diff --git a/examples/server/public/index.html b/examples/server/public/index.html index d65c540cf..27570a1ee 100644 --- a/examples/server/public/index.html +++ b/examples/server/public/index.html @@ -119,36 +119,6 @@ - - -
-
- Templates - - - -
- -
@@ -319,7 +289,6 @@ // Note: in order not to introduce breaking changes, please keep the same data type (number, string, etc) if you want to change the default value. Do not use null or undefined for default value. apiKey: '', systemMessage: 'You are a helpful assistant.', - chat_template: 'chatml', input_prefix: '', input_suffix: '', // make sure these default values are in sync with `common.h` @@ -347,7 +316,6 @@ const CONFIG_INFO = { apiKey: 'Set the API Key if you are using --api-key option for the server.', systemMessage: 'The starting message that defines how model should behave.', - chat_template: 'The fromat used for messages.', input_prefix: 'Prefix for user messages in custom chat templates.', input_suffix: 'Suffix for user messages in custom chat templates.', samplers: 'The order at which samplers are applied, in simplified way. Default is "dkypmxt": dry->top_k->typ_p->top_p->min_p->xtc->temperature', @@ -375,7 +343,6 @@ const CONFIG_NUMERIC_KEYS = Object.entries(CONFIG_DEFAULT).filter(e => isNumeric(e[1])).map(e => e[0]); // list of themes supported by daisyui const THEMES = ['light', 'dark', 'cupcake', 'bumblebee', 'emerald', 'corporate', 'synthwave', 'retro', 'cyberpunk', 'valentine', 'halloween', 'garden', 'forest', 'aqua', 'lofi', 'pastel', 'fantasy', 'wireframe', 'black', 'luxury', 'dracula', 'cmyk', 'autumn', 'business', 'acid', 'lemonade', 'night', 'coffee', 'winter', 'dim', 'nord', 'sunset']; - const CHAT_TEMPLATES = ['chatml', 'llama2', 'mistral', 'phi3', 'zephyr', 'monarch', 'gemma', 'gemma2', 'orion', 'openchat', 'vicuna', 'vicuna-orca', 'deepseek', 'command-r', 'llama3', 'chatglm3', 'chatglm4', 'minicpm', 'deepseek2', 'exaone3', 'rwkv-world', 'granite', 'custom']; // markdown support const VueMarkdown = defineComponent( @@ -522,7 +489,6 @@ editingMsg: null, // const themes: THEMES, - templates: CHAT_TEMPLATES, configDefault: {...CONFIG_DEFAULT}, configInfo: {...CONFIG_INFO}, } @@ -542,9 +508,6 @@ this.selectedTheme = theme; StorageUtils.setTheme(theme); }, - setSelectedTemplate(template) { - this.config.chat_template = template; - }, newConversation() { if (this.isGenerating) return; this.viewingConvId = StorageUtils.getNewConvId(); @@ -604,7 +567,6 @@ stream: true, cache_prompt: true, samplers: this.config.samplers, - chat_template: this.config.chat_template, input_prefix: this.config.input_prefix, input_suffix: this.config.input_suffix, temperature: this.config.temperature, diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 644649715..85a65efb9 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -798,6 +798,7 @@ struct server_context { slot.oaicompat = false; slot.oaicompat_model = ""; } + std::string default_empty = ""; slot.params.stream = json_value(data, "stream", false); slot.params.cache_prompt = json_value(data, "cache_prompt", false); @@ -1147,9 +1148,6 @@ struct server_context { {"model", params.model_alias}, {"seed", slot.sparams.seed}, {"seed_cur", slot.smpl ? common_sampler_get_seed(slot.smpl) : 0}, - {"chat_template", params.chat_template}, - {"input_prefix", params.input_prefix}, - {"input_suffix", params.input_suffix}, {"temperature", slot.sparams.temp}, {"dynatemp_range", slot.sparams.dynatemp_range}, {"dynatemp_exponent", slot.sparams.dynatemp_exponent}, @@ -3221,22 +3219,18 @@ int main(int argc, char ** argv) { LOG_INF("%s: model loaded\n", __func__); - // if a standard chat template is not chosen, check prefix and suffix to switch to custom template + // if a standard chat template is not chosen, check prefix and suffix to switch to custom formatting // otherwise use the one that comes with the model (if any) // if a standard chat template is