
openai.api_base = f'{base_url}/v1/chat'
chat_completion = openai.Completion.create(
messages=payload['messages'],
model=model,
max_tokens=n_predict,
stream=enable_streaming,
seed=seed
)
except openai.error.APIError as e:
if expect_api_error is not None and expect_api_error:
return 401
else:
assert False, f'error raised: {e}'
if enable_streaming:
for chunk in chat_completion:
assert len(chunk.choices) == 1
delta = chunk.choices[0].delta
if 'content' in delta:
completion_response['content'] += delta['content']
completion_response['timings']['predicted_n'] += 1
else:
assert len(chat_completion.choices) == 1
completion_response = {
'content': chat_completion.choices[0].message.content,
'timings': {
'predicted_n': chat_completion.usage.completion_tokens
}
}
if debug:
print("OAI response formatted to llama.cpp:", completion_response)
return completion_response
async def request_embedding(content, base_url=None):
async with aiohttp.ClientSession() as session:
async with session.post(f'{base_url}/embedding',
json={
"content": content,
}) as response:
assert response.status == 200
response_json = await response.json()
return response_json['embedding']
async def request_oai_embeddings(input,
base_url=None, user_api_key=None,
model=None, async_client=False):
# openai client always expects an api_key
user_api_key = user_api_key if user_api_key is not None else 'nope'
if async_client:
origin = 'llama.cpp'
if user_api_key is not None:
headers = {'Authorization': f'Bearer {user_api_key}', 'Origin': origin}
async with aiohttp.ClientSession() as session:
async with session.post(f'{base_url}/v1/embeddings',
json={
"input": input,
"model": model,
},
headers=headers) as response:
assert response.status == 200, f"received status code not expected: {response.status}"
assert response.headers['Access-Control-Allow-Origin'] == origin
assert response.headers['Content-Type'] == "application/json; charset=utf-8"
response_json = await response.json()
assert response_json['model'] == model, f"invalid model received: {response_json['model']}"
assert response_json['object'] == 'list'
return response_json['data']
else:
openai.api_key = user_api_key
openai.api_base = f'{base_url}/v1'
oai_embeddings = openai.Embedding.create(
model=model,
input=input,
)
if isinstance(input, collections.abc.Sequence):
embeddings = []
for an_oai_embeddings in oai_embeddings.data:
embeddings.append(an_oai_embeddings.embedding)
else:
embeddings = oai_embeddings.data.embedding
return embeddings
def assert_n_tokens_predicted(completion_response, expected_predicted_n=None, re_content=None):
content = completion_response['content']
n_predicted = completion_response['timings']['predicted_n']
assert len(content) > 0, "no token predicted"
if expected_predicted_n is not None:
assert n_predicted == expected_predicted_n, (f'invalid number of tokens predicted:'
f' {n_predicted} <> {expected_predicted_n}')
if re_content is not None:
re_content = '^.*' + re_content.replace('<or>', '|') + '.*$'
assert re.match(re_content, content, flags=RegexFlag.IGNORECASE | RegexFlag.MULTILINE | RegexFlag.DOTALL), (
f'invalid tokens predicted:'
f' ```\n{content}\n``` do not match /{re_content}/')
async def gather_tasks_results(context):
n_tasks = len(context.concurrent_tasks)
if context.debug:
print(f"Waiting for all {n_tasks} tasks results...")
for task_no in range(n_tasks):
context.tasks_result.append(await context.concurrent_tasks.pop())
n_completions = len(context.tasks_result)
return n_completions
async def wait_for_health_status(context,
base_url,
expected_http_status_code,
expected_health_status,
params=None,
slots_idle=None,
slots_processing=None,
expected_slots=None):
if context.debug:
print(f"Starting checking for health for expected_health_status={expected_health_status}")
timeout = 3 # seconds
interval = 0.5
counter = 0
async with aiohttp.ClientSession() as session:
while True:
async with await session.get(f'{base_url}/health', params=params) as health_response:
status_code = health_response.status
health = await health_response.json()
if context.debug:
print(f"HEALTH - response for expected health status='{expected_health_status}' on "
f"'{base_url}/health'?{params} is {health}")
if (status_code == expected_http_status_code
and health['status'] == expected_health_status
and (slots_idle is None or health['slots_idle'] == slots_idle)
and (slots_processing is None or health['slots_processing'] == slots_processing)):
if expected_slots is not None:
assert_slots_status(health['slots'], expected_slots)
return
if (status_code == expected_http_status_code
and health['status'] == expected_health_status
and (slots_idle is None or health['slots_idle'] == slots_idle)
