From 5872e4f4da577d0dff43346f4db3d965cb61d437 Mon Sep 17 00:00:00 2001 From: "Wile E. Coyote" Date: Sun, 22 Oct 2023 21:45:30 -0400 Subject: [PATCH] server support for system, prefix, and suffix prompts with special tokens --- examples/server/README.md | 26 + examples/server/server.cpp | 3473 +++++++++++++++++++----------------- 2 files changed, 1813 insertions(+), 1686 deletions(-) diff --git a/examples/server/README.md b/examples/server/README.md index 715007735..138b87f8c 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -200,6 +200,24 @@ node index.js `system_prompt`: Change the system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime) + `system' : Set the system prompt added before text prompt (arrays). It is independent of system_prompt above + and should not be used together with it. + + `input_prefix`: Set the prefix added to input text prompt lines. + + `input_suffix`: Set the suffix added to input text prompt lines. + + The system, input_prefix, and input_suffix are tokenized with special + tokens required by some models to work correctly. Using these three + prompts enables the server API to support a full externally accumulated + chat history toggling between user inputs and generated outputs line by + line with the desired system header, input_prefix, and input_suffix to + delineate user and genrated lines, without relying on any context memory + in the server. In order for this to work right, input prompts must + not have any hard lfs so the prompt array toggles between user input + and generated output every line. Hard lfs in input prompts need to + be replaced with ascii \n sequence or space. + - **POST** `/tokenize`: Tokenize a given text. *Options:* @@ -208,6 +226,14 @@ node index.js Note that the special `BOS` token is not added in front of the text and also a space character is not inserted automatically as it is for `/completion`. +- **POST** `/tokenizes`: Tokenize a given text with special tokens. + + *Options:* + + `content`: Set the text to tokenize with special tokens. + + Note that the special `BOS` token is not added in front of the text and also a space character is not inserted automatically as it is for `/completion`. + - **POST** `/detokenize`: Convert tokens to text. *Options:* diff --git a/examples/server/server.cpp b/examples/server/server.cpp index c3279dbc9..fc36abbeb 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -49,10 +49,10 @@ static bool server_verbose = false; #define LOG_VERBOSE(MSG, ...) \ do \ { \ - if (server_verbose) \ - { \ - server_log("VERBOSE", __func__, __LINE__, MSG, __VA_ARGS__); \ - } \ + if (server_verbose) \ + { \ + server_log("VERBOSE", __func__, __LINE__, MSG, __VA_ARGS__); \ + } \ } while (0) #endif @@ -65,9 +65,9 @@ static bool server_verbose = false; // static const std::string base64_chars = - "ABCDEFGHIJKLMNOPQRSTUVWXYZ" - "abcdefghijklmnopqrstuvwxyz" - "0123456789+/"; + "ABCDEFGHIJKLMNOPQRSTUVWXYZ" + "abcdefghijklmnopqrstuvwxyz" + "0123456789+/"; static inline bool is_base64(uint8_t c) { @@ -89,46 +89,46 @@ static std::vector base64_decode(std::string const &encoded_string) while (in_len-- && (encoded_string[in_] != '=') && is_base64(encoded_string[in_])) { - char_array_4[i++] = encoded_string[in_]; in_++; - if (i == 4) - { - for (i = 0; i <4; i++) - { - char_array_4[i] = base64_chars.find(char_array_4[i]); - } + char_array_4[i++] = encoded_string[in_]; in_++; + if (i == 4) + { + for (i = 0; i <4; i++) + { + char_array_4[i] = base64_chars.find(char_array_4[i]); + } - char_array_3[0] = ((char_array_4[0] ) << 2) + ((char_array_4[1] & 0x30) >> 4); - char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2); - char_array_3[2] = ((char_array_4[2] & 0x3) << 6) + char_array_4[3]; + char_array_3[0] = ((char_array_4[0] ) << 2) + ((char_array_4[1] & 0x30) >> 4); + char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2); + char_array_3[2] = ((char_array_4[2] & 0x3) << 6) + char_array_4[3]; - for (i = 0; (i < 3); i++) - { - ret.push_back(char_array_3[i]); - } - i = 0; - } + for (i = 0; (i < 3); i++) + { + ret.push_back(char_array_3[i]); + } + i = 0; + } } if (i) { - for (j = i; j <4; j++) - { - char_array_4[j] = 0; - } + for (j = i; j <4; j++) + { + char_array_4[j] = 0; + } - for (j = 0; j <4; j++) - { - char_array_4[j] = base64_chars.find(char_array_4[j]); - } + for (j = 0; j <4; j++) + { + char_array_4[j] = base64_chars.find(char_array_4[j]); + } - char_array_3[0] = ((char_array_4[0] ) << 2) + ((char_array_4[1] & 0x30) >> 4); - char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2); - char_array_3[2] = ((char_array_4[2] & 0x3) << 6) + char_array_4[3]; + char_array_3[0] = ((char_array_4[0] ) << 2) + ((char_array_4[1] & 0x30) >> 4); + char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2); + char_array_3[2] = ((char_array_4[2] & 0x3) << 6) + char_array_4[3]; - for (j = 0; (j < i - 1); j++) - { - ret.push_back(char_array_3[j]); - } + for (j = 0; (j < i - 1); j++) + { + ret.push_back(char_array_3[j]); + } } return ret; @@ -183,6 +183,7 @@ struct slot_params std::vector antiprompt; + json system; json input_prefix; json input_suffix; }; @@ -205,8 +206,8 @@ struct completion_token_output { struct token_prob { - llama_token tok; - float prob; + llama_token tok; + float prob; }; std::vector probs; @@ -232,26 +233,26 @@ enum stop_type static bool ends_with(const std::string &str, const std::string &suffix) { return str.size() >= suffix.size() && - 0 == str.compare(str.size() - suffix.size(), suffix.size(), suffix); + 0 == str.compare(str.size() - suffix.size(), suffix.size(), suffix); } static size_t find_partial_stop_string(const std::string &stop, - const std::string &text) + const std::string &text) { if (!text.empty() && !stop.empty()) { - const char text_last_char = text.back(); - for (int64_t char_index = stop.size() - 1; char_index >= 0; char_index--) - { - if (stop[char_index] == text_last_char) - { - const std::string current_partial = stop.substr(0, char_index + 1); - if (ends_with(text, current_partial)) - { - return text.size() - char_index - 1; - } - } - } + const char text_last_char = text.back(); + for (int64_t char_index = stop.size() - 1; char_index >= 0; char_index--) + { + if (stop[char_index] == text_last_char) + { + const std::string current_partial = stop.substr(0, char_index + 1); + if (ends_with(text, current_partial)) + { + return text.size() - char_index - 1; + } + } + } } return std::string::npos; } @@ -263,26 +264,26 @@ static std::string tokens_to_str(llama_context *ctx, Iter begin, Iter end) std::string ret; for (; begin != end; ++begin) { - ret += llama_token_to_piece(ctx, *begin); + ret += llama_token_to_piece(ctx, *begin); } return ret; } static void server_log(const char *level, const char *function, int line, - const char *message, const nlohmann::ordered_json &extra) + const char *message, const nlohmann::ordered_json &extra) { nlohmann::ordered_json log { - {"timestamp", time(nullptr)}, - {"level", level}, - {"function", function}, - {"line", line}, - {"message", message}, + {"timestamp", time(nullptr)}, + {"level", level}, + {"function", function}, + {"line", line}, + {"message", message}, }; if (!extra.empty()) { - log.merge_patch(extra); + log.merge_patch(extra); } const std::string str = log.dump(-1, ' ', false, json::error_handler_t::replace); @@ -298,10 +299,10 @@ static std::string tokens_to_output_formatted_string(const llama_context *ctx, c // (size > 1 meaning it's already a known token) if (out.size() == 1 && (out[0] & 0x80) == 0x80) { - std::stringstream ss; - ss << std::hex << (out[0] & 0xff); - std::string res(ss.str()); - out = "byte: \\x" + res; + std::stringstream ss; + ss << std::hex << (out[0] & 0xff); + std::string res(ss.str()); + out = "byte: \\x" + res; } return out; } @@ -312,21 +313,21 @@ static json probs_vector_to_json(const llama_context *ctx, const std::vector 0 || n_remaining == -1; // no budget || limitless + n_remaining = -1; + if(params.n_predict != -1) + { + n_remaining = params.n_predict - n_decoded; + } + else if (global_params.n_predict != -1) + { + n_remaining = global_params.n_predict - n_decoded; + } + return n_remaining > 0 || n_remaining == -1; // no budget || limitless } bool available() const { - return state == IDLE && command == NONE; + return state == IDLE && command == NONE; } bool is_processing() const { - return (state == IDLE && command == LOAD_PROMPT) || state == PROCESSING; + return (state == IDLE && command == LOAD_PROMPT) || state == PROCESSING; } void add_token_string(const completion_token_output &token) { - if (command == RELEASE) - { - return; - } - cache_tokens.push_back(token.tok); - generated_token_probs.push_back(token); + if (command == RELEASE) + { + return; + } + cache_tokens.push_back(token.tok); + generated_token_probs.push_back(token); } void release() { - if (state == PROCESSING) - { - t_token_generation = (ggml_time_us() - t_start_genereration) / 1e3; - command = RELEASE; - } + if (state == PROCESSING) + { + t_token_generation = (ggml_time_us() - t_start_genereration) / 1e3; + command = RELEASE; + } } json get_formated_timings() { - return json - { - {"prompt_n", num_prompt_tokens_processed}, - {"prompt_ms", t_prompt_processing}, - {"prompt_per_token_ms", t_prompt_processing / num_prompt_tokens_processed}, - {"prompt_per_second", 1e3 / t_prompt_processing * num_prompt_tokens_processed}, + return json + { + {"prompt_n", num_prompt_tokens_processed}, + {"prompt_ms", t_prompt_processing}, + {"prompt_per_token_ms", t_prompt_processing / num_prompt_tokens_processed}, + {"prompt_per_second", 1e3 / t_prompt_processing * num_prompt_tokens_processed}, - {"predicted_n", n_decoded}, - {"predicted_ms", t_token_generation}, - {"predicted_per_token_ms", t_token_generation / n_decoded}, - {"predicted_per_second", 1e3 / t_token_generation * n_decoded}, - }; + {"predicted_n", n_decoded}, + {"predicted_ms", t_token_generation}, + {"predicted_per_token_ms", t_token_generation / n_decoded}, + {"predicted_per_second", 1e3 / t_token_generation * n_decoded}, + }; } void print_timings() { - 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); + 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); } }; @@ -523,1245 +524,1333 @@ struct llama_server_context ~llama_server_context() { - if (ctx) - { - llama_free(ctx); - ctx = nullptr; - } - if (model) - { - llama_free_model(model); - model = nullptr; - } + if (ctx) + { + llama_free(ctx); + ctx = nullptr; + } + if (model) + { + llama_free_model(model); + model = nullptr; + } } bool load_model(const gpt_params ¶ms_) { - params = params_; - if (!params.mmproj.empty()) { - multimodal = true; - LOG_TEE("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}}); - return false; - } + params = params_; + if (!params.mmproj.empty()) { + multimodal = true; + LOG_TEE("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}}); + return false; + } - if (params.n_ctx < 2048) { // request larger context for the image embedding - params.n_ctx = 2048; - } - } + if (params.n_ctx < 2048) { // request larger context for the image embedding + params.n_ctx = 2048; + } + } - std::tie(model, ctx) = llama_init_from_gpt_params(params); - if (model == nullptr) - { - LOG_ERROR("unable to load model", {{"model", params.model}}); - return false; - } + std::tie(model, ctx) = llama_init_from_gpt_params(params); + if (model == nullptr) + { + LOG_ERROR("unable to load model", {{"model", params.model}}); + return false; + } - if (multimodal) { - const int n_embd_clip = clip_n_mmproj_embd(clp_ctx); - const int n_embd_llm = llama_n_embd(model); - if (n_embd_clip != n_embd_llm) { - LOG_TEE("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_embd_clip, n_embd_llm); - llama_free(ctx); - llama_free_model(model); - return false; - } - } + if (multimodal) { + const int n_embd_clip = clip_n_mmproj_embd(clp_ctx); + const int n_embd_llm = llama_n_embd(model); + if (n_embd_clip != n_embd_llm) { + LOG_TEE("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_embd_clip, n_embd_llm); + llama_free(ctx); + llama_free_model(model); + return false; + } + } - n_ctx = llama_n_ctx(ctx); + n_ctx = llama_n_ctx(ctx); - return true; + return true; } void initialize() { - id_gen = 0; + id_gen = 0; - // create slots - all_slots_are_idle = true; + // create slots + all_slots_are_idle = true; - const int32_t n_ctx_slot = n_ctx / params.n_parallel; + const int32_t n_ctx_slot = n_ctx / params.n_parallel; - LOG_TEE("Available slots:\n"); - for (int i = 0; i < params.n_parallel; i++) - { - llama_client_slot slot; + LOG_TEE("Available slots:\n"); + for (int i = 0; i < params.n_parallel; i++) + { + llama_client_slot slot; - slot.id = i; - slot.n_ctx = n_ctx_slot; - slot.reset(); + slot.id = i; + slot.n_ctx = n_ctx_slot; + slot.reset(); - LOG_TEE(" -> Slot %i - max context: %i\n", slot.id, n_ctx_slot); - slots.push_back(slot); - } + LOG_TEE(" -> Slot %i - max context: %i\n", slot.id, n_ctx_slot); + slots.push_back(slot); + } - batch = llama_batch_init(n_ctx, 0, params.n_parallel); + batch = llama_batch_init(n_ctx, 0, params.n_parallel); - // empty system prompt - system_prompt = ""; - system_tokens.clear(); + // empty system prompt + system_prompt = ""; + system_tokens.clear(); } - std::vector tokenize(const json & json_prompt, bool add_bos) const + std::vector tokenize(const json & json_prompt, + bool add_bos, bool special=false, + const json & json_system=NULL, + const json & json_prefix=NULL, + const json & json_suffix=NULL) const { - // If `add_bos` is true, we only add BOS, when json_prompt is a string, - // or the first element of the json_prompt array is a string. - std::vector prompt_tokens; + // If `add_bos` is true, we only add BOS, when json_prompt is a string, + // or the first element of the json_prompt array is a string. + std::vector prompt_tokens; - if (json_prompt.is_array()) - { - bool first = true; - for (const auto& p : json_prompt) - { - if (p.is_string()) - { - auto s = p.template get(); - std::vector p; - if (first) - { - p = ::llama_tokenize(ctx, s, add_bos); - first = false; - } - else - { - p = ::llama_tokenize(ctx, s, false); - } - prompt_tokens.insert(prompt_tokens.end(), p.begin(), p.end()); - } - else - { - if (first) - { - first = false; - } - prompt_tokens.push_back(p.template get()); - } - } - } - else - { - auto s = json_prompt.template get(); - prompt_tokens = ::llama_tokenize(ctx, s, add_bos); - } + // to support short term learning from chat context, + // first line in array is query, next line + // is generated text, next line is next query ... + bool user_input = true; - return prompt_tokens; + // don't add sys/prefix/suffix if not a normal tokenize + bool add_params = add_bos; + + std::string params_system="",params_input_prefix="",params_input_suffix=""; + if (json_system != NULL) + if (json_system.is_string()) + params_system = json_system.template get(); + if (json_prefix != NULL) + if (json_prefix.is_string()) + params_input_prefix = json_prefix.template get(); + if (json_suffix != NULL) + if (json_suffix.is_string()) + params_input_suffix = json_suffix.template get(); + + if (add_params && (params_system.size() > 1)) { + // add the system prompt before the conversation input + LOG("system: '%s'\n", params_system.c_str()); + std::vector system; + system = ::llama_tokenize(ctx,params_system,false,true); + prompt_tokens.insert(prompt_tokens.end(),system.begin(), system.end()); + LOG("prompt: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, system).c_str()); + } + + if (json_prompt.is_array()) + { + for (const auto& p : json_prompt) + { + if (p.is_string()) + { + auto s = p.template get(); + std::vector p; + + if (add_params && user_input && (params_input_prefix.size() > 1)) { + LOG("input prefix: '%s'\n", params_input_prefix.c_str()); + std::vector line_pfx; + line_pfx = ::llama_tokenize(ctx,params_input_prefix,add_bos,true); + prompt_tokens.insert(prompt_tokens.end(),line_pfx.begin(), line_pfx.end()); + LOG("prefix tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_pfx).c_str()); + + // bos has been added + add_bos = false; + } + + p = ::llama_tokenize(ctx, s, add_bos, special); + // bos has been added + add_bos = false; + + LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, p).c_str()); + prompt_tokens.insert(prompt_tokens.end(), p.begin(), p.end()); + + if (add_params && user_input && (params_input_suffix.size() > 1)) { + LOG("input suffix: '%s'\n", params_input_suffix.c_str()); + std::vector line_sfx; + line_sfx = ::llama_tokenize(ctx,params_input_suffix,false,true); + LOG("suffix tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_sfx).c_str()); + prompt_tokens.insert(prompt_tokens.end(),line_sfx.begin(), line_sfx.end()); + } + + user_input = !user_input; + } + else + { + prompt_tokens.push_back(p.template get()); + } + } + } + else + { + auto s = json_prompt.template get(); + + std::vector p; + + if (add_params && (params_input_prefix.size() > 1)) { + LOG("input prefix: '%s'\n", params_input_prefix.c_str()); + std::vector line_pfx; + line_pfx = ::llama_tokenize(ctx,params_input_prefix,add_bos,true); + prompt_tokens.insert(prompt_tokens.end(),line_pfx.begin(), line_pfx.end()); + LOG("prefix tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_pfx).c_str()); + // bos has been added + add_bos = false; + } + + p = ::llama_tokenize(ctx, s, add_bos, special); + add_bos = false; + + LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, p).c_str()); + prompt_tokens.insert(prompt_tokens.end(), p.begin(), p.end()); + + // Add the suffix if defined + if (add_params && (params_input_suffix.size() > 1)) { + LOG("input suffix: '%s'\n", params_input_suffix.c_str()); + std::vector line_sfx; + line_sfx = ::llama_tokenize(ctx,params_input_suffix,false,true); + LOG("suffix tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_sfx).c_str()); + prompt_tokens.insert(prompt_tokens.end(),line_sfx.begin(), line_sfx.end()); + } + } + + return prompt_tokens; } llama_client_slot* get_slot(int id) { - int64_t t_last = ggml_time_us(); - llama_client_slot *last_used = nullptr; + int64_t t_last = ggml_time_us(); + llama_client_slot *last_used = nullptr; - for (llama_client_slot & slot : slots) - { - if (slot.id == id && slot.available()) - { - return &slot; - } + for (llama_client_slot & slot : slots) + { + if (slot.id == id && slot.available()) + { + return &slot; + } - if (slot.available() && slot.t_last_used < t_last) - { - last_used = &slot; - t_last = slot.t_last_used; - } - } + if (slot.available() && slot.t_last_used < t_last) + { + last_used = &slot; + t_last = slot.t_last_used; + } + } - return last_used; + return last_used; } bool launch_slot_with_data(llama_client_slot* &slot, json data) { - slot_params default_params; - llama_sampling_params default_sparams; + slot_params default_params; + llama_sampling_params default_sparams; - slot->params.stream = json_value(data, "stream", false); - slot->params.cache_prompt = json_value(data, "cache_prompt", false); - slot->params.n_predict = json_value(data, "n_predict", default_params.n_predict); - slot->sparams.top_k = json_value(data, "top_k", default_sparams.top_k); - slot->sparams.top_p = json_value(data, "top_p", default_sparams.top_p); - slot->sparams.tfs_z = json_value(data, "tfs_z", default_sparams.tfs_z); - slot->sparams.typical_p = json_value(data, "typical_p", default_sparams.typical_p); - slot->sparams.temp = json_value(data, "temperature", default_sparams.temp); - slot->sparams.penalty_last_n = json_value(data, "repeat_last_n", default_sparams.penalty_last_n); - slot->sparams.penalty_repeat = json_value(data, "repeat_penalty", default_sparams.penalty_repeat); - slot->sparams.penalty_freq = json_value(data, "frequency_penalty", default_sparams.penalty_freq); - slot->sparams.penalty_present = json_value(data, "presence_penalty", default_sparams.penalty_present); - slot->sparams.mirostat = json_value(data, "mirostat", default_sparams.mirostat); - slot->sparams.mirostat_tau = json_value(data, "mirostat_tau", default_sparams.mirostat_tau); - slot->sparams.mirostat_eta = json_value(data, "mirostat_eta", default_sparams.mirostat_eta); - slot->sparams.penalize_nl = json_value(data, "penalize_nl", default_sparams.penalize_nl); - slot->params.n_keep = json_value(data, "n_keep", slot->params.n_keep); - slot->params.seed = json_value(data, "seed", default_params.seed); - slot->sparams.grammar = json_value(data, "grammar", default_sparams.grammar); - slot->sparams.n_probs = json_value(data, "n_probs", default_sparams.n_probs); + slot->params.stream = json_value(data, "stream", false); + slot->params.cache_prompt = json_value(data, "cache_prompt", false); + slot->params.n_predict = json_value(data, "n_predict", default_params.n_predict); + slot->sparams.top_k = json_value(data, "top_k", default_sparams.top_k); + slot->sparams.top_p = json_value(data, "top_p", default_sparams.top_p); + slot->sparams.tfs_z = json_value(data, "tfs_z", default_sparams.tfs_z); + slot->sparams.typical_p = json_value(data, "typical_p", default_sparams.typical_p); + slot->sparams.temp = json_value(data, "temperature", default_sparams.temp); + slot->sparams.penalty_last_n = json_value(data, "repeat_last_n", default_sparams.penalty_last_n); + slot->sparams.penalty_repeat = json_value(data, "repeat_penalty", default_sparams.penalty_repeat); + slot->sparams.penalty_freq = json_value(data, "frequency_penalty", default_sparams.penalty_freq); + slot->sparams.penalty_present = json_value(data, "presence_penalty", default_sparams.penalty_present); + slot->sparams.mirostat = json_value(data, "mirostat", default_sparams.mirostat); + slot->sparams.mirostat_tau = json_value(data, "mirostat_tau", default_sparams.mirostat_tau); + slot->sparams.mirostat_eta = json_value(data, "mirostat_eta", default_sparams.mirostat_eta); + slot->sparams.penalize_nl = json_value(data, "penalize_nl", default_sparams.penalize_nl); + slot->params.n_keep = json_value(data, "n_keep", slot->params.n_keep); + slot->params.seed = json_value(data, "seed", default_params.seed); + slot->sparams.grammar = json_value(data, "grammar", default_sparams.grammar); + slot->sparams.n_probs = json_value(data, "n_probs", default_sparams.n_probs); - // infill - if (data.count("input_prefix") != 0) - { - slot->params.input_prefix = data["input_prefix"]; - } - else - { - slot->params.input_prefix = ""; - } + // system prompt + if (data.count("system") != 0) + { + slot->params.system = data["system"]; + } + else + { + slot->params.system = ""; + } - if (data.count("input_suffix") != 0) - { - slot->params.input_suffix = data["input_suffix"]; - } - else - { - slot->params.input_suffix = ""; - } + // infill, prompt prefix/suffix + if (data.count("input_prefix") != 0) + { + slot->params.input_prefix = data["input_prefix"]; + } + else + { + slot->params.input_prefix = ""; + } - if (data.count("prompt") != 0) - { - slot->prompt = data["prompt"]; - } - else - { - slot->prompt = ""; - } + if (data.count("input_suffix") != 0) + { + slot->params.input_suffix = data["input_suffix"]; + } + else + { + slot->params.input_suffix = ""; + } - slot->sparams.logit_bias.clear(); + if (data.count("prompt") != 0) + { + slot->prompt = data["prompt"]; + } + else + { + slot->prompt = ""; + } - if (json_value(data, "ignore_eos", false)) - { - slot->sparams.logit_bias[llama_token_eos(ctx)] = -INFINITY; - } + slot->sparams.logit_bias.clear(); - const auto &logit_bias = data.find("logit_bias"); - if (logit_bias != data.end() && logit_bias->is_array()) - { - const int n_vocab = llama_n_vocab(model); - for (const auto &el : *logit_bias) - { - if (el.is_array() && el.size() == 2 && el[0].is_number_integer()) - { - llama_token tok = el[0].get(); - if (tok >= 0 && tok < n_vocab) - { - if (el[1].is_number()) - { - slot->sparams.logit_bias[tok] = el[1].get(); - } - else if (el[1].is_boolean() && !el[1].get()) - { - slot->sparams.logit_bias[tok] = -INFINITY; - } - } - } - } - } + if (json_value(data, "ignore_eos", false)) + { + slot->sparams.logit_bias[llama_token_eos(ctx)] = -INFINITY; + } - slot->params.antiprompt.clear(); - const auto &stop = data.find("stop"); - if (stop != data.end() && stop->is_array()) - { - for (const auto &word : *stop) - { - if (!word.empty()) - { - slot->params.antiprompt.push_back(word); - } - } - } + const auto &logit_bias = data.find("logit_bias"); + if (logit_bias != data.end() && logit_bias->is_array()) + { + const int n_vocab = llama_n_vocab(model); + for (const auto &el : *logit_bias) + { + if (el.is_array() && el.size() == 2 && el[0].is_number_integer()) + { + llama_token tok = el[0].get(); + if (tok >= 0 && tok < n_vocab) + { + if (el[1].is_number()) + { + slot->sparams.logit_bias[tok] = el[1].get(); + } + else if (el[1].is_boolean() && !el[1].get()) + { + slot->sparams.logit_bias[tok] = -INFINITY; + } + } + } + } + } - if (multimodal) - { - const auto &images_data = data.find("image_data"); - if (images_data != data.end() && images_data->is_array()) - { - for (const auto &img : *images_data) - { - std::string data_b64 = img["data"].get(); - slot_image img_sl; - img_sl.id = img.count("id") != 0 ? img["id"].get() : slot->images.size(); - int width, height, channels; - std::vector image_buffer = base64_decode(data_b64); - data_b64.clear(); - auto data = stbi_load_from_memory(image_buffer.data(), image_buffer.size(), &width, &height, &channels, 3); - if (!data) { - LOG_TEE("slot %i - failed to load image [id: %i]\n", slot->id, img_sl.id); - return false; - } - LOG_TEE("slot %i - image loaded [id: %i] resolution (%i x %i)\n", slot->id, img_sl.id, width, height); - img_sl.img_data.nx = width; - img_sl.img_data.ny = height; - img_sl.img_data.size = width * height * 3; - img_sl.img_data.data = new uint8_t[width * height * 3](); - memcpy(img_sl.img_data.data, data, width * height * 3); - stbi_image_free(data); - img_sl.request_encode_image = true; - slot->images.push_back(img_sl); - } - // process prompt - // example: system prompt [img-102] user [img-103] describe [img-134] -> [{id: 102, prefix: 'system prompt '}, {id: 103, prefix: ' user '}, {id: 134, prefix: ' describe '}]} - if (slot->images.size() > 0 && !slot->prompt.is_array()) - { - std::string prompt = slot->prompt.get(); - size_t pos = 0, begin_prefix = 0; - std::string pattern = "[img-"; - while ((pos = prompt.find(pattern, pos)) != std::string::npos) { - size_t end_prefix = pos; - pos += pattern.length(); - size_t end_pos = prompt.find("]", pos); - if (end_pos != std::string::npos) - { - std::string image_id = prompt.substr(pos, end_pos - pos); - try - { - int img_id = std::stoi(image_id); - bool found = false; - for (slot_image &img : slot->images) - { - if (img.id == img_id) { - found = true; - img.prefix_prompt = prompt.substr(begin_prefix, end_prefix - begin_prefix); - begin_prefix = end_pos + 1; - break; - } - } - if (!found) { - LOG_TEE("ERROR: Image with id: %i, not found.