chosen, warn about prefix and suffix not being used if (params.chat_template.empty()) { if (!params.input_prefix.empty() || !params.input_suffix.empty()) { - LOG_WRN("%s: Prefix and suffix are used instead of a chat template. This may cause the model to output suboptimal responses\n", __func__); - params.chat_template = "custom"; + LOG_WRN("%s: Prefix and suffix will be used for a custom chat template. This may cause the model to output suboptimal responses\n", __func__); } else if (!ctx_server.validate_model_chat_template()) { LOG_WRN("%s: The chat template that comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses\n", __func__); params.chat_template = "chatml"; } - } else if (params.chat_template != "custom" && - (!params.input_prefix.empty() || !params.input_suffix.empty())) { - LOG_WRN("%s: Prefix and suffix are defined, but will not be used because a standard chat template is chosen.\n", __func__); - } else { - LOG_WRN("%s: Custom chat template is chosen. This may cause the model to output suboptimal responses\n", __func__); + } else if (!params.input_prefix.empty() || !params.input_suffix.empty()) { + LOG_WRN("%s: Prefix and suffix are defined, but will not be used because a chat template '%s' is chosen.\n", __func__, params.chat_template.c_str()); } // print sample chat example to make it clear which template is used diff --git a/examples/server/utils.hpp b/examples/server/utils.hpp index 89ec4f335..6c51b9f12 100644 --- a/examples/server/utils.hpp +++ b/examples/server/utils.hpp @@ -304,6 +304,8 @@ inline std::string format_chat(const struct llama_model * model, const std::stri std::vector chat; std::string formatted_chat; + bool is_custom = !prefix.empty() || !suffix.empty(); + for (size_t i = 0; i < messages.size(); ++i) { const auto & curr_msg = messages[i]; @@ -326,7 +328,7 @@ inline std::string format_chat(const struct llama_model * model, const std::stri throw std::runtime_error("Missing 'content' (ref: https://github.com/ggerganov/llama.cpp/issues/8367)"); } - if (tmpl == "custom") { + if (is_custom) { // simple format using prefix and suffix if (role == "user") formatted_chat += prefix + content + suffix; else formatted_chat += content; @@ -335,7 +337,7 @@ inline std::string format_chat(const struct llama_model * model, const std::stri } } - if (tmpl != "custom") formatted_chat = common_chat_apply_template(model, tmpl, chat, true); + if (!is_custom) formatted_chat = common_chat_apply_template(model, tmpl, chat, true); LOG_WRN("formatted_chat using '%s': '%s'\n", tmpl.c_str(), formatted_chat.c_str()); return formatted_chat; @@ -351,7 +353,7 @@ inline std::string format_chat_example(const struct llama_model * model, const s std::string formatted_example; - if (tmpl == "custom") { + if (!prefix.empty() || !suffix.empty()) { for (auto message : msgs) { if (message.role == "user") formatted_example += prefix + message.content + suffix; else formatted_example += message.content; @@ -634,7 +636,20 @@ static json oaicompat_completion_params_parse( llama_params["__oaicompat"] = true; // Apply chat template to the list of messages - llama_params["prompt"] = format_chat(model, chat_template, input_prefix, input_suffix, body.at("messages")); + std::string chat_tmpl = chat_template; + std::string prefix = ""; + std::string suffix = ""; + + // if template is sent in data, ignore prefix and suffix + if (body.contains("chat_template")) { + chat_tmpl = body.at("chat_template").get(); + LOG_WRN("\nUsing '%s' template, prefix and suffix are ignored.\n", chat_tmpl.c_str()); + } else { + prefix = (body.contains("input_prefix") ? body.at("input_prefix").get() : input_prefix); + suffix = (body.contains("input_suffix") ? body.at("input_suffix").get() : input_suffix); + } + + llama_params["prompt"] = format_chat(model, chat_tmpl, prefix, suffix, body.at("messages")); // Handle "stop" field if (body.contains("stop") && body.at("stop").is_string()) {