and (slots_processing is None or health['slots_processing'] == slots_processing)):
if expected_slots is not None:
assert_slots_status(health['slots'], expected_slots)
return
await asyncio.sleep(interval)
counter += interval
if counter >= timeout:
# Sometimes health requests are triggered after completions are predicted
if expected_http_status_code == 503:
if len(context.tasks_result) == 0:
print("\x1b[5;37;43mWARNING: forcing concurrent tasks,"
" busy health check missed, probably too fast inference\x1b[0m")
n_completions = await gather_tasks_results(context)
if n_completions > 0:
return
assert False, 'timeout exceeded'
def assert_embeddings(embeddings):
assert len(embeddings) > 0
embeddings_computed = False
for emb in embeddings:
if emb != 0:
embeddings_computed = True
assert embeddings_computed, f"Embeddings: {embeddings}"
async def request_slots_status(context, expected_slots):
async with aiohttp.ClientSession() as session:
async with await session.get(f'{context.base_url}/slots') as slots_response:
assert slots_response.status == 200
slots = await slots_response.json()
assert_slots_status(slots, expected_slots)
def assert_slots_status(slots, expected_slots):
assert len(slots) == len(expected_slots)
for slot_id, (expected, slot) in enumerate(zip(expected_slots, slots)):
for key in expected:
assert expected[key] == slot[key], (f"invalid slot {slot_id}"
f" expected[{key}] != slot[{key}]"
f" = {expected[key]} != {slot[key]}")
def start_server_background(context):
context.server_path = '../../../build/bin/server'
if 'LLAMA_SERVER_BIN_PATH' in os.environ:
context.server_path = os.environ['LLAMA_SERVER_BIN_PATH']
server_args = [
'--host', context.server_fqdn,
'--port', context.server_port,
'--model', context.model_file
]
if context.server_continuous_batching:
server_args.append('--cont-batching')
if context.server_embeddings:
server_args.append('--embedding')
if context.server_metrics:
server_args.append('--metrics')
if context.model_alias is not None:
server_args.extend(['--alias', context.model_alias])
if context.n_ctx is not None:
server_args.extend(['--ctx-size', context.n_ctx])
if context.n_slots is not None:
server_args.extend(['--parallel', context.n_slots])
if context.n_server_predict is not None:
server_args.extend(['--n-predict', context.n_server_predict])
if context.server_api_key is not None:
server_args.extend(['--api-key', context.server_api_key])
if context.debug:
server_args.append('--verbose')
if 'SERVER_LOG_FORMAT_JSON' not in os.environ:
server_args.extend(['--log-format', "text"])
print(f"starting server with: {context.server_path}", *server_args)
context.server_process = subprocess.Popen(
[str(arg) for arg in [context.server_path, *server_args]],
close_fds=True)
print(f"server pid={context.server_process.pid}")
examples/server/tests/features/wrong_usages.feature
+21
View File
@@ -0,0 +1,21 @@
# run with ./test.sh --tags wrong_usage
@wrong_usage
Feature: Wrong usage of llama.cpp server
#3969 The user must always set --n-predict option
# to cap the number of tokens any completion request can generate
# or pass n_predict/max_tokens in the request.
Scenario: Infinite loop
Given a server listening on localhost:8080
And a model file stories260K.gguf
# Uncomment below to fix the issue
#And 64 server max tokens to predict
Then the server is starting
Given a prompt:
"""
Go to: infinite loop
"""
# Uncomment below to fix the issue
#And 128 max tokens to predict
Given concurrent completion requests
Then all prompts are predicted
examples/server/tests/requirements.txt
+4
View File
@@ -0,0 +1,4 @@
aiohttp~=3.9.3
behave~=1.2.6
openai~=0.25.0
prometheus-client~=0.20.0
examples/server/tests/tests.sh
Executable
+12
View File
@@ -0,0 +1,12 @@
#!/bin/bash
set -eu
if [ $# -lt 1 ]
then
# Start @llama.cpp scenario
behave --summary --stop --no-capture --exclude 'issues|wrong_usages' --tags llama.cpp
else
behave "$@"
fi
examples/server/utils.hpp
+55 -29
View File
@@ -14,6 +14,7 @@
using json = nlohmann::json;
extern bool server_verbose;
extern bool server_log_json;
#ifndef SERVER_VERBOSE
#define SERVER_VERBOSE 1
@@ -27,14 +28,14 @@ extern bool server_verbose;
{ \
if (server_verbose) \
{ \
server_log("VERBOSE", __func__, __LINE__, MSG, __VA_ARGS__); \
server_log("VERB", __func__, __LINE__, MSG, __VA_ARGS__); \
} \
} while (0)
#endif
#define LOG_ERROR( MSG, ...) server_log("ERROR", __func__, __LINE__, MSG, __VA_ARGS__)
#define LOG_WARNING(MSG, ...) server_log("WARNING", __func__, __LINE__, MSG, __VA_ARGS__)
#define LOG_INFO( MSG, ...) server_log("INFO", __func__, __LINE__, MSG, __VA_ARGS__)
#define LOG_ERROR( MSG, ...) server_log("ERR", __func__, __LINE__, MSG, __VA_ARGS__)
