Merge pull request #4380 from ollama/mxyng/tokenize

use tokenize/detokenize
This commit is contained in:
Michael Yang 2024-05-29 12:00:59 -07:00 committed by GitHub
commit 32cb1960c1
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 66 additions and 242 deletions

View file

@ -140,7 +140,6 @@ struct server_slot {
std::vector<llama_token> cache_tokens;
std::vector<completion_token_output> generated_token_probs;
bool infill = false;
bool embedding = false;
bool has_next_token = true;
bool truncated = false;
@ -187,7 +186,6 @@ struct server_slot {
n_past = 0;
n_sent_text = 0;
n_sent_token_probs = 0;
infill = false;
ga_i = 0;
n_past_se = 0;
@ -600,16 +598,6 @@ struct llama_server_context
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)
{
slot->params.input_suffix = data["input_suffix"];
@ -897,15 +885,6 @@ struct llama_server_context
system_need_update = true;
}
void system_prompt_process(const json &sys_props) {
system_prompt = sys_props.value("prompt", "");
name_user = sys_props.value("anti_prompt", "");
name_assistant = sys_props.value("assistant_name", "");
system_prompt_notify();
}
static size_t find_stopping_strings(const std::string &text, const size_t last_token_size,
const stop_type type, server_slot &slot)
{
@ -1263,13 +1242,12 @@ struct llama_server_context
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.id = task_id;
task.target_id = 0;
task.data = std::move(data);
task.infill_mode = infill;
task.embedding_mode = embedding;
task.type = TASK_TYPE_COMPLETION;
task.multitask_id = multitask_id;
@ -1415,8 +1393,8 @@ struct llama_server_context
json subtask_data = multiprompt_task.data;
subtask_data["prompt"] = subtask_data["prompt"][i];
// subtasks inherit everything else (infill mode, embedding mode, etc.)
request_completion(subtask_ids[i], subtask_data, multiprompt_task.infill_mode, multiprompt_task.embedding_mode, multitask_id);
// subtasks inherit everything else (embedding mode, etc.)
request_completion(subtask_ids[i], subtask_data, multiprompt_task.embedding_mode, multitask_id);
}
}
@ -1434,26 +1412,8 @@ struct llama_server_context
break;
}
if (task.data.contains("system_prompt"))
{
if (!all_slots_are_idle) {
send_error(task, "system prompt can only be updated when all slots are idle");
break;
}
system_prompt_process(task.data["system_prompt"]);
// reset cache_tokens for all slots
for (server_slot &slot : slots)
{
slot.cache_tokens.clear();
slot.n_past = 0;
slot.n_past_se = 0;
}
}
slot->reset();
slot->infill = task.infill_mode;
slot->embedding = task.embedding_mode;
slot->task_id = task.id;
slot->multitask_id = task.multitask_id;
@ -1679,8 +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();
// 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 && !slot.infill)
if (slot.state == IDLE && slot.command == LOAD_PROMPT && !has_prompt)
{
slot.release();
slot.print_timings();
@ -1697,33 +1656,7 @@ struct llama_server_context
slot.t_start_process_prompt = ggml_time_us();
slot.t_start_genereration = 0;
if (slot.infill)
{
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();
@ -2130,8 +2063,7 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
printf("\n");
}
static void server_params_parse(int argc, char **argv, server_params &sparams,
gpt_params &params, llama_server_context& llama)
static void server_params_parse(int argc, char **argv, server_params &sparams, gpt_params &params)
{
gpt_params default_params;
server_params default_sparams;
@ -2546,27 +2478,6 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
}
params.n_predict = std::stoi(argv[i]);
}
else if (arg == "-spf" || arg == "--system-prompt-file")
{
if (++i >= argc)
{
invalid_param = true;
break;
}
std::ifstream file(argv[i]);
if (!file) {
fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
invalid_param = true;
break;
}
std::string systm_content;
std::copy(
std::istreambuf_iterator<char>(file),
std::istreambuf_iterator<char>(),
std::back_inserter(systm_content)
);
llama.system_prompt_process(json::parse(systm_content));
}
else if (arg == "-ctk" || arg == "--cache-type-k") {
params.cache_type_k = argv[++i];
}
@ -2714,21 +2625,6 @@ static json format_partial_response(
return res;
}
static json format_tokenizer_response(const std::vector<llama_token> &tokens)
{
return json {
{"tokens", tokens}
};
}
static json format_detokenized_response(std::string 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
@ -2818,7 +2714,7 @@ int main(int argc, char **argv) {
// struct that contains llama context and inference
llama_server_context llama;
server_params_parse(argc, argv, sparams, params, llama);
server_params_parse(argc, argv, sparams, params);
if (params.model_alias == "unknown")
{
@ -3150,7 +3046,7 @@ int main(int argc, char **argv) {
json data = json::parse(req.body);
const int task_id = llama.queue_tasks.get_new_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)) {
std::string completion_text;
task_result result = llama.queue_results.recv(task_id);
@ -3218,34 +3114,6 @@ int main(int argc, char **argv) {
}
});
svr.Post("/tokenize", [&llama](const httplib::Request &req, httplib::Response &res)
{
res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
const json body = json::parse(req.body);
std::vector<llama_token> tokens;
if (body.count("content") != 0)
{
tokens = llama.tokenize(body["content"], false);
}
const json data = format_tokenizer_response(tokens);
return res.set_content(data.dump(), "application/json; charset=utf-8");
});
svr.Post("/detokenize", [&llama](const httplib::Request &req, httplib::Response &res)
{
res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
const json body = json::parse(req.body);
std::string content;
if (body.count("tokens") != 0)
{
const std::vector<llama_token> tokens = body["tokens"];
content = tokens_to_str(llama.ctx, tokens.cbegin(), tokens.cend());
}
const json data = format_detokenized_response(content);
return res.set_content(data.dump(), "application/json; charset=utf-8");
});
svr.Post("/embedding", [&llama](const httplib::Request &req, httplib::Response &res)
{
res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
@ -3272,7 +3140,7 @@ int main(int argc, char **argv) {
// create and queue the task
const int task_id = llama.queue_tasks.get_new_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
task_result result = llama.queue_results.recv(task_id);