\n", img_id); - slot->images.clear(); - return false; - } - } catch (const std::invalid_argument& e) { - LOG_TEE("Invalid image number id in prompt\n"); - slot->images.clear(); - return false; - } - } - } - slot->prompt = ""; - slot->params.input_suffix = prompt.substr(begin_prefix); - slot->params.cache_prompt = false; // multimodal doesn't support cache prompt - } - } - } + slot->params.antiprompt.clear(); + const auto &stop = data.find("stop"); + if (stop != data.end() && stop->is_array()) + { + for (const auto &word : *stop) + { + if (!word.empty()) + { + slot->params.antiprompt.push_back(word); + } + } + } - if (slot->ctx_sampling != nullptr) - { - llama_sampling_free(slot->ctx_sampling); - } - slot->ctx_sampling = llama_sampling_init(slot->sparams); - slot->command = LOAD_PROMPT; + if (multimodal) + { + const auto &images_data = data.find("image_data"); + if (images_data != data.end() && images_data->is_array()) + { + for (const auto &img : *images_data) + { + std::string data_b64 = img["data"].get(); + slot_image img_sl; + img_sl.id = img.count("id") != 0 ? img["id"].get() : slot->images.size(); + int width, height, channels; + std::vector image_buffer = base64_decode(data_b64); + data_b64.clear(); + auto data = stbi_load_from_memory(image_buffer.data(), image_buffer.size(), &width, &height, &channels, 3); + if (!data) { + LOG_TEE("slot %i - failed to load image [id: %i]\n", slot->id, img_sl.id); + return false; + } + LOG_TEE("slot %i - image loaded [id: %i] resolution (%i x %i)\n", slot->id, img_sl.id, width, height); + img_sl.img_data.nx = width; + img_sl.img_data.ny = height; + img_sl.img_data.size = width * height * 3; + img_sl.img_data.data = new uint8_t[width * height * 3](); + memcpy(img_sl.img_data.data, data, width * height * 3); + stbi_image_free(data); + img_sl.request_encode_image = true; + slot->images.push_back(img_sl); + } + // process prompt + // example: system prompt [img-102] user [img-103] describe [img-134] -> [{id: 102, prefix: 'system prompt '}, {id: 103, prefix: ' user '}, {id: 134, prefix: ' describe '}]} + if (slot->images.size() > 0 && !slot->prompt.is_array()) + { + std::string prompt = slot->prompt.get(); + size_t pos = 0, begin_prefix = 0; + std::string pattern = "[img-"; + while ((pos = prompt.find(pattern, pos)) != std::string::npos) { + size_t end_prefix = pos; + pos += pattern.length(); + size_t end_pos = prompt.find("]", pos); + if (end_pos != std::string::npos) + { + std::string image_id = prompt.substr(pos, end_pos - pos); + try + { + int img_id = std::stoi(image_id); + bool found = false; + for (slot_image &img : slot->images) + { + if (img.id == img_id) { + found = true; + img.prefix_prompt = prompt.substr(begin_prefix, end_prefix - begin_prefix); + begin_prefix = end_pos + 1; + break; + } + } + if (!found) { + LOG_TEE("ERROR: Image with id: %i, not found.\n", img_id); + slot->images.clear(); + return false; + } + } catch (const std::invalid_argument& e) { + LOG_TEE("Invalid image number id in prompt\n"); + slot->images.clear(); + return false; + } + } + } + slot->prompt = ""; + slot->params.input_suffix = prompt.substr(begin_prefix); + slot->params.cache_prompt = false; // multimodal doesn't support cache prompt + } + } + } - all_slots_are_idle = false; + if (slot->ctx_sampling != nullptr) + { + llama_sampling_free(slot->ctx_sampling); + } + slot->ctx_sampling = llama_sampling_init(slot->sparams); + slot->command = LOAD_PROMPT; - LOG_TEE("slot %i is processing [task id: %i]\n", slot->id, slot->task_id); + all_slots_are_idle = false; - return true; + LOG_TEE("slot %i is processing [task id: %i]\n", slot->id, slot->task_id); + + return true; } void kv_cache_clear() { - // clear the entire KV cache - llama_kv_cache_tokens_rm(ctx, -1, -1); - clean_kv_cache = false; + // clear the entire KV cache + llama_kv_cache_tokens_rm(ctx, -1, -1); + clean_kv_cache = false; } void update_system_prompt() { - system_tokens = ::llama_tokenize(ctx, system_prompt, true); + system_tokens = ::llama_tokenize(ctx, system_prompt, true); - llama_batch_clear(batch); + llama_batch_clear(batch); - kv_cache_clear(); + kv_cache_clear(); - for (int32_t i = 0; i < batch.n_tokens; ++i) - { - llama_batch_add(batch, system_tokens[i], i, { 0 }, false); - } + for (int32_t i = 0; i < batch.n_tokens; ++i) + { + llama_batch_add(batch, system_tokens[i], i, { 0 }, false); + } - if (llama_decode(ctx, batch) != 0) - { - LOG_TEE("%s: llama_decode() failed\n", __func__); - return; - } + if (llama_decode(ctx, batch) != 0) + { + LOG_TEE("%s: llama_decode() failed\n", __func__); + return; + } - // assign the system KV cache to all parallel sequences - for (int32_t i = 1; i < params.n_parallel; ++i) - { - llama_kv_cache_seq_cp(ctx, 0, i, 0, system_tokens.size()); - } + // assign the system KV cache to all parallel sequences + for (int32_t i = 1; i < params.n_parallel; ++i) + { + llama_kv_cache_seq_cp(ctx, 0, i, 0, system_tokens.size()); + } - LOG_TEE("system prompt updated\n"); - system_need_update = false; + LOG_TEE("system prompt updated\n"); + system_need_update = false; } void notify_system_prompt_changed() { - // release all slots - for (llama_client_slot &slot : slots) - { - slot.release(); - } - wait_all_are_idle(); - all_slots_are_idle = true; + // release all slots + for (llama_client_slot &slot : slots) + { + slot.release(); + } + wait_all_are_idle(); + all_slots_are_idle = true; - // wait until system prompt load - system_need_update = true; - while (system_need_update) - { - std::this_thread::sleep_for(std::chrono::milliseconds(5)); - } - // system prompt loaded, continue + // wait until system prompt load + system_need_update = true; + while (system_need_update) + { + std::this_thread::sleep_for(std::chrono::milliseconds(5)); + } + // system prompt loaded, continue } void process_system_prompt_data(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 = sys_props.value("prompt", ""); + name_user = sys_props.value("anti_prompt", ""); + name_assistant = sys_props.value("assistant_name", ""); - if (slots.size() > 0) - { - notify_system_prompt_changed(); - } - else - { - system_need_update = true; - } + if (slots.size() > 0) + { + notify_system_prompt_changed(); + } + else + { + system_need_update = true; + } } void wait_all_are_idle() { - bool wait = true; - while (wait) - { - wait = false; - for (auto &slot : slots) - { - if (!slot.available()) - { - wait = true; - break; - } - } - } + bool wait = true; + while (wait) + { + wait = false; + for (auto &slot : slots) + { + if (!slot.available()) + { + wait = true; + break; + } + } + } } static size_t find_stopping_strings(const std::string &text, const size_t last_token_size, - const stop_type type, llama_client_slot &slot) + const stop_type type, llama_client_slot &slot) { - size_t stop_pos = std::string::npos; + size_t stop_pos = std::string::npos; - for (const std::string &word : slot.params.antiprompt) - { - size_t pos; - if (type == STOP_FULL) - { - const size_t tmp = word.size() + last_token_size; - const size_t from_pos = text.size() > tmp ? text.size() - tmp : 0; - pos = text.find(word, from_pos); - } - else - { - pos = find_partial_stop_string(word, text); - } - if (pos != std::string::npos && - (stop_pos == std::string::npos || pos < stop_pos)) - { - if (type == STOP_FULL) - { - slot.stopped_word = true; - slot.stopping_word = word; - slot.has_next_token = false; - } - stop_pos = pos; + for (const std::string &word : slot.params.antiprompt) + { + size_t pos; + if (type == STOP_FULL) + { + const size_t tmp = word.size() + last_token_size; + const size_t from_pos = text.size() > tmp ? text.size() - tmp : 0; + pos = text.find(word, from_pos); + } + else + { + pos = find_partial_stop_string(word, text); + } + if (pos != std::string::npos && + (stop_pos == std::string::npos || pos < stop_pos)) + { + if (type == STOP_FULL) + { + slot.stopped_word = true; + slot.stopping_word = word; + slot.has_next_token = false; + } + stop_pos = pos; - } - } + } + } - return stop_pos; + return stop_pos; } bool process_token(completion_token_output &result, llama_client_slot &slot) { - // remember which tokens were sampled - used for repetition penalties during sampling - const std::string token_str = llama_token_to_piece(ctx, result.tok); - slot.sampled = result.tok; + // remember which tokens were sampled - used for repetition penalties during sampling + const std::string token_str = llama_token_to_piece(ctx, result.tok); + slot.sampled = result.tok; - // search stop word and delete it - slot.generated_text += token_str; - slot.has_next_token = true; + // search stop word and delete it + slot.generated_text += token_str; + slot.has_next_token = true; - if (slot.multibyte_pending > 0) - { - slot.multibyte_pending -= token_str.size(); - } - else if (token_str.size() == 1) - { - const char c = token_str[0]; - // 2-byte characters: 110xxxxx 10xxxxxx - if ((c & 0xE0) == 0xC0) - { - slot.multibyte_pending = 1; - // 3-byte characters: 1110xxxx 10xxxxxx 10xxxxxx - } - else if ((c & 0xF0) == 0xE0) - { - slot.multibyte_pending = 2; - // 4-byte characters: 11110xxx 10xxxxxx 10xxxxxx 10xxxxxx - } - else if ((c & 0xF8) == 0xF0) - { - slot.multibyte_pending = 3; - } - else - { - slot.multibyte_pending = 0; - } - } + if (slot.multibyte_pending > 0) + { + slot.multibyte_pending -= token_str.size(); + } + else if (token_str.size() == 1) + { + const char c = token_str[0]; + // 2-byte characters: 110xxxxx 10xxxxxx + if ((c & 0xE0) == 0xC0) + { + slot.multibyte_pending = 1; + // 3-byte characters: 1110xxxx 10xxxxxx 10xxxxxx + } + else if ((c & 0xF0) == 0xE0) + { + slot.multibyte_pending = 2; + // 4-byte characters: 11110xxx 10xxxxxx 10xxxxxx 10xxxxxx + } + else if ((c & 0xF8) == 0xF0) + { + slot.multibyte_pending = 3; + } + else + { + slot.multibyte_pending = 0; + } + } - if (slot.multibyte_pending == 0) - { - size_t pos = std::min(slot.sent_count, slot.generated_text.size()); - const std::string str_test = slot.generated_text.substr(pos); - bool is_stop_full = false; - size_t stop_pos = find_stopping_strings(str_test, token_str.size(), STOP_FULL, slot); - if (stop_pos != std::string::npos) - { - is_stop_full = true; - slot.generated_text.erase( - slot.generated_text.begin() + pos + stop_pos, - slot.generated_text.end()); - pos = std::min(slot.sent_count, slot.generated_text.size()); - } - else - { - is_stop_full = false; - stop_pos = find_stopping_strings(str_test, token_str.size(), STOP_PARTIAL, slot); - } + if (slot.multibyte_pending == 0) + { + size_t pos = std::min(slot.sent_count, slot.generated_text.size()); + const std::string str_test = slot.generated_text.substr(pos); + bool is_stop_full = false; + size_t stop_pos = find_stopping_strings(str_test, token_str.size(), STOP_FULL, slot); + if (stop_pos != std::string::npos) + { + is_stop_full = true; + slot.generated_text.erase( + slot.generated_text.begin() + pos + stop_pos, + slot.generated_text.end()); + pos = std::min(slot.sent_count, slot.generated_text.size()); + } + else + { + is_stop_full = false; + stop_pos = find_stopping_strings(str_test, token_str.size(), STOP_PARTIAL, slot); + } - // check if there is any token to predict - if (stop_pos == std::string::npos || (!slot.has_next_token && !is_stop_full && stop_pos > 0)) - { - // no send the stop word in the response - result.text_to_send = slot.generated_text.substr(pos, std::string::npos); - slot.sent_count += result.text_to_send.size(); - // add the token to slot queue and cache - } - slot.add_token_string(result); - if (slot.params.stream) - { - send_partial_response(slot, result); - } - } + // check if there is any token to predict + if (stop_pos == std::string::npos || (!slot.has_next_token && !is_stop_full && stop_pos > 0)) + { + // no send the stop word in the response + result.text_to_send = slot.generated_text.substr(pos, std::string::npos); + slot.sent_count += result.text_to_send.size(); + // add the token to slot queue and cache + } + slot.add_token_string(result); + if (slot.params.stream) + { + send_partial_response(slot, result); + } + } - if (slot.multibyte_pending > 0 && !slot.has_next_token) - { - slot.has_next_token = true; - } + if (slot.multibyte_pending > 0 && !slot.has_next_token) + { + slot.has_next_token = true; + } - // check the limits - if (slot.n_decoded > 2 && slot.has_next_token && !slot.has_budget(params)) - { - slot.stopped_limit = true; - slot.has_next_token = false; - } + // check the limits + if (slot.n_decoded > 2 && slot.has_next_token && !slot.has_budget(params)) + { + slot.stopped_limit = true; + slot.has_next_token = false; + } - if (!slot.cache_tokens.empty() && result.tok == llama_token_eos(ctx)) - { - slot.stopped_eos = true; - slot.has_next_token = false; - LOG_VERBOSE("eos token found", {}); - } + if (!slot.cache_tokens.empty() && result.tok == llama_token_eos(ctx)) + { + slot.stopped_eos = true; + slot.has_next_token = false; + LOG_VERBOSE("eos token found", {}); + } - LOG_VERBOSE("next token", { - {"token", result.tok}, - {"token_text", tokens_to_output_formatted_string(ctx, result.tok)}, - {"has_next_token", slot.has_next_token}, - {"n_remain", slot.n_remaining}, - {"num_tokens_predicted", slot.n_decoded}, - {"stopped_eos", slot.stopped_eos}, - {"stopped_word", slot.stopped_word}, - {"stopped_limit", slot.stopped_limit}, - {"stopping_word", slot.stopping_word}, - }); + LOG_VERBOSE("next token", { + {"token", result.tok}, + {"token_text", tokens_to_output_formatted_string(ctx, result.tok)}, + {"has_next_token", slot.has_next_token}, + {"n_remain", slot.n_remaining}, + {"num_tokens_predicted", slot.n_decoded}, + {"stopped_eos", slot.stopped_eos}, + {"stopped_word", slot.stopped_word}, + {"stopped_limit", slot.stopped_limit}, + {"stopping_word", slot.stopping_word}, + }); - return slot.has_next_token; // continue + return slot.has_next_token; // continue } bool process_images(llama_client_slot &slot) const { - for (slot_image &img : slot.images) - { - if (!img.request_encode_image) - { - continue; - } - clip_image_f32 img_res; - if (!clip_image_preprocess(clp_ctx, &img.img_data, &img_res, /*pad2square =*/ true)) - { - LOG_TEE("Error processing the given image"); - clip_free(clp_ctx); - return false; - } - img.image_tokens = clip_n_patches(clp_ctx); - img.image_embedding = (float *)malloc(clip_embd_nbytes(clp_ctx)); - if (!img.image_embedding) - { - LOG_TEE("Unable to allocate memory for image embeddings\n"); - clip_free(clp_ctx); - return false; - } - LOG_TEE("slot %i - encoding image [id: %i]\n", slot.id, img.id); - if (!clip_image_encode(clp_ctx, params.n_threads, &img_res, img.image_embedding)) - { - LOG_TEE("Unable to encode image\n"); - return false; - } - img.request_encode_image = false; - } + for (slot_image &img : slot.images) + { + if (!img.request_encode_image) + { + continue; + } + clip_image_f32 img_res; + if (!clip_image_preprocess(clp_ctx, &img.img_data, &img_res, /*pad2square =*/ true)) + { + LOG_TEE("Error processing the given image"); + clip_free(clp_ctx); + return false; + } + img.image_tokens = clip_n_patches(clp_ctx); + img.image_embedding = (float *)malloc(clip_embd_nbytes(clp_ctx)); + if (!img.image_embedding) + { + LOG_TEE("Unable to allocate memory for image embeddings\n"); + clip_free(clp_ctx); + return false; + } + LOG_TEE("slot %i - encoding image [id: %i]\n", slot.id, img.id); + if (!clip_image_encode(clp_ctx, params.n_threads, &img_res, img.image_embedding)) + { + LOG_TEE("Unable to encode image\n"); + return false; + } + img.request_encode_image = false; + } - return slot.images.size() > 0; + return slot.images.size() > 0; } void send_error(int id, std::string error) { - std::lock_guard lock(mutex_results); - task_result res; - res.id = id; - res.error = true; - res.result_json = { { "content", error } }; - queue_results.push_back(res); + std::lock_guard lock(mutex_results); + task_result res; + res.id = id; + res.error = true; + res.result_json = { { "content", error } }; + queue_results.push_back(res); } json get_model_props() { - return get_formated_generation(slots[0]); + return get_formated_generation(slots[0]); } json get_formated_generation(llama_client_slot &slot) { - const auto eos_bias = slot.sparams.logit_bias.find(llama_token_eos(ctx)); - const bool ignore_eos = eos_bias != slot.sparams.logit_bias.end() && - eos_bias->second < 0.0f && std::isinf(eos_bias->second); - return json { - {"n_ctx", slot.n_ctx}, - {"model", params.model_alias}, - {"seed", slot.params.seed}, - {"temp", slot.sparams.temp}, - {"top_k", slot.sparams.top_k}, - {"top_p", slot.sparams.top_p}, - {"tfs_z", slot.sparams.tfs_z}, - {"typical_p", slot.sparams.typical_p}, - {"repeat_last_n", slot.sparams.penalty_last_n}, - {"repeat_penalty", slot.sparams.penalty_repeat}, - {"presence_penalty", slot.sparams.penalty_present}, - {"frequency_penalty", slot.sparams.penalty_freq}, - {"mirostat", slot.sparams.mirostat}, - {"mirostat_tau", slot.sparams.mirostat_tau}, - {"mirostat_eta", slot.sparams.mirostat_eta}, - {"penalize_nl", slot.sparams.penalize_nl}, - {"stop", slot.params.antiprompt}, - {"n_predict", slot.params.n_predict}, - {"n_keep", params.n_keep}, - {"ignore_eos", ignore_eos}, - {"stream", slot.params.stream}, - {"logit_bias", slot.sparams.logit_bias}, - {"n_probs", slot.sparams.n_probs}, - {"grammar", slot.sparams.grammar}, - }; + const auto eos_bias = slot.sparams.logit_bias.find(llama_token_eos(ctx)); + const bool ignore_eos = eos_bias != slot.sparams.logit_bias.end() && + eos_bias->second < 0.0f && std::isinf(eos_bias->second); + return json { + {"n_ctx", slot.n_ctx}, + {"model", params.model_alias}, + {"seed", slot.params.seed}, + {"temp", slot.sparams.temp}, + {"top_k", slot.sparams.top_k}, + {"top_p", slot.sparams.top_p}, + {"tfs_z", slot.sparams.tfs_z}, + {"typical_p", slot.sparams.typical_p}, + {"repeat_last_n", slot.sparams.penalty_last_n}, + {"repeat_penalty", slot.sparams.penalty_repeat}, + {"presence_penalty", slot.sparams.penalty_present}, + {"frequency_penalty", slot.sparams.penalty_freq}, + {"mirostat", slot.sparams.mirostat}, + {"mirostat_tau", slot.sparams.mirostat_tau}, + {"mirostat_eta", slot.sparams.mirostat_eta}, + {"penalize_nl", slot.sparams.penalize_nl}, + {"system", slot.params.system}, + {"input_prefix", slot.params.input_prefix}, + {"input_suffix", slot.params.input_suffix}, + {"stop", slot.params.antiprompt}, + {"n_predict", slot.params.n_predict}, + {"n_keep", params.n_keep}, + {"ignore_eos", ignore_eos}, + {"stream", slot.params.stream}, + {"logit_bias", slot.sparams.logit_bias}, + {"n_probs", slot.sparams.n_probs}, + {"grammar", slot.sparams.grammar}, + }; } void send_partial_response(llama_client_slot &slot, completion_token_output tkn) { - std::lock_guard lock(mutex_results); - task_result res; - res.id = slot.task_id; - res.error = false; - res.stop = false; + std::lock_guard lock(mutex_results); + task_result res; + res.id = slot.task_id; + res.error = false; + res.stop = false; - res.result_json = json - { - {"content", tkn.text_to_send}, - {"stop", false}, - {"slot_id", slot.id}, - {"multimodal", multimodal} - }; + res.result_json = json + { + {"content", tkn.text_to_send}, + {"stop", false}, + {"slot_id", slot.id}, + {"multimodal", multimodal} + }; - if (slot.sparams.n_probs > 0) - { - std::vector probs_output = {}; - const std::vector to_send_toks = llama_tokenize(ctx, tkn.text_to_send, false); - size_t probs_pos = std::min(slot.sent_token_probs_index, slot.generated_token_probs.size()); - size_t probs_stop_pos = std::min(slot.sent_token_probs_index + to_send_toks.size(), slot.generated_token_probs.size()); - if (probs_pos < probs_stop_pos) - { - probs_output = std::vector(slot.generated_token_probs.begin() + probs_pos, slot.generated_token_probs.begin() + probs_stop_pos); - } - slot.sent_token_probs_index = probs_stop_pos; - res.result_json["completion_probabilities"] = probs_vector_to_json(ctx, probs_output); - } + if (slot.sparams.n_probs > 0) + { + std::vector probs_output = {}; + const std::vector to_send_toks = llama_tokenize(ctx, tkn.text_to_send, false); + size_t probs_pos = std::min(slot.sent_token_probs_index, slot.generated_token_probs.size()); + size_t probs_stop_pos = std::min(slot.sent_token_probs_index + to_send_toks.size(), slot.generated_token_probs.size()); + if (probs_pos < probs_stop_pos) + { + probs_output = std::vector(slot.generated_token_probs.begin() + probs_pos, slot.generated_token_probs.begin() + probs_stop_pos); + } + slot.sent_token_probs_index = probs_stop_pos; + res.result_json["completion_probabilities"] = probs_vector_to_json(ctx, probs_output); + } - queue_results.push_back(res); + queue_results.push_back(res); } void send_final_response(llama_client_slot &slot) { - std::lock_guard lock(mutex_results); - task_result res; - res.id = slot.task_id; - res.error = false; - res.stop = true; + std::lock_guard lock(mutex_results); + task_result res; + res.id = slot.task_id; + res.error = false; + res.stop = true; - res.result_json = json - { - {"content", !slot.params.stream ? slot.generated_text : ""}, - {"slot_id", slot.id}, - {"stop", true}, - {"model", params.model_alias}, - {"tokens_predicted", slot.n_decoded}, - {"tokens_evaluated", slot.num_prompt_tokens}, - {"generation_settings", get_formated_generation(slot)}, - {"prompt", slot.prompt}, - {"truncated", slot.truncated}, - {"stopped_eos", slot.stopped_eos}, - {"stopped_word", slot.stopped_word}, - {"stopped_limit", slot.stopped_limit}, - {"stopping_word", slot.stopping_word}, - {"tokens_cached", slot.n_past}, - {"timings", slot.get_formated_timings()} - }; + res.result_json = json + { + {"content", !slot.params.stream ? slot.generated_text : ""}, + {"slot_id", slot.id}, + {"stop", true}, + {"model", params.model_alias}, + {"tokens_predicted", slot.n_decoded}, + {"tokens_evaluated", slot.num_prompt_tokens}, + {"generation_settings", get_formated_generation(slot)}, + {"prompt", slot.prompt}, + {"truncated", slot.truncated}, + {"stopped_eos", slot.stopped_eos}, + {"stopped_word", slot.stopped_word}, + {"stopped_limit", slot.stopped_limit}, + {"stopping_word", slot.stopping_word}, + {"tokens_cached", slot.n_past}, + {"timings", slot.get_formated_timings()} + }; - if (slot.sparams.n_probs > 0) - { - std::vector probs = {}; - if (!slot.params.stream && slot.stopped_word) - { - const std::vector stop_word_toks = llama_tokenize(ctx, slot.stopping_word, false); - probs = std::vector(slot.generated_token_probs.begin(), slot.generated_token_probs.end() - stop_word_toks.size()); - } - else - { - probs = std::vector( - slot.generated_token_probs.begin(), - slot.generated_token_probs.begin() + slot.sent_token_probs_index); - } - res.result_json["completion_probabilities"] = probs_vector_to_json(ctx, probs); - } + if (slot.sparams.n_probs > 0) + { + std::vector probs = {}; + if (!slot.params.stream && slot.stopped_word) + { + const std::vector stop_word_toks = llama_tokenize(ctx, slot.stopping_word, false); + probs = std::vector(slot.generated_token_probs.begin(), slot.generated_token_probs.end() - stop_word_toks.size()); + } + else + { + probs = std::vector( + slot.generated_token_probs.begin(), + slot.generated_token_probs.begin() + slot.sent_token_probs_index); + } + res.result_json["completion_probabilities"] = probs_vector_to_json(ctx, probs); + } - queue_results.push_back(res); + queue_results.push_back(res); } void send_embedding(llama_client_slot &slot) { - std::lock_guard lock(mutex_results); - task_result res; - res.id = slot.task_id; - res.error = false; - res.stop = true; + std::lock_guard lock(mutex_results); + task_result res; + res.id = slot.task_id; + res.error = false; + res.stop = true; - const int n_embd = llama_n_embd(model); - if (!params.embedding) - { - LOG_WARNING("embedding disabled", { - {"params.embedding", params.embedding}, - }); - res.result_json = json - { - {"embedding", std::vector(n_embd, 0.0f)}, - }; - } - else - { - const float *data = llama_get_embeddings(ctx); - std::vector embedding(data, data + n_embd); - res.result_json = json - { - {"embedding", embedding }, - }; - } - queue_results.push_back(res); + const