#define LOG_WARNING(MSG, ...) server_log("WARN", __func__, __LINE__, MSG, __VA_ARGS__)
#define LOG_INFO( MSG, ...) server_log("INFO", __func__, __LINE__, MSG, __VA_ARGS__)
//
// parallel
@@ -50,7 +51,7 @@ enum task_type {
TASK_TYPE_COMPLETION,
TASK_TYPE_CANCEL,
TASK_TYPE_NEXT_RESPONSE,
TASK_TYPE_SLOTS_DATA
TASK_TYPE_METRICS
};
struct task_server {
@@ -133,26 +134,48 @@ struct completion_token_output
std::string text_to_send;
};
static inline void server_log(const char *level, const char *function, int line,
const char *message, const nlohmann::ordered_json &extra)
static inline void server_log(const char *level, const char *function, int line, const char *message, const nlohmann::ordered_json &extra)
{
nlohmann::ordered_json log
{
std::stringstream ss_tid;
ss_tid << std::this_thread::get_id();
json log = nlohmann::ordered_json{
{"tid", ss_tid.str()},
{"timestamp", time(nullptr)},
{"level", level},
{"function", function},
{"line", line},
{"message", message},
};
if (!extra.empty())
{
log.merge_patch(extra);
}
if (server_log_json) {
log.merge_patch(
{
{"level", level},
{"function", function},
{"line", line},
{"msg", message},
});
if (!extra.empty()) {
log.merge_patch(extra);
}
const std::string str = log.dump(-1, ' ', false, json::error_handler_t::replace);
printf("%.*s\n", (int)str.size(), str.data());
fflush(stdout);
std::cout << log.dump(-1, ' ', false, json::error_handler_t::replace) << "\n" << std::flush;
} else {
char buf[1024];
snprintf(buf, 1024, "%4s [%24s] %s", level, function, message);
if (!extra.empty()) {
log.merge_patch(extra);
}
std::stringstream ss;
ss << buf << " |";
for (const auto& el : log.items())
{
const std::string value = el.value().dump(-1, ' ', false, json::error_handler_t::replace);
snprintf(buf, 1024, " %s=%s", el.key().c_str(), value.c_str());
ss << buf;
}
const std::string str = ss.str();
printf("%.*s\n", (int)str.size(), str.data());
fflush(stdout);
}
}
//
@@ -234,6 +257,7 @@ struct llama_server_queue {
std::unique_lock<std::mutex> lock(mutex_tasks);
if (task.id == -1) {
task.id = id++;
LOG_VERBOSE("new task id", {{"new_id", task.id}});
}
queue_tasks.push_back(std::move(task));
condition_tasks.notify_one();
@@ -249,7 +273,9 @@ struct llama_server_queue {
// Get the next id for creating anew task
int get_new_id() {
std::unique_lock<std::mutex> lock(mutex_tasks);
return id++;
int new_id = id++;
LOG_VERBOSE("new task id", {{"new_id", new_id}});
return new_id;
}
// Register function to process a new task
@@ -290,8 +316,7 @@ struct llama_server_queue {
void start_loop() {
running = true;
while (true) {
// new task arrived
LOG_VERBOSE("have new task", {});
LOG_VERBOSE("new task may arrive", {});
{
while (true)
{
@@ -303,7 +328,7 @@ struct llama_server_queue {
task_server task = queue_tasks.front();
queue_tasks.erase(queue_tasks.begin());
lock.unlock();
LOG_VERBOSE("callback_new_task", {});
LOG_VERBOSE("callback_new_task", {{"task_id", task.id}});
callback_new_task(task);
}
LOG_VERBOSE("callback_all_task_finished", {});
@@ -384,11 +409,13 @@ struct llama_server_response {
std::condition_variable condition_results;
void add_waiting_task_id(int task_id) {
LOG_VERBOSE("waiting for task id", {{"task_id", task_id}});
std::unique_lock<std::mutex> lock(mutex_results);
waiting_task_ids.insert(task_id);
}
void remove_waiting_task_id(int task_id) {
LOG_VERBOSE("remove waiting for task id", {{"task_id", task_id}});
std::unique_lock<std::mutex> lock(mutex_results);
waiting_task_ids.erase(task_id);
}
@@ -401,7 +428,6 @@ struct llama_server_response {
condition_results.wait(lock, [&]{
return !queue_results.empty();
});
LOG_VERBOSE("condition_results unblock", {});
for (int i = 0; i < (int) queue_results.size(); i++)
{
@@ -426,22 +452,22 @@ struct llama_server_response {
// Send a new result to a waiting task_id
void send(task_result result) {
std::unique_lock<std::mutex> lock(mutex_results);
LOG_VERBOSE("send new result", {});
LOG_VERBOSE("send new result", {{"task_id", result.id}});
for (auto& task_id : waiting_task_ids) {
// LOG_TEE("waiting task id %i \n", task_id);
// for now, tasks that have associated parent multitasks just get erased once multitask picks up the result
if (result.multitask_id == task_id)
{
LOG_VERBOSE("callback_update_multitask", {});
LOG_VERBOSE("callback_update_multitask", {{"task_id", task_id}});
callback_update_multitask(task_id, result.id, result);
continue;
}
if (result.id == task_id)
{
LOG_VERBOSE("queue_results.push_back", {});
LOG_VERBOSE("queue_results.push_back", {{"task_id", task_id}});
queue_results.push_back(result);
condition_results.notify_one();
condition_results.notify_all();
return;
}
}