vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge branch 'master' into xsn/vision_2

Browse source
This commit is contained in:
Xuan Son Nguyen 2025-01-23 23:07:20 +01:00
commit b72d7557f3
10 changed files with 162 additions and 89 deletions
Download patch file Download diff file Expand all files Collapse all files

4
examples/main/README.md
View file

@ -310,9 +310,9 @@ These options help improve the performance and memory usage of the LLaMA models.
### Batch Size
- `-b N, --batch-size N`: Set the batch size for prompt processing (default: `2048`). This large batch size benefits users who have BLAS installed and enabled it during the build. If you don't have BLAS enabled ("BLAS=0"), you can use a smaller number, such as 8, to see the prompt progress as it's evaluated in some situations.
- `-ub N`, `--ubatch-size N`: Physical batch size. This is the maximum number of tokens that may be processed at a time. Increasing this value may improve performance during prompt processing, at the expense of higher memory usage. Default: `512`.
- `-ub N`, `--ubatch-size N`: physical maximum batch size. This is for pipeline parallelization. Default: `512`.
- `-b N`, `--batch-size N`: Logical batch size. Increasing this value above the value of the physical batch size may improve prompt processing performance when using multiple GPUs with pipeline parallelism. Default: `2048`.
### Prompt Caching

22
examples/run/run.cpp
View file

@ -147,7 +147,8 @@ class Opt {
if (handle_option_with_value(argc, argv, i, context_size) == 1) {
return 1;
}
} else if (options_parsing && (strcmp(argv[i], "-n") == 0 || strcmp(argv[i], "--ngl") == 0)) {
} else if (options_parsing &&
(strcmp(argv[i], "-n") == 0 || strcmp(argv[i], "-ngl") == 0 || strcmp(argv[i], "--ngl") == 0)) {
if (handle_option_with_value(argc, argv, i, ngl) == 1) {
return 1;
}
@ -194,7 +195,7 @@ class Opt {
"Options:\n"
" -c, --context-size <value>\n"
" Context size (default: %d)\n"
" -n, --ngl <value>\n"
" -n, -ngl, --ngl <value>\n"
" Number of GPU layers (default: %d)\n"
" --temp <value>\n"
" Temperature (default: %.1f)\n"
@ -634,20 +635,20 @@ class LlamaData {
return path.substr(pos + 1);
}
int remove_proto(std::string & model_) {
const std::string::size_type pos = model_.find("://");
int rm_until_substring(std::string & model_, const std::string & substring) {
const std::string::size_type pos = model_.find(substring);
if (pos == std::string::npos) {
return 1;
}
model_ = model_.substr(pos + 3); // Skip past "://"
model_ = model_.substr(pos + substring.size()); // Skip past the substring
return 0;
}
int resolve_model(std::string & model_) {
int ret = 0;
if (string_starts_with(model_, "file://") || std::filesystem::exists(model_)) {
remove_proto(model_);
rm_until_substring(model_, "://");
return ret;
}
@ -656,13 +657,16 @@ class LlamaData {
const std::vector<std::string> headers = { "--header",
"Accept: application/vnd.docker.distribution.manifest.v2+json" };
if (string_starts_with(model_, "hf://") || string_starts_with(model_, "huggingface://")) {
remove_proto(model_);
rm_until_substring(model_, "://");
ret = huggingface_dl(model_, headers, bn);
} else if (string_starts_with(model_, "hf.co/")) {
rm_until_substring(model_, "hf.co/");
ret = huggingface_dl(model_, headers, bn);
} else if (string_starts_with(model_, "ollama://")) {
remove_proto(model_);
rm_until_substring(model_, "://");
ret = ollama_dl(model_, headers, bn);
} else if (string_starts_with(model_, "https://")) {
download(model_, headers, bn, true);
ret = download(model_, headers, bn, true);
} else {
ret = ollama_dl(model_, headers, bn);
}

46
examples/server/server.cpp
View file

@ -1433,6 +1433,10 @@ struct server_queue {
} else {
queue_tasks.push_back(std::move(task));
}
// if this is cancel task make sure to clean up pending tasks
if (task.type == SERVER_TASK_TYPE_CANCEL) {
cleanup_pending_task(task.id_target);
}
condition_tasks.notify_one();
return task.id;
}
@ -1450,6 +1454,10 @@ struct server_queue {
} else {
queue_tasks.push_back(std::move(task));
}
// if this is cancel task make sure to clean up pending tasks
if (task.type == SERVER_TASK_TYPE_CANCEL) {
cleanup_pending_task(task.id_target);
}
}
condition_tasks.notify_one();
return 0;
@ -1544,6 +1552,20 @@ struct server_queue {
}
}
}
private:
void cleanup_pending_task(int id_task) {
// no need lock because this is called exclusively by post()
auto rm_func = [id_task](const server_task & task) {
return task.id_target == id_task;
};
queue_tasks.erase(
std::remove_if(queue_tasks.begin(), queue_tasks.end(), rm_func),
queue_tasks.end());
queue_tasks_deferred.erase(
std::remove_if(queue_tasks_deferred.begin(), queue_tasks_deferred.end(), rm_func),
queue_tasks_deferred.end());
}
};
struct server_response {
@ -1579,6 +1601,12 @@ struct server_response {
std::unique_lock<std::mutex> lock(mutex_results);
waiting_task_ids.erase(id_task);
// make sure to clean up all pending results
queue_results.erase(
std::remove_if(queue_results.begin(), queue_results.end(), [id_task](const server_task_result_ptr & res) {
return res->id == id_task;
}),
queue_results.end());
}
void remove_waiting_task_ids(const std::unordered_set<int> & id_tasks) {
@ -1598,7 +1626,7 @@ struct server_response {
return !queue_results.empty();
});
for (int i = 0; i < (int) queue_results.size(); i++) {
for (size_t i = 0; i < queue_results.size(); i++) {
if (id_tasks.find(queue_results[i]->id) != id_tasks.end()) {
server_task_result_ptr res = std::move(queue_results[i]);
queue_results.erase(queue_results.begin() + i);
@ -1615,12 +1643,6 @@ struct server_response {
server_task_result_ptr recv_with_timeout(const std::unordered_set<int> & id_tasks, int timeout) {
while (true) {
std::unique_lock<std::mutex> lock(mutex_results);
bool cr_res = condition_results.wait_for(lock, std::chrono::seconds(timeout), [&]{
return !queue_results.empty();
});
if (!cr_res) {
return nullptr;
}
for (int i = 0; i < (int) queue_results.size(); i++) {
if (id_tasks.find(queue_results[i]->id) != id_tasks.end()) {
@ -1629,6 +1651,11 @@ struct server_response {
return res;
}
}
std::cv_status cr_res = condition_results.wait_for(lock, std::chrono::seconds(timeout));
if (cr_res == std::cv_status::timeout) {
return nullptr;
}
}
// should never reach here
@ -1772,6 +1799,9 @@ struct server_context {
// force F16 KV cache for the draft model for extra performance
cparams_dft.type_k = GGML_TYPE_F16;
cparams_dft.type_v = GGML_TYPE_F16;
// the context is not needed - we will create one for each slot
llama_init_dft.context.reset();
}
chat_templates = common_chat_templates_from_model(model, params_base.chat_template);
@ -2373,8 +2403,8 @@ struct server_context {
server_task task(SERVER_TASK_TYPE_CANCEL);
task.id_target = id_task;
cancel_tasks.push_back(task);
queue_results.remove_waiting_task_id(id_task);
cancel_tasks.push_back(task);
}
// push to beginning of the queue, so it has highest priority
queue_tasks.post(cancel_tasks, true);