int n_embd = llama_n_embd(model); + if (!params.embedding) + { + LOG_WARNING("embedding disabled", { + {"params.embedding", params.embedding}, + }); + res.result_json = json + { + {"embedding", std::vector(n_embd, 0.0f)}, + }; + } + else + { + const float *data = llama_get_embeddings(ctx); + std::vector embedding(data, data + n_embd); + res.result_json = json + { + {"embedding", embedding }, + }; + } + queue_results.push_back(res); } int request_completion(json data, bool infill) { - std::lock_guard lock(mutex_tasks); - task_server task; - task.id = id_gen++; - task.data = data; - task.infill_mode = infill; - task.type = COMPLETION_TASK; - queue_tasks.push_back(task); - return task.id; + std::lock_guard lock(mutex_tasks); + task_server task; + task.id = id_gen++; + task.data = data; + task.infill_mode = infill; + task.type = COMPLETION_TASK; + queue_tasks.push_back(task); + return task.id; } task_result next_result(int task_id) { - while (true) - { - std::this_thread::sleep_for(std::chrono::microseconds(5)); - std::lock_guard lock(mutex_results); + while (true) + { + std::this_thread::sleep_for(std::chrono::microseconds(5)); + std::lock_guard lock(mutex_results); - if (queue_results.empty()) - { - continue; - } + if (queue_results.empty()) + { + continue; + } - for (int i = 0; i < (int) queue_results.size(); i++) - { - if (queue_results[i].id == task_id) - { - task_result res = queue_results[i]; - queue_results.erase(queue_results.begin() + i); - return res; - } - } - } + for (int i = 0; i < (int) queue_results.size(); i++) + { + if (queue_results[i].id == task_id) + { + task_result res = queue_results[i]; + queue_results.erase(queue_results.begin() + i); + return res; + } + } + } - // never reached - //return task_result{-1, false, false, {}}; + // never reached + //return task_result{-1, false, false, {}}; } // for multiple images processing bool ingest_images(llama_client_slot &slot, int n_batch) { - int image_idx = 0; + int image_idx = 0; - while (image_idx < (int) slot.images.size()) - { - slot_image &img = slot.images[image_idx]; + while (image_idx < (int) slot.images.size()) + { + slot_image &img = slot.images[image_idx]; - // process prefix prompt - for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) - { - const int32_t n_tokens = std::min(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)) - { - LOG_TEE("%s : failed to eval\n", __func__); - return false; - } - } + // process prefix prompt + for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) + { + const int32_t n_tokens = std::min(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)) + { + LOG_TEE("%s : failed to eval\n", __func__); + return false; + } + } - // process image with llm - for (int i = 0; i < img.image_tokens; i += n_batch) - { - int n_eval = img.image_tokens - i; - if (n_eval > n_batch) - { - n_eval = n_batch; - } + // process image with llm + for (int i = 0; i < img.image_tokens; i += n_batch) + { + int n_eval = img.image_tokens - i; + if (n_eval > n_batch) + { + n_eval = n_batch; + } - const int n_embd = llama_n_embd(model); - llama_batch batch_img = { n_eval, nullptr, (img.image_embedding + i * n_embd), nullptr, nullptr, nullptr, nullptr, slot.n_past, 1, 0, }; - if (llama_decode(ctx, batch_img)) - { - LOG_TEE("%s : failed to eval image\n", __func__); - return false; - } - slot.n_past += n_eval; - } - image_idx++; + const int n_embd = llama_n_embd(model); + llama_batch batch_img = { n_eval, nullptr, (img.image_embedding + i * n_embd), nullptr, nullptr, nullptr, nullptr, slot.n_past, 1, 0, }; + if (llama_decode(ctx, batch_img)) + { + LOG_TEE("%s : failed to eval image\n", __func__); + return false; + } + slot.n_past += n_eval; + } + image_idx++; - llama_batch_clear(batch); + llama_batch_clear(batch); - // append prefix of next image - const auto json_prompt = (image_idx >= (int) slot.images.size()) ? - slot.params.input_suffix : // no more images, then process suffix prompt - (json)(slot.images[image_idx].prefix_prompt); + // append prefix of next image + const auto json_prompt = (image_idx >= (int) slot.images.size()) ? + slot.params.input_suffix : // no more images, then process suffix prompt + (json)(slot.images[image_idx].prefix_prompt); - std::vector append_tokens = tokenize(json_prompt, false); // has next image - for (int i = 0; i < (int) append_tokens.size(); ++i) - { - llama_batch_add(batch, append_tokens[i], slot.n_past, { slot.id }, true); - slot.n_past += 1; - } - } + std::vector append_tokens = tokenize(json_prompt, false); // has next image + for (int i = 0; i < (int) append_tokens.size(); ++i) + { + llama_batch_add(batch, append_tokens[i], slot.n_past, { slot.id }, true); + slot.n_past += 1; + } + } - return true; + return true; } void request_cancel(int task_id) { - std::lock_guard lock(mutex_tasks); - task_server task; - task.id = id_gen++; - task.type = CANCEL_TASK; - task.target_id = task_id; - queue_tasks.push_back(task); + std::lock_guard lock(mutex_tasks); + task_server task; + task.id = id_gen++; + task.type = CANCEL_TASK; + task.target_id = task_id; + queue_tasks.push_back(task); } void process_tasks() { - std::lock_guard lock(mutex_tasks); - while (!queue_tasks.empty()) - { - task_server task = queue_tasks.front(); - queue_tasks.erase(queue_tasks.begin()); - switch (task.type) - { - case COMPLETION_TASK: { - llama_client_slot *slot = get_slot(json_value(task.data, "slot_id", -1)); - if (slot == nullptr) - { - LOG_TEE("slot unavailable\n"); - // send error result - send_error(task.id, "slot unavaliable"); - return; - } + std::lock_guard lock(mutex_tasks); + while (!queue_tasks.empty()) + { + task_server task = queue_tasks.front(); + queue_tasks.erase(queue_tasks.begin()); + switch (task.type) + { + case COMPLETION_TASK: { + llama_client_slot *slot = get_slot(json_value(task.data, "slot_id", -1)); + if (slot == nullptr) + { + LOG_TEE("slot unavailable\n"); + // send error result + send_error(task.id, "slot unavaliable"); + return; + } - if (task.data.contains("system_prompt")) - { - process_system_prompt_data(task.data["system_prompt"]); - } + if (task.data.contains("system_prompt")) + { + process_system_prompt_data(task.data["system_prompt"]); + } - slot->reset(); + slot->reset(); - slot->infill = task.infill_mode; - slot->task_id = task.id; + slot->infill = task.infill_mode; + slot->task_id = task.id; - if (!launch_slot_with_data(slot, task.data)) - { - // send error result - send_error(task.id, "internal_error"); - break; - } - } break; - case CANCEL_TASK: { // release slot linked with the task id - for (auto & slot : slots) - { - if (slot.task_id == task.target_id) - { - slot.release(); - break; - } - } - } break; - } - } + if (!launch_slot_with_data(slot, task.data)) + { + // send error result + send_error(task.id, "internal_error"); + break; + } + } break; + case CANCEL_TASK: { // release slot linked with the task id + for (auto & slot : slots) + { + if (slot.task_id == task.target_id) + { + slot.release(); + break; + } + } + } break; + } + } } bool update_slots() { - // attend tasks - process_tasks(); + // attend tasks + process_tasks(); - // update the system prompt wait until all slots are idle state - if (system_need_update) - { - LOG_TEE("updating system prompt\n"); - update_system_prompt(); - } + // update the system prompt wait until all slots are idle state + if (system_need_update) + { + LOG_TEE("updating system prompt\n"); + update_system_prompt(); + } - llama_batch_clear(batch); + llama_batch_clear(batch); - if (all_slots_are_idle) - { - if (system_prompt.empty() && clean_kv_cache) - { - LOG_TEE("all slots are idle and system prompt is empty, clear the KV cache\n"); - kv_cache_clear(); - } - // avoid 100% usage of cpu all time - std::this_thread::sleep_for(std::chrono::milliseconds(5)); - } + if (all_slots_are_idle) + { + if (system_prompt.empty() && clean_kv_cache) + { + LOG_TEE("all slots are idle and system prompt is empty, clear the KV cache\n"); + kv_cache_clear(); + } + // avoid 100% usage of cpu all time + std::this_thread::sleep_for(std::chrono::milliseconds(5)); + } - for (llama_client_slot &slot : slots) - { - if (slot.is_processing() && slot.cache_tokens.size() >= (size_t) slot.n_ctx) - { - // Shift context - const int n_left = slot.n_past - slot.params.n_keep - 1; - const int n_discard = n_left / 2; + for (llama_client_slot &slot : slots) + { + if (slot.is_processing() && slot.cache_tokens.size() >= (size_t) slot.n_ctx) + { + // Shift context + const int n_left = slot.n_past - slot.params.n_keep - 1; + 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, slot.params.n_keep, n_left, n_discard); - llama_kv_cache_seq_rm (ctx, slot.id, slot.params.n_keep + 1 , slot.params.n_keep + n_discard + 1); - llama_kv_cache_seq_shift(ctx, slot.id, slot.params.n_keep + 1 + n_discard, slot.n_past, -n_discard); + LOG_TEE("slot %d: context shift - n_keep = %d, n_left = %d, n_discard = %d\n", slot.id, slot.params.n_keep, n_left, n_discard); + llama_kv_cache_seq_rm (ctx, slot.id, slot.params.n_keep + 1 , slot.params.n_keep + n_discard + 1); + llama_kv_cache_seq_shift(ctx, slot.id, slot.params.n_keep + 1 + n_discard, slot.n_past, -n_discard); - for (size_t i = slot.params.n_keep + 1 + n_discard; i < slot.cache_tokens.size(); i++) - { - slot.cache_tokens[i - n_discard] = slot.cache_tokens[i]; - } + for (size_t i = slot.params.n_keep + 1 + n_discard; i < slot.cache_tokens.size(); i++) + { + slot.cache_tokens[i - n_discard] = slot.cache_tokens[i]; + } - slot.cache_tokens.resize(slot.cache_tokens.size() - n_discard); + slot.cache_tokens.resize(slot.cache_tokens.size() - n_discard); - slot.n_past -= n_discard; + slot.n_past -= n_discard; - slot.truncated = true; + slot.truncated = true; - LOG_VERBOSE("context shift", { - {"n_ctx", n_ctx}, - {"n_keep", params.n_keep}, - {"n_left", n_left}, - }); - } - } + LOG_VERBOSE("context shift", { + {"n_ctx", n_ctx}, + {"n_keep", params.n_keep}, + {"n_left", n_left}, + }); + } + } - // decode any currently ongoing sequences - for (auto & slot : slots) - { - // release the slot - if (slot.state == PROCESSING && slot.command == RELEASE) - { - slot.state = IDLE; - slot.command = NONE; - slot.t_last_used = ggml_time_us(); + // decode any currently ongoing sequences + for (auto & slot : slots) + { + // release the slot + if (slot.state == PROCESSING && slot.command == RELEASE) + { + slot.state = IDLE; + 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_TEE("slot %d released (%d tokens in cache)\n", slot.id, (int) slot.cache_tokens.size()); - continue; - } + continue; + } - if (slot.state == IDLE || slot.command == RELEASE) - { - continue; - } + if (slot.state == IDLE || slot.command == RELEASE) + { + continue; + } - slot.i_batch = batch.n_tokens; + slot.i_batch = batch.n_tokens; - llama_batch_add(batch, slot.sampled, system_tokens.size() + slot.n_past, { slot.id }, true); + llama_batch_add(batch, slot.sampled, system_tokens.size() + slot.n_past, { slot.id }, true); - slot.n_decoded += 1; - slot.n_past += 1; - } + slot.n_decoded += 1; + slot.n_past += 1; + } - // process in chunks of params.n_batch - int32_t n_batch = params.n_batch; + // process in chunks of params.n_batch + int32_t n_batch = params.n_batch; - // assign workload to the slots - if (params.cont_batching || batch.n_tokens == 0) - { - for (auto & slot : slots) - { - // need process the prompt - if (slot.state == IDLE && slot.command == LOAD_PROMPT) - { - slot.state = PROCESSING; - slot.command = NONE; - std::vector prompt_tokens; - slot.t_start_process_prompt = ggml_time_us(); - slot.t_start_genereration = 0; + // assign workload to the slots + if (params.cont_batching || batch.n_tokens == 0) + { + for (auto & slot : slots) + { + // need process the prompt + if (slot.state == IDLE && slot.command == LOAD_PROMPT) + { + slot.state = PROCESSING; + slot.command = NONE; + std::vector prompt_tokens; + 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); + 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()); - } + 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(ctx)); - prefix_tokens.insert(prefix_tokens.begin(), llama_token_bos(ctx)); // always add BOS - prefix_tokens.insert(prefix_tokens.end(), llama_token_suffix(ctx)); - prefix_tokens.insert(prefix_tokens.end(), suffix_tokens.begin(), suffix_tokens.end()); - prefix_tokens.push_back(llama_token_middle(ctx)); - prompt_tokens = prefix_tokens; - } - else - { - prompt_tokens = tokenize(slot.prompt, system_prompt.empty()); // add BOS if there isn't system prompt - } + prefix_tokens.insert(prefix_tokens.begin(), llama_token_prefix(ctx)); + prefix_tokens.insert(prefix_tokens.begin(), llama_token_bos(ctx)); // always add BOS + prefix_tokens.insert(prefix_tokens.end(), llama_token_suffix(ctx)); + prefix_tokens.insert(prefix_tokens.end(), suffix_tokens.begin(), suffix_tokens.end()); + prefix_tokens.push_back(llama_token_middle(ctx)); + prompt_tokens = prefix_tokens; + } + else + { + prompt_tokens = tokenize(slot.prompt, system_prompt.empty(), + false, + slot.params.system, + slot.params.input_prefix, + slot.params.input_suffix); // add BOS if there isn't system prompt + } - slot.num_prompt_tokens = prompt_tokens.size(); + slot.num_prompt_tokens = prompt_tokens.size(); - if (!slot.params.cache_prompt) - { - llama_sampling_reset(slot.ctx_sampling); + if (!slot.params.cache_prompt) + { + llama_sampling_reset(slot.ctx_sampling); - slot.n_past = 0; - slot.num_prompt_tokens_processed = slot.num_prompt_tokens; - } - else - { - if (slot.params.n_keep < 0) - { - slot.params.n_keep = slot.num_prompt_tokens; - } - slot.params.n_keep = std::min(slot.n_ctx - 4, slot.params.n_keep); + slot.n_past = 0; + slot.num_prompt_tokens_processed = slot.num_prompt_tokens; + } + else + { + if (slot.params.n_keep < 0) + { + slot.params.n_keep = slot.num_prompt_tokens; + } + slot.params.n_keep = std::min(slot.n_ctx - 4, slot.params.n_keep); - // if input prompt is too big, truncate it - if (slot.num_prompt_tokens >= slot.n_ctx) - { - const int n_left = slot.n_ctx - slot.params.n_keep; - const int n_block_size = n_left / 2; - const int erased_blocks = (slot.num_prompt_tokens - slot.params.n_keep - n_block_size) / n_block_size; + // if input prompt is too big, truncate it + if (slot.num_prompt_tokens >= slot.n_ctx) + { + const int n_left = slot.n_ctx - slot.params.n_keep; + const int n_block_size = n_left / 2; + const int erased_blocks = (slot.num_prompt_tokens - slot.params.n_keep - n_block_size) / n_block_size; - std::vector new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + slot.params.n_keep); - new_tokens.insert(new_tokens.end(), prompt_tokens.begin() + slot.params.n_keep + erased_blocks * n_block_size, prompt_tokens.end()); + std::vector new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + slot.params.n_keep); + new_tokens.insert(new_tokens.end(), prompt_tokens.begin() + slot.params.n_keep + erased_blocks * n_block_size, prompt_tokens.end()); - LOG_VERBOSE("input truncated", { - {"n_ctx", slot.n_ctx}, - {"n_keep", slot.params.n_keep}, - {"n_left", n_left}, - {"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())}, - }); - slot.truncated = true; - prompt_tokens = new_tokens; + LOG_VERBOSE("input truncated", { + {"n_ctx", slot.n_ctx}, + {"n_keep", slot.params.n_keep}, + {"n_left", n_left}, + {"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())}, + }); + slot.truncated = true; + prompt_tokens = new_tokens; - slot.num_prompt_tokens = prompt_tokens.size(); - GGML_ASSERT(slot.num_prompt_tokens < slot.n_ctx); - } + slot.num_prompt_tokens = prompt_tokens.size(); + GGML_ASSERT(slot.num_prompt_tokens < slot.n_ctx); + } - // push the prompt into the sampling context (do not apply grammar) - for (auto &token : prompt_tokens) - { - llama_sampling_accept(slot.ctx_sampling, ctx, token, false); - } + // push the prompt into the sampling context (do not apply grammar) + for (auto &token : prompt_tokens) + { + llama_sampling_accept(slot.ctx_sampling, ctx, token, false); + } - slot.n_past = common_part(slot.cache_tokens, prompt_tokens); - slot.num_prompt_tokens_processed = slot.num_prompt_tokens - slot.n_past; + slot.n_past = common_part(slot.cache_tokens, prompt_tokens); + slot.num_prompt_tokens_processed = slot.num_prompt_tokens - slot.n_past; - LOG_TEE("slot %d : in cache: %i tokens | to process: %i tokens\n", slot.id, slot.n_past, slot.num_prompt_tokens_processed); - } + LOG_TEE("slot %d : in cache: %i tokens | to process: %i tokens\n", slot.id, slot.n_past, slot.num_prompt_tokens_processed); + } - LOG_TEE("slot %d : kv cache rm - [%d, end)\n", slot.id, (int) system_tokens.size() + slot.n_past); + 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); + llama_kv_cache_seq_rm(ctx, slot.id, system_tokens.size() + slot.n_past, -1); - slot.cache_tokens = prompt_tokens; + slot.cache_tokens = prompt_tokens; - if (slot.n_past == slot.num_prompt_tokens) - { - // 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); - slot.n_past--; - } + if (slot.n_past == slot.num_prompt_tokens) + { + // 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); + slot.n_past--; + } - LOG_VERBOSE("prompt ingested", { - {"n_past", slot.n_past}, - {"cached", tokens_to_str(ctx, slot.cache_tokens.cbegin(), slot.cache_tokens.cbegin() + slot.n_past)}, - {"to_eval", tokens_to_str(ctx, slot.cache_tokens.cbegin() + slot.n_past, slot.cache_tokens.cend())}, - }); + LOG_VERBOSE("prompt ingested", { + {"n_past", slot.n_past}, + {"cached", tokens_to_str(ctx, slot.cache_tokens.cbegin(), slot.cache_tokens.cbegin() + slot.n_past)}, + {"to_eval", tokens_to_str(ctx, slot.cache_tokens.cbegin() + slot.n_past, slot.cache_tokens.cend())}, + }); - const bool has_images = process_images(slot); + const bool has_images = process_images(slot); - // process the prefix of first image - std::vector prefix_tokens = has_images ? tokenize(slot.images[0].prefix_prompt, true) : prompt_tokens; - for (; slot.n_past < (int) prefix_tokens.size(); ++slot.n_past) - { - llama_batch_add(batch, prefix_tokens[slot.n_past], system_tokens.size() + slot.n_past, { slot.id }, false); - } + // process the prefix of first image + std::vector prefix_tokens = has_images ? tokenize(slot.images[0].prefix_prompt, true) : prompt_tokens; + for (; slot.n_past < (int) prefix_tokens.size(); ++slot.n_past) + { + llama_batch_add(batch, prefix_tokens[slot.n_past], system_tokens.size() + slot.n_past, { slot.id }, false); + } - if (has_images && !ingest_images(slot, n_batch)) - { - LOG_TEE("failed processing images\n"); - return false; - } + if (has_images && !ingest_images(slot, n_batch)) + { + LOG_TEE("failed processing images\n"); + return false; + } - // extract the logits only for the last token - if (batch.n_tokens > 0) - { - batch.logits[batch.n_tokens - 1] = true; - } + // extract the logits only for the last token + if (batch.n_tokens > 0) + { + batch.logits[batch.n_tokens - 1] = true; + } - slot.n_decoded = 0; - slot.i_batch = batch.n_tokens - 1; - } - } - } + slot.n_decoded = 0; + slot.i_batch = batch.n_tokens - 1; + } + } + } - if (batch.n_tokens == 0) - { - all_slots_are_idle = true; - return true; - } + if (batch.n_tokens == 0) + { + all_slots_are_idle = true; + return true; + } - for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) - { - const int32_t n_tokens = std::min(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 - }; + for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) + { + const int32_t n_tokens = std::min(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 + }; - const int ret = llama_decode(ctx, batch_view); - if (ret != 0) - { - if (n_batch == 1 || ret < 0) - { - // if you get here, it means the KV cache is full - try increasing it via the context size - LOG_TEE("%s : failed to decode the batch, n_batch = %d, ret = %d\n", __func__, n_batch, ret); - return false; - } + const int ret = llama_decode(ctx, batch_view); + if (ret != 0) + { + if (n_batch == 1 || ret < 0) + { + // if you get here, it means the KV cache is full - try increasing it via the context size + LOG_TEE("%s : failed to decode the batch, n_batch = %d, ret = %d\n", __func__, n_batch, ret); + return false; + } - LOG_TEE("%s : failed to find free space in the KV cache, retrying with smaller n_batch = %d\n", __func__, n_batch / 2); + LOG_TEE("%s : failed to find free space in the KV cache, retrying with smaller n_batch = %d\n", __func__, n_batch / 2); - // retry with half the batch size to try to find a free slot in the KV cache - n_batch /= 2; - i -= n_batch; - continue; - } + // retry with half the batch size to try to find a free slot in the KV cache + n_batch /= 2; + i -= n_batch; + continue; + } - for (auto & slot : slots) - { - if (slot.i_batch < (int) i || slot.i_batch >= (int) (i + n_tokens)) - { - continue; - } + for (auto & slot : slots) + { + if (slot.i_batch < (int) i || slot.i_batch >= (int) (i + n_tokens)) + { + continue; + } - // prompt evaluated for embedding - if (params.embedding) - { - send_embedding(slot); - slot.release(); - slot.i_batch = -1; - return true; - } + // prompt evaluated for embedding + if (params.embedding) + { + send_embedding(slot); + slot.release(); + slot.i_batch = -1; + return true; + } - completion_token_output result; - const llama_token id = llama_sampling_sample(slot.ctx_sampling, ctx, NULL, slot.i_batch - i); + completion_token_output result; + const llama_token id = llama_sampling_sample(slot.ctx_sampling, ctx, NULL, slot.i_batch - i); - llama_sampling_accept(slot.ctx_sampling, ctx, id, true); + llama_sampling_accept(slot.ctx_sampling, ctx, id, true); - if (slot.n_decoded == 1) - { - slot.t_start_genereration = ggml_time_us(); - slot.t_prompt_processing = (slot.t_start_genereration - slot.t_start_process_prompt) / 1e3; - } + if (slot.n_decoded == 1) + { + slot.t_start_genereration = ggml_time_us(); + slot.t_prompt_processing = (slot.t_start_genereration - slot.t_start_process_prompt) / 1e3; + } - llama_token_data_array cur_p = { slot.ctx_sampling->cur.data(), slot.ctx_sampling->cur.size(), false }; - result.tok = id; + llama_token_data_array cur_p = { slot.ctx_sampling->cur.data(), slot.ctx_sampling->cur.size(), false }; + result.tok = id; - const int32_t n_probs = slot.sparams.n_probs; - if (slot.sparams.temp <= 0 && n_probs > 0) - { - // for llama_sample_token_greedy we need to sort candidates - llama_sample_softmax(ctx, &cur_p); - } + const int32_t n_probs = slot.sparams.n_probs; + if (slot.sparams.temp <= 0 && n_probs > 0) + { + // for llama_sample_token_greedy we need to sort candidates + llama_sample_softmax(ctx, &cur_p); + } - for (size_t i = 0; i < std::min(cur_p.size, (size_t)n_probs); ++i) - { - result.probs.push_back({cur_p.data[i].id, cur_p.data[i].p}); - } + for (size_t i = 0; i < std::min(cur_p.size, (size_t)n_probs); ++i) + { + result.probs.push_back({cur_p.data[i].id, cur_p.data[i].p}); + } - if (!process_token(result, slot)) - { - slot.release(); - send_final_response(slot); - slot.print_timings(); - } + if (!process_token(result, slot)) + { + slot.release(); + send_final_response(slot); + slot.print_timings(); + } - slot.i_batch = -1; - } - } - return true; + slot.i_batch = -1; + } + } + return true; } }; static void server_print_usage(const char *argv0, const gpt_params ¶ms, - const server_params &sparams) + const server_params &sparams) { printf("usage: %s [options]\n", argv0); printf("\n"); @@ -1777,11 +1866,11 @@ static void server_print_usage(const char *argv0, const gpt_params ¶ms, printf(" not recommended: doubles context memory required and no measurable increase in quality\n"); if (llama_mlock_supported()) { - printf(" --mlock force system to keep model in RAM rather than swapping or compressing\n"); + printf(" --mlock force system to keep model in RAM rather than swapping or compressing\n"); } if (llama_mmap_supported()) { - printf(" --no-mmap do not memory-map model (slower load but may reduce pageouts if not using mlock)\n"); + printf(" --no-mmap do not memory-map model (slower load but may reduce pageouts if not using mlock)\n"); } printf(" --numa attempt optimizations that help on some NUMA systems\n"); #ifdef LLAMA_SUPPORTS_GPU_OFFLOAD @@ -1814,7 +1903,7 @@ static void server_print_usage(const char *argv0, const gpt_params ¶ms, } static void server_params_parse(int argc, char **argv, server_params &sparams, - gpt_params ¶ms, llama_server_context& llama) + gpt_params ¶ms, llama_server_context& llama) { gpt_params default_params; server_params default_sparams; @@ -1823,306 +1912,306 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, for (int i = 1; i < argc; i++) { - arg = argv[i]; - if (arg == "--port") - { - if (++i >= argc) - { - invalid_param = true; - break; - } - sparams.port = std::stoi(argv[i]); - } - else if (arg == "--host") - { - if (++i >= argc) - { - invalid_param = true; - break; - } - sparams.hostname = argv[i]; - } - else if (arg == "--path") - { - if (++i >= argc) - { - invalid_param = true; - break; - } - sparams.public_path = argv[i]; - } - else if (arg == "--timeout" || arg == "-to") - { - if (++i >= argc) - { - invalid_param = true; - break; - } - sparams.read_timeout = std::stoi(argv[i]); - sparams.write_timeout = std::stoi(argv[i]); - } - else if (arg == "-m" || arg == "--model") - { - if (++i >= argc) - { - invalid_param = true; - break; - } - params.model = argv[i]; - } - else if (arg == "-a" || arg == "--alias") - { - if (++i >= argc) - { - invalid_param = true; - break; - } - params.model_alias = argv[i]; - } - else