View file

@ -12,6 +12,7 @@ package llm
import "C"
import (
"fmt"
"strings"
"unsafe"
)
@ -37,3 +38,47 @@ func Quantize(infile, outfile string, ftype fileType) error {
return nil
}
type llamaModel struct {
m *C.struct_llama_model
}
func newLlamaModel(p string) *llamaModel {
cs := C.CString(p)
defer C.free(unsafe.Pointer(cs))
return &llamaModel{
C.llama_load_model_from_file(
cs,
C.llama_model_default_params(),
),
}
}
func (llm *llamaModel) Close() {
C.llama_free_model(llm.m)
}
func (llm *llamaModel) Tokenize(s string) []int {
cs := C.CString(s)
defer C.free(unsafe.Pointer(cs))
tokens := make([]int, len(s)+2)
if n := C.llama_tokenize(llm.m, cs, C.int(len(s)), (*C.llama_token)(unsafe.Pointer(&tokens[0])), C.int(len(s)+2), false, true); n > 0 {
return tokens[:n]
}
return nil
}
func (llm *llamaModel) Detokenize(i32s []int) string {
var sb strings.Builder
for _, i32 := range i32s {
c := make([]byte, 512)
if n := C.llama_token_to_piece(llm.m, C.llama_token(i32), (*C.char)(unsafe.Pointer(&c[0])), C.int(len(c)), false); n > 0 {
sb.WriteString(unsafe.String(&c[0], n))
}
}
return sb.String()
}

