baalajimaestro/ollama
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

rm unused infill

Browse source
This commit is contained in:
Michael Yang 2024-05-12 09:21:35 -07:00
parent 3e21799377
commit de781b37c8
1 changed files with 7 additions and 48 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

55
llm/ext_server/server.cpp vendored
View file

@ -140,7 +140,6 @@ struct server_slot {
std::vector<llama_token> cache_tokens; std::vector<llama_token> cache_tokens;
std::vector<completion_token_output> generated_token_probs; std::vector<completion_token_output> generated_token_probs;
bool infill = false;
bool embedding = false; bool embedding = false;
bool has_next_token = true; bool has_next_token = true;
bool truncated = false; bool truncated = false;
@ -187,7 +186,6 @@ struct server_slot {
n_past = 0; n_past = 0;
n_sent_text = 0; n_sent_text = 0;
n_sent_token_probs = 0; n_sent_token_probs = 0;
infill = false;
ga_i = 0; ga_i = 0;
n_past_se = 0; n_past_se = 0;
@ -600,16 +598,6 @@ struct llama_server_context
slot->params.n_predict = slot->n_predict; slot->params.n_predict = slot->n_predict;
} }
// infill
if (data.count("input_prefix") != 0)
{
slot->params.input_prefix = data["input_prefix"];
}
else
{
slot->params.input_prefix = "";
}
if (data.count("input_suffix") != 0) if (data.count("input_suffix") != 0)
{ {
slot->params.input_suffix = data["input_suffix"]; slot->params.input_suffix = data["input_suffix"];
@ -1254,13 +1242,12 @@ struct llama_server_context
queue_results.send(res); queue_results.send(res);
} }
void request_completion(int task_id, json data, bool infill, bool embedding, int multitask_id) void request_completion(int task_id, json data, bool embedding, int multitask_id)
{ {
task_server task; task_server task;
task.id = task_id; task.id = task_id;
task.target_id = 0; task.target_id = 0;
task.data = std::move(data); task.data = std::move(data);
task.infill_mode = infill;
task.embedding_mode = embedding; task.embedding_mode = embedding;
task.type = TASK_TYPE_COMPLETION; task.type = TASK_TYPE_COMPLETION;
task.multitask_id = multitask_id; task.multitask_id = multitask_id;
@ -1406,8 +1393,8 @@ struct llama_server_context
json subtask_data = multiprompt_task.data; json subtask_data = multiprompt_task.data;
subtask_data["prompt"] = subtask_data["prompt"][i]; subtask_data["prompt"] = subtask_data["prompt"][i];
// subtasks inherit everything else (infill mode, embedding mode, etc.) // subtasks inherit everything else (embedding mode, etc.)
request_completion(subtask_ids[i], subtask_data, multiprompt_task.infill_mode, multiprompt_task.embedding_mode, multitask_id); request_completion(subtask_ids[i], subtask_data, multiprompt_task.embedding_mode, multitask_id);
} }
} }
@ -1427,7 +1414,6 @@ struct llama_server_context
slot->reset(); slot->reset();
slot->infill = task.infill_mode;
slot->embedding = task.embedding_mode; slot->embedding = task.embedding_mode;
slot->task_id = task.id; slot->task_id = task.id;
slot->multitask_id = task.multitask_id; slot->multitask_id = task.multitask_id;
@ -1653,8 +1639,7 @@ struct llama_server_context
const bool has_prompt = slot.prompt.is_array() || (slot.prompt.is_string() && !slot.prompt.get<std::string>().empty()) || !slot.images.empty(); const bool has_prompt = slot.prompt.is_array() || (slot.prompt.is_string() && !slot.prompt.get<std::string>().empty()) || !slot.images.empty();
// empty prompt passed -> release the slot and send empty response // empty prompt passed -> release the slot and send empty response
// note: infill mode allows empty prompt if (slot.state == IDLE && slot.command == LOAD_PROMPT && !has_prompt)
if (slot.state == IDLE && slot.command == LOAD_PROMPT && !has_prompt && !slot.infill)
{ {
slot.release(); slot.release();
slot.print_timings(); slot.print_timings();
@ -1671,33 +1656,7 @@ struct llama_server_context
slot.t_start_process_prompt = ggml_time_us(); slot.t_start_process_prompt = ggml_time_us();
slot.t_start_genereration = 0; slot.t_start_genereration = 0;
if (slot.infill) prompt_tokens = tokenize(slot.prompt, system_prompt.empty() && add_bos_token); // add BOS if there isn't system prompt
{
bool suff_rm_leading_spc = true;
if (params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1)
{
params.input_suffix.erase(0, 1);
suff_rm_leading_spc = false;
}
auto prefix_tokens = tokenize(slot.params.input_prefix, false);
auto suffix_tokens = tokenize(slot.params.input_suffix, false);
const int space_token = 29871; // TODO: this should not be hardcoded
if (suff_rm_leading_spc && !suffix_tokens.empty() && suffix_tokens[0] == space_token) {
suffix_tokens.erase(suffix_tokens.begin());
}
prefix_tokens.insert(prefix_tokens.begin(), llama_token_prefix(model));
prefix_tokens.insert(prefix_tokens.begin(), llama_token_bos(model)); // always add BOS
prefix_tokens.insert(prefix_tokens.end(), llama_token_suffix(model));
prefix_tokens.insert(prefix_tokens.end(), suffix_tokens.begin(), suffix_tokens.end());
prefix_tokens.push_back(llama_token_middle(model));
prompt_tokens = prefix_tokens;
}
else
{
prompt_tokens = tokenize(slot.prompt, system_prompt.empty() && add_bos_token); // add BOS if there isn't system prompt
}
slot.n_prompt_tokens = prompt_tokens.size(); slot.n_prompt_tokens = prompt_tokens.size();
@ -3087,7 +3046,7 @@ int main(int argc, char **argv) {
json data = json::parse(req.body); json data = json::parse(req.body);
const int task_id = llama.queue_tasks.get_new_id(); const int task_id = llama.queue_tasks.get_new_id();
llama.queue_results.add_waiting_task_id(task_id); llama.queue_results.add_waiting_task_id(task_id);
llama.request_completion(task_id, data, false, false, -1); llama.request_completion(task_id, data, false, -1);
if (!json_value(data, "stream", false)) { if (!json_value(data, "stream", false)) {
std::string completion_text; std::string completion_text;
task_result result = llama.queue_results.recv(task_id); task_result result = llama.queue_results.recv(task_id);
@ -3181,7 +3140,7 @@ int main(int argc, char **argv) {
// create and queue the task // create and queue the task
const int task_id = llama.queue_tasks.get_new_id(); const int task_id = llama.queue_tasks.get_new_id();
llama.queue_results.add_waiting_task_id(task_id); llama.queue_results.add_waiting_task_id(task_id);
llama.request_completion(task_id, { {"prompt", prompt}, { "n_predict", 0}, {"image_data", image_data} }, false, true, -1); llama.request_completion(task_id, { {"prompt", prompt}, { "n_predict", 0}, {"image_data", image_data} }, true, -1);
// get the result // get the result
task_result result = llama.queue_results.recv(task_id); task_result result = llama.queue_results.recv(task_id);