if (arg == "-h" || arg == "--help") - { - server_print_usage(argv[0], default_params, default_sparams); - exit(0); - } - else if (arg == "-c" || arg == "--ctx-size" || arg == "--ctx_size") - { - if (++i >= argc) - { - invalid_param = true; - break; - } - params.n_ctx = std::stoi(argv[i]); - } - else if (arg == "--rope-freq-base") - { - if (++i >= argc) - { - invalid_param = true; - break; - } - params.rope_freq_base = std::stof(argv[i]); - } - else if (arg == "--rope-freq-scale") - { - if (++i >= argc) - { - invalid_param = true; - break; - } - params.rope_freq_scale = std::stof(argv[i]); - } - else if (arg == "--memory-f32" || arg == "--memory_f32") - { - params.memory_f16 = false; - } - else if (arg == "--threads" || arg == "-t") - { - if (++i >= argc) - { - invalid_param = true; - break; - } - params.n_threads = std::stoi(argv[i]); - } - else if (arg == "-b" || arg == "--batch-size") - { - if (++i >= argc) - { - invalid_param = true; - break; - } - params.n_batch = std::stoi(argv[i]); - params.n_batch = std::min(512, params.n_batch); - } - else if (arg == "--gpu-layers" || arg == "-ngl" || arg == "--n-gpu-layers") - { - if (++i >= argc) - { - invalid_param = true; - break; - } + arg = argv[i]; + if (arg == "--port") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + sparams.port = std::stoi(argv[i]); + } + else if (arg == "--host") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + sparams.hostname = argv[i]; + } + else if (arg == "--path") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + sparams.public_path = argv[i]; + } + else if (arg == "--timeout" || arg == "-to") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + sparams.read_timeout = std::stoi(argv[i]); + sparams.write_timeout = std::stoi(argv[i]); + } + else if (arg == "-m" || arg == "--model") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + params.model = argv[i]; + } + else if (arg == "-a" || arg == "--alias") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + params.model_alias = argv[i]; + } + else if (arg == "-h" || arg == "--help") + { + server_print_usage(argv[0], default_params, default_sparams); + exit(0); + } + else if (arg == "-c" || arg == "--ctx-size" || arg == "--ctx_size") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + params.n_ctx = std::stoi(argv[i]); + } + else if (arg == "--rope-freq-base") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + params.rope_freq_base = std::stof(argv[i]); + } + else if (arg == "--rope-freq-scale") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + params.rope_freq_scale = std::stof(argv[i]); + } + else if (arg == "--memory-f32" || arg == "--memory_f32") + { + params.memory_f16 = false; + } + else if (arg == "--threads" || arg == "-t") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + params.n_threads = std::stoi(argv[i]); + } + else if (arg == "-b" || arg == "--batch-size") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + params.n_batch = std::stoi(argv[i]); + params.n_batch = std::min(512, params.n_batch); + } + else if (arg == "--gpu-layers" || arg == "-ngl" || arg == "--n-gpu-layers") + { + if (++i >= argc) + { + invalid_param = true; + break; + } #ifdef LLAMA_SUPPORTS_GPU_OFFLOAD - params.n_gpu_layers = std::stoi(argv[i]); + params.n_gpu_layers = std::stoi(argv[i]); #else - LOG_WARNING("Not compiled with GPU offload support, --n-gpu-layers option will be ignored. " - "See main README.md for information on enabling GPU BLAS support", - {{"n_gpu_layers", params.n_gpu_layers}}); + LOG_WARNING("Not compiled with GPU offload support, --n-gpu-layers option will be ignored. " + "See main README.md for information on enabling GPU BLAS support", + {{"n_gpu_layers", params.n_gpu_layers}}); #endif - } - else if (arg == "--tensor-split" || arg == "-ts") - { - if (++i >= argc) - { - invalid_param = true; - break; - } + } + else if (arg == "--tensor-split" || arg == "-ts") + { + if (++i >= argc) + { + invalid_param = true; + break; + } #ifdef GGML_USE_CUBLAS - std::string arg_next = argv[i]; + std::string arg_next = argv[i]; - // split string by , and / - const std::regex regex{R"([,/]+)"}; - std::sregex_token_iterator it{arg_next.begin(), arg_next.end(), regex, -1}; - std::vector split_arg{it, {}}; - GGML_ASSERT(split_arg.size() <= LLAMA_MAX_DEVICES); + // split string by , and / + const std::regex regex{R"([,/]+)"}; + std::sregex_token_iterator it{arg_next.begin(), arg_next.end(), regex, -1}; + std::vector split_arg{it, {}}; + GGML_ASSERT(split_arg.size() <= LLAMA_MAX_DEVICES); - for (size_t i_device = 0; i_device < LLAMA_MAX_DEVICES; ++i_device) - { - if (i_device < split_arg.size()) - { - params.tensor_split[i_device] = std::stof(split_arg[i_device]); - } - else - { - params.tensor_split[i_device] = 0.0f; - } - } + for (size_t i_device = 0; i_device < LLAMA_MAX_DEVICES; ++i_device) + { + if (i_device < split_arg.size()) + { + params.tensor_split[i_device] = std::stof(split_arg[i_device]); + } + else + { + params.tensor_split[i_device] = 0.0f; + } + } #else - LOG_WARNING("llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n", {}); + LOG_WARNING("llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n", {}); #endif // GGML_USE_CUBLAS - } - else if (arg == "--no-mul-mat-q" || arg == "-nommq") - { + } + else if (arg == "--no-mul-mat-q" || arg == "-nommq") + { #ifdef GGML_USE_CUBLAS - params.mul_mat_q = false; + params.mul_mat_q = false; #else - LOG_WARNING("warning: llama.cpp was compiled without cuBLAS. Disabling mul_mat_q kernels has no effect.\n", {}); + LOG_WARNING("warning: llama.cpp was compiled without cuBLAS. Disabling mul_mat_q kernels has no effect.\n", {}); #endif // GGML_USE_CUBLAS - } - else if (arg == "--main-gpu" || arg == "-mg") - { - if (++i >= argc) - { - invalid_param = true; - break; - } + } + else if (arg == "--main-gpu" || arg == "-mg") + { + if (++i >= argc) + { + invalid_param = true; + break; + } #ifdef GGML_USE_CUBLAS - params.main_gpu = std::stoi(argv[i]); + params.main_gpu = std::stoi(argv[i]); #else - LOG_WARNING("llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.", {}); + LOG_WARNING("llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.", {}); #endif - } - else if (arg == "--lora") - { - if (++i >= argc) - { - invalid_param = true; - break; - } - params.lora_adapter.push_back(std::make_tuple(argv[i], 1.0f)); - params.use_mmap = false; - } - else if (arg == "--lora-scaled") - { - if (++i >= argc) - { - invalid_param = true; - break; - } - const char * lora_adapter = argv[i]; - if (++i >= argc) - { - invalid_param = true; - break; - } - params.lora_adapter.push_back(std::make_tuple(lora_adapter, std::stof(argv[i]))); - params.use_mmap = false; - } - else if (arg == "--lora-base") - { - if (++i >= argc) - { - invalid_param = true; - break; - } - params.lora_base = argv[i]; - } - else if (arg == "-v" || arg == "--verbose") - { + } + else if (arg == "--lora") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + params.lora_adapter.push_back(std::make_tuple(argv[i], 1.0f)); + params.use_mmap = false; + } + else if (arg == "--lora-scaled") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + const char * lora_adapter = argv[i]; + if (++i >= argc) + { + invalid_param = true; + break; + } + params.lora_adapter.push_back(std::make_tuple(lora_adapter, std::stof(argv[i]))); + params.use_mmap = false; + } + else if (arg == "--lora-base") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + params.lora_base = argv[i]; + } + else if (arg == "-v" || arg == "--verbose") + { #if SERVER_VERBOSE != 1 - LOG_WARNING("server.cpp is not built with verbose logging.", {}); + LOG_WARNING("server.cpp is not built with verbose logging.", {}); #else - server_verbose = true; + server_verbose = true; #endif - } - else if (arg == "--mlock") - { - params.use_mlock = true; - } - else if (arg == "--no-mmap") - { - params.use_mmap = false; - } - else if (arg == "--numa") - { - params.numa = true; - } - else if (arg == "--embedding") - { - params.embedding = true; - } - else if (arg == "-cb" || arg == "--cont-batching") - { - params.cont_batching = true; - } - else if (arg == "-np" || arg == "--parallel") - { - if (++i >= argc) - { - invalid_param = true; - break; - } - params.n_parallel = std::stoi(argv[i]); - } else if (arg == "-n" || arg == "--n-predict") - { - if (++i >= argc) - { - invalid_param = true; - break; - } - 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(file), - std::istreambuf_iterator(), - std::back_inserter(systm_content) - ); - llama.process_system_prompt_data(json::parse(systm_content)); - } - else if(arg == "--mmproj") - { - if (++i >= argc) - { - invalid_param = true; - break; - } - params.mmproj = argv[i]; - } - else - { - fprintf(stderr, "error: unknown argument: %s\n", arg.c_str()); - server_print_usage(argv[0], default_params, default_sparams); - exit(1); - } + } + else if (arg == "--mlock") + { + params.use_mlock = true; + } + else if (arg == "--no-mmap") + { + params.use_mmap = false; + } + else if (arg == "--numa") + { + params.numa = true; + } + else if (arg == "--embedding") + { + params.embedding = true; + } + else if (arg == "-cb" || arg == "--cont-batching") + { + params.cont_batching = true; + } + else if (arg == "-np" || arg == "--parallel") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + params.n_parallel = std::stoi(argv[i]); + } else if (arg == "-n" || arg == "--n-predict") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + 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(file), + std::istreambuf_iterator(), + std::back_inserter(systm_content) + ); + llama.process_system_prompt_data(json::parse(systm_content)); + } + else if(arg == "--mmproj") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + params.mmproj = argv[i]; + } + else + { + fprintf(stderr, "error: unknown argument: %s\n", arg.c_str()); + server_print_usage(argv[0], default_params, default_sparams); + exit(1); + } } if (invalid_param) { - fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str()); - server_print_usage(argv[0], default_params, default_sparams); - exit(1); + fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str()); + server_print_usage(argv[0], default_params, default_sparams); + exit(1); } } @@ -2131,15 +2220,15 @@ static json format_partial_response( ) { json res = json { - {"content", content }, - {"stop", false}, - {"slot_id", slot->id }, - {"multimodal", llama.multimodal } + {"content", content }, + {"stop", false}, + {"slot_id", slot->id }, + {"multimodal", llama.multimodal } }; if (slot->sparams.n_probs > 0) { - res["completion_probabilities"] = probs_vector_to_json(llama.ctx, probs); + res["completion_probabilities"] = probs_vector_to_json(llama.ctx, probs); } return res; @@ -2148,31 +2237,31 @@ static json format_partial_response( static json format_tokenizer_response(const std::vector &tokens) { return json{ - {"tokens", tokens}}; + {"tokens", tokens}}; } static json format_detokenized_response(std::string content) { return json{ - {"content", content}}; + {"content", content}}; } static void log_server_request(const httplib::Request &req, const httplib::Response &res) { 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 @@ -2190,11 +2279,11 @@ static void append_to_generated_text_from_generated_token_probs(llama_server_con const size_t len = std::accumulate(gtps.begin(), gtps.end(), size_t(0), add_strlen); if (slot->generated_text.capacity() < slot->generated_text.size() + len) { - slot->generated_text.reserve(slot->generated_text.size() + len); + slot->generated_text.reserve(slot->generated_text.size() + len); } for (const completion_token_output & cto : gtps) { - slot->generated_text += translator(cto); + slot->generated_text += translator(cto); } } @@ -2211,25 +2300,25 @@ int main(int argc, char **argv) if (params.model_alias == "unknown") { - params.model_alias = params.model; + params.model_alias = params.model; } llama_backend_init(params.numa); LOG_INFO("build info", {{"build", BUILD_NUMBER}, - {"commit", BUILD_COMMIT}}); + {"commit", BUILD_COMMIT}}); LOG_INFO("system info", { - {"n_threads", params.n_threads}, - {"n_threads_batch", params.n_threads_batch}, - {"total_threads", std::thread::hardware_concurrency()}, - {"system_info", llama_print_system_info()}, - }); + {"n_threads", params.n_threads}, + {"n_threads_batch", params.n_threads_batch}, + {"total_threads", std::thread::hardware_concurrency()}, + {"system_info", llama_print_system_info()}, + }); // load the model if (!llama.load_model(params)) { - return 1; + return 1; } llama.initialize(); @@ -2237,248 +2326,260 @@ int main(int argc, char **argv) httplib::Server