View file

@ -57,6 +57,8 @@ type llmServer struct {
loadDuration time.Duration // Record how long it took the model to load
loadProgress float32
*llamaModel
sem *semaphore.Weighted
}
@ -306,6 +308,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
totalLayers: ggml.KV().BlockCount() + 1,
gpuCount: gpuCount,
done: make(chan error, 1),
llamaModel: newLlamaModel(model),
}
s.cmd.Env = os.Environ()
@ -843,12 +846,12 @@ func (s *llmServer) Embedding(ctx context.Context, prompt string) ([]float64, er
return nil, fmt.Errorf("unexpected server status: %s", status.ToString())
}
data, err := json.Marshal(TokenizeRequest{Content: prompt})
if err != nil {
var b bytes.Buffer
if err := json.NewEncoder(&b).Encode(EmbeddingRequest{Content: prompt}); err != nil {
return nil, fmt.Errorf("error marshaling embed data: %w", err)
}
req, err := http.NewRequestWithContext(ctx, http.MethodPost, fmt.Sprintf("http://127.0.0.1:%d/embedding", s.port), bytes.NewBuffer(data))
req, err := http.NewRequestWithContext(ctx, http.MethodPost, fmt.Sprintf("http://127.0.0.1:%d/embedding", s.port), &b)
if err != nil {
return nil, fmt.Errorf("error creating embed request: %w", err)
}
@ -878,108 +881,12 @@ func (s *llmServer) Embedding(ctx context.Context, prompt string) ([]float64, er
return embedding.Embedding, nil
}
type TokenizeRequest struct {
Content string `json:"content"`
}
type TokenizeResponse struct {
Tokens []int `json:"tokens"`
}
func (s *llmServer) Tokenize(ctx context.Context, content string) ([]int, error) {
// Make sure the server is ready
status, err := s.getServerStatus(ctx)
if err != nil {
return nil, err
} else if status != ServerStatusReady && status != ServerStatusNoSlotsAvailable {
return nil, fmt.Errorf("unexpected server status: %s", status.ToString())
}
data, err := json.Marshal(TokenizeRequest{Content: content})
if err != nil {
return nil, fmt.Errorf("marshaling encode data: %w", err)
}
req, err := http.NewRequestWithContext(ctx, http.MethodPost, fmt.Sprintf("http://127.0.0.1:%d/tokenize", s.port), bytes.NewBuffer(data))
if err != nil {
return nil, fmt.Errorf("encode request: %w", err)
}
req.Header.Set("Content-Type", "application/json")
resp, err := http.DefaultClient.Do(req)
if err != nil {
return nil, fmt.Errorf("do encode request: %w", err)
}
defer resp.Body.Close()
body, err := io.ReadAll(resp.Body)
if err != nil {
return nil, fmt.Errorf("read encode request: %w", err)
}
if resp.StatusCode >= 400 {
log.Printf("llm encode error: %s", body)
return nil, fmt.Errorf("%s", body)
}
var encoded TokenizeResponse
if err := json.Unmarshal(body, &encoded); err != nil {
return nil, fmt.Errorf("unmarshal encode response: %w", err)
}
return encoded.Tokens, nil
}
type DetokenizeRequest struct {
Tokens []int `json:"tokens"`
}
type DetokenizeResponse struct {
Content string `json:"content"`
return s.llamaModel.Tokenize(content), nil
}
func (s *llmServer) Detokenize(ctx context.Context, tokens []int) (string, error) {
// Make sure the server is ready
status, err := s.getServerStatus(ctx)
if err != nil {
return "", err
} else if status != ServerStatusReady && status != ServerStatusNoSlotsAvailable {
return "", fmt.Errorf("unexpected server status: %s", status.ToString())
}
data, err := json.Marshal(DetokenizeRequest{Tokens: tokens})
if err != nil {
return "", fmt.Errorf("marshaling decode data: %w", err)
}
req, err := http.NewRequestWithContext(ctx, http.MethodPost, fmt.Sprintf("http://127.0.0.1:%d/detokenize", s.port), bytes.NewBuffer(data))
if err != nil {
return "", fmt.Errorf("decode request: %w", err)
}
req.Header.Set("Content-Type", "application/json")
resp, err := http.DefaultClient.Do(req)
if err != nil {
return "", fmt.Errorf("do decode request: %w", err)
}
defer resp.Body.Close()
body, err := io.ReadAll(resp.Body)
if err != nil {
return "", fmt.Errorf("read decode request: %w", err)
}
if resp.StatusCode >= 400 {
log.Printf("llm decode error: %s", body)
return "", fmt.Errorf("%s", body)
}
var decoded DetokenizeResponse
if err := json.Unmarshal(body, &decoded); err != nil {
return "", fmt.Errorf("unmarshal encode response: %w", err)
}
return decoded.Content, nil
return s.llamaModel.Detokenize(tokens), nil
}
func (s *llmServer) Close() error {
@ -997,6 +904,10 @@ func (s *llmServer) Close() error {
slog.Debug("llama server stopped")
}
if s.llamaModel != nil {
s.llamaModel.Close()
}
return nil
}