svr; svr.set_default_headers({{"Server", "llama.cpp"}, - {"Access-Control-Allow-Origin", "*"}, - {"Access-Control-Allow-Headers", "content-type"}}); + {"Access-Control-Allow-Origin", "*"}, + {"Access-Control-Allow-Headers", "content-type"}}); // this is only called if no index.html is found in the public --path svr.Get("/", [](const httplib::Request &, httplib::Response &res) - { - res.set_content(reinterpret_cast(&index_html), index_html_len, "text/html"); - return false; - }); + { + res.set_content(reinterpret_cast(&index_html), index_html_len, "text/html"); + return false; + }); // this is only called if no index.js is found in the public --path svr.Get("/index.js", [](const httplib::Request &, httplib::Response &res) - { - res.set_content(reinterpret_cast(&index_js), index_js_len, "text/javascript"); - return false; - }); + { + res.set_content(reinterpret_cast(&index_js), index_js_len, "text/javascript"); + return false; + }); // this is only called if no index.html is found in the public --path svr.Get("/completion.js", [](const httplib::Request &, httplib::Response &res) - { - res.set_content(reinterpret_cast(&completion_js), completion_js_len, "application/javascript"); - return false; - }); + { + res.set_content(reinterpret_cast(&completion_js), completion_js_len, "application/javascript"); + return false; + }); // this is only called if no index.html is found in the public --path svr.Get("/json-schema-to-grammar.mjs", [](const httplib::Request &, httplib::Response &res) - { - res.set_content(reinterpret_cast(&json_schema_to_grammar_mjs), json_schema_to_grammar_mjs_len, "application/javascript"); - return false; - }); + { + res.set_content(reinterpret_cast(&json_schema_to_grammar_mjs), json_schema_to_grammar_mjs_len, "application/javascript"); + return false; + }); svr.Get("/props", [&llama](const httplib::Request & /*req*/, httplib::Response &res) - { - res.set_header("Access-Control-Allow-Origin", "*"); - json data = { - { "user_name", llama.name_user.c_str() }, - { "assistant_name", llama.name_assistant.c_str() } - }; - res.set_content(data.dump(), "application/json"); - }); + { + res.set_header("Access-Control-Allow-Origin", "*"); + json data = { + { "user_name", llama.name_user.c_str() }, + { "assistant_name", llama.name_assistant.c_str() } + }; + res.set_content(data.dump(), "application/json"); + }); svr.Post("/completion", [&llama](const httplib::Request &req, httplib::Response &res) - { - json data = json::parse(req.body); - const int task_id = llama.request_completion(data, false); - if (!json_value(data, "stream", false)) { - std::string completion_text; - task_result result = llama.next_result(task_id); - if(!result.error && result.stop) { - res.set_content(result.result_json.dump(-1, ' ', false, json::error_handler_t::replace), "application/json"); - } - else - { - res.status = 404; - res.set_content(result.result_json["content"], "text/plain"); - return; - } - } else { - const auto chunked_content_provider = [task_id, &llama](size_t, httplib::DataSink & sink) - { - while (true) - { - task_result result = llama.next_result(task_id); - if (!result.error) { - const std::string str = - "data: " + - result.result_json.dump(-1, ' ', false, json::error_handler_t::replace) + - "\n\n"; - LOG_VERBOSE("data stream", { - { "to_send", str } - }); - if (!sink.write(str.c_str(), str.size())) - { - return false; - } - if(result.stop) { - break; - } - } else { - break; - } - } - sink.done(); - return true; - }; + { + json data = json::parse(req.body); + const int task_id = llama.request_completion(data, false); + if (!json_value(data, "stream", false)) { + std::string completion_text; + task_result result = llama.next_result(task_id); + if(!result.error && result.stop) { + res.set_content(result.result_json.dump(-1, ' ', false, json::error_handler_t::replace), "application/json"); + } + else + { + res.status = 404; + res.set_content(result.result_json["content"], "text/plain"); + return; + } + } else { + const auto chunked_content_provider = [task_id, &llama](size_t, httplib::DataSink & sink) + { + while (true) + { + task_result result = llama.next_result(task_id); + if (!result.error) { + const std::string str = + "data: " + + result.result_json.dump(-1, ' ', false, json::error_handler_t::replace) + + "\n\n"; + LOG_VERBOSE("data stream", { + { "to_send", str } + }); + if (!sink.write(str.c_str(), str.size())) + { + return false; + } + if(result.stop) { + break; + } + } else { + break; + } + } + sink.done(); + return true; + }; - auto on_complete = [task_id, &llama] (bool) - { - // cancel - llama.request_cancel(task_id); - }; + auto on_complete = [task_id, &llama] (bool) + { + // cancel + llama.request_cancel(task_id); + }; - res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete); - } - }); + res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete); + } + }); svr.Post("/infill", [&llama](const httplib::Request &req, httplib::Response &res) - { - json data = json::parse(req.body); - const int task_id = llama.request_completion(data, true); - if (!json_value(data, "stream", false)) { - std::string completion_text; - task_result result = llama.next_result(task_id); - if (!result.error && result.stop) - { - res.set_content(result.result_json.dump(-1, ' ', false, json::error_handler_t::replace), "application/json"); - } - else - { - res.status = 404; - res.set_content(result.result_json["content"], "text/plain"); - return; - } - } else { - const auto chunked_content_provider = [task_id, &llama](size_t, httplib::DataSink & sink) { - while (true) - { - task_result result = llama.next_result(task_id); - if (!result.error) { - const std::string str = - "data: " + - result.result_json.dump(-1, ' ', false, json::error_handler_t::replace) + - "\n\n"; - LOG_VERBOSE("data stream", { - { "to_send", str } - }); - if (!sink.write(str.c_str(), str.size())) - { - return false; - } - if (result.stop) - { - break; - } - } - else - { - break; - } - } + { + json data = json::parse(req.body); + const int task_id = llama.request_completion(data, true); + if (!json_value(data, "stream", false)) { + std::string completion_text; + task_result result = llama.next_result(task_id); + if (!result.error && result.stop) + { + res.set_content(result.result_json.dump(-1, ' ', false, json::error_handler_t::replace), "application/json"); + } + else + { + res.status = 404; + res.set_content(result.result_json["content"], "text/plain"); + return; + } + } else { + const auto chunked_content_provider = [task_id, &llama](size_t, httplib::DataSink & sink) { + while (true) + { + task_result result = llama.next_result(task_id); + if (!result.error) { + const std::string str = + "data: " + + result.result_json.dump(-1, ' ', false, json::error_handler_t::replace) + + "\n\n"; + LOG_VERBOSE("data stream", { + { "to_send", str } + }); + if (!sink.write(str.c_str(), str.size())) + { + return false; + } + if (result.stop) + { + break; + } + } + else + { + break; + } + } - sink.done(); + sink.done(); - return true; - }; + return true; + }; - auto on_complete = [task_id, &llama] (bool) - { - // cancel - llama.request_cancel(task_id); - }; + auto on_complete = [task_id, &llama] (bool) + { + // cancel + llama.request_cancel(task_id); + }; - res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete); - } - }); + res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete); + } + }); svr.Get("/model.json", [&llama](const httplib::Request &, httplib::Response &res) - { - const json data = llama.get_model_props(); - return res.set_content(data.dump(), "application/json"); - }); + { + const json data = llama.get_model_props(); + return res.set_content(data.dump(), "application/json"); + }); svr.Options(R"(/.*)", [](const httplib::Request &, httplib::Response &res) - { return res.set_content("", "application/json"); }); + { return res.set_content("", "application/json"); }); svr.Post("/tokenize", [&llama](const httplib::Request &req, httplib::Response &res) - { - const json body = json::parse(req.body); - std::vector 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"); - }); + { + const json body = json::parse(req.body); + std::vector 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"); + }); + + svr.Post("/tokenizes", [&llama](const httplib::Request &req, httplib::Response &res) + { + const json body = json::parse(req.body); + std::vector tokens; + if (body.count("content") != 0) + { + tokens = llama.tokenize(body["content"], false, true); + } + const json data = format_tokenizer_response(tokens); + return res.set_content(data.dump(), "application/json"); + }); svr.Post("/detokenize", [&llama](const httplib::Request &req, httplib::Response &res) - { - const json body = json::parse(req.body); - std::string content; - if (body.count("tokens") != 0) - { - const std::vector tokens = body["tokens"]; - content = tokens_to_str(llama.ctx, tokens.cbegin(), tokens.cend()); - } + { + const json body = json::parse(req.body); + std::string content; + if (body.count("tokens") != 0) + { + const std::vector 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"); - }); + const json data = format_detokenized_response(content); + return res.set_content(data.dump(), "application/json"); + }); svr.Post("/embedding", [&llama](const httplib::Request &req, httplib::Response &res) - { - const json body = json::parse(req.body); - json prompt; - if (body.count("content") != 0) - { - prompt = body["content"]; - } - else - { - prompt = ""; - } - const int task_id = llama.request_completion({ {"prompt", prompt}, { "n_predict", 0} }, false); - task_result result = llama.next_result(task_id); - return res.set_content(result.result_json.dump(), "application/json"); - }); + { + const json body = json::parse(req.body); + json prompt; + if (body.count("content") != 0) + { + prompt = body["content"]; + } + else + { + prompt = ""; + } + const int task_id = llama.request_completion({ {"prompt", prompt}, { "n_predict", 0} }, false); + task_result result = llama.next_result(task_id); + return res.set_content(result.result_json.dump(), "application/json"); + }); svr.set_logger(log_server_request); svr.set_exception_handler([](const httplib::Request &, httplib::Response &res, std::exception_ptr ep) - { - const char fmt[] = "500 Internal Server Error\n%s"; - char buf[BUFSIZ]; - try - { - std::rethrow_exception(std::move(ep)); - } - catch (std::exception &e) - { - snprintf(buf, sizeof(buf), fmt, e.what()); - } - catch (...) - { - snprintf(buf, sizeof(buf), fmt, "Unknown Exception"); - } - res.set_content(buf, "text/plain"); - res.status = 500; - }); + { + const char fmt[] = "500 Internal Server Error\n%s"; + char buf[BUFSIZ]; + try + { + std::rethrow_exception(std::move(ep)); + } + catch (std::exception &e) + { + snprintf(buf, sizeof(buf), fmt, e.what()); + } + catch (...) + { + snprintf(buf, sizeof(buf), fmt, "Unknown Exception"); + } + res.set_content(buf, "text/plain"); + res.status = 500; + }); svr.set_error_handler([](const httplib::Request &, httplib::Response &res) - { - if (res.status == 400) - { - res.set_content("Invalid request", "text/plain"); - } - else if (res.status != 500) - { - res.set_content("File Not Found", "text/plain"); - res.status = 404; - } - }); + { + if (res.status == 400) + { + res.set_content("Invalid request", "text/plain"); + } + else if (res.status != 500) + { + res.set_content("File Not Found", "text/plain"); + res.status = 404; + } + }); // set timeouts and change hostname and port svr.set_read_timeout (sparams.read_timeout); @@ -2486,8 +2587,8 @@ int main(int argc, char **argv) if (!svr.bind_to_port(sparams.hostname, sparams.port)) { - fprintf(stderr, "\ncouldn't bind to server socket: hostname=%s port=%d\n\n", sparams.hostname.c_str(), sparams.port); - return 1; + fprintf(stderr, "\ncouldn't bind to server socket: hostname=%s port=%d\n\n", sparams.hostname.c_str(), sparams.port); + return 1; } // Set the base directory for serving static files @@ -2497,30 +2598,30 @@ int main(int argc, char **argv) LOG_TEE("\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port); LOG_INFO("HTTP server listening", { - {"hostname", sparams.hostname}, - {"port", sparams.port}, - }); + {"hostname", sparams.hostname}, + {"port", sparams.port}, + }); // run the HTTP server in a thread - see comment below std::thread t([&]() - { - if (!svr.listen_after_bind()) - { - return 1; - } + { + if (!svr.listen_after_bind()) + { + return 1; + } - return 0; - }); + return 0; + }); // GG: if I put the main loop inside a thread, it crashes on the first request when build in Debug!? // "Bus error: 10" - this is on macOS, it does not crash on Linux //std::thread t2([&]() { - bool running = true; - while (running) - { - running = llama.update_slots(); - } + bool running = true; + while (running) + { + running = llama.update_slots(); + } } //);