add streaming via server-sent events

Removes /next-token endpoint and adds a "stream" parameter to the
/completion one.

examples/server/server.cpp

@@ -13,7 +13,7 @@ struct server_params
 
 struct llama_server_context
 {
-  bool as_loop = false;
+  bool stream = false;
   bool has_next_token = false;
   std::string generated_text = "";
 
@@ -35,7 +35,6 @@ struct llama_server_context
   std::string stopping_word;
 
   void rewind() {
-    as_loop = false;
     params.antiprompt.clear();
     num_tokens_predicted = 0;
     generated_text = "";
@@ -253,9 +252,6 @@ struct llama_server_context
     if (token == -1) {
       return "";
     }
-    if(as_loop) {
-      generated_text = "";
-    }
 
     std::string token_text = llama_token_to_str(ctx, token);
     generated_text += token_text;
@@ -270,7 +266,7 @@ struct llama_server_context
       }
     }
 
-    return generated_text;
+    return token_text;
   }
 
   std::vector<float> embedding(std::string content, int threads) {
@@ -478,9 +474,13 @@ bool server_params_parse(int argc, char **argv, server_params &sparams, gpt_para
 
 bool parse_options_completion(json body, llama_server_context& llama, Response &res) {
   gpt_params default_params;
-  if (!body["as_loop"].is_null())
+  if (!body["stream"].is_null())
   {
-    llama.as_loop = body["as_loop"].get<bool>();
+      llama.stream = body["stream"].get<bool>();
+  }
+  else
+  {
+      llama.stream = false;
   }
   if (!body["n_predict"].is_null())
   {
@@ -671,8 +671,6 @@ int main(int argc, char **argv)
   llama_server_context llama;
   params.model = "ggml-model.bin";
 
-  std::string final_text;
-
   if (server_params_parse(argc, argv, sparams, params) == false)
   {
     return 1;
@@ -689,65 +687,80 @@ int main(int argc, char **argv)
   svr.Get("/", [](const Request &, Response &res)
           { res.set_content("<h1>llama.cpp server works</h1>", "text/html"); });
 
-  svr.Post("/completion", [&llama, &final_text](const Request &req, Response &res)
+  svr.Post("/completion", [&llama](const Request &req, Response &res) {
-            {
       if (llama.params.embedding) {
           json data = {
               {"status", "error"},
               {"reason", "To use completion function, disable embedding mode"}};
-                res.set_content(data.dump(), "application/json");
+          res.set_content(data.dump(-1, ' ', false, json::error_handler_t::replace),
+                          "application/json");
           res.status = 400;
           return;
       }
 
       llama.rewind();
-              final_text = "";
 
       if (parse_options_completion(json::parse(req.body), llama, res) == false) {
           return;
       }
 
-              if (!llama.loadPrompt())
+      if (!llama.loadPrompt()) {
-              {
+          json data = {{"status", "error"}, {"reason", "Context too long."}};
-                json data = {
+          res.set_content(data.dump(-1, ' ', false, json::error_handler_t::replace),
-                    {"status", "error"},
+                          "application/json");
-                    {"reason", "Context too long."}};
-                res.set_content(data.dump(), "application/json");
           res.status = 400;
           return;
       }
 
       llama.beginCompletion();
-              if(llama.as_loop) {
+
-                json data = {
+      if (!llama.stream) {
-                    {"status", "done" } };
+          while (llama.has_next_token) {
-                return res.set_content(data.dump(), "application/json");
-              } else {
-                // loop inference until finish completion
-                while (llama.has_next_token)
-                {
               llama.doCompletion();
           }
-                try
+
-                {
+          json data = {{"content", llama.generated_text},
-                  json data = {
+                       {"stop", true},
                        {"model", llama.params.model_alias },
-                      {"content", llama.generated_text },
                        {"tokens_predicted", llama.num_tokens_predicted},
                        {"generation_settings", format_generation_settings(llama)},
                        {"prompt", llama.params.prompt},
                        {"stopping_word", llama.stopping_word}};
-                  return res.set_content(data.dump(), "application/json");
+          return res.set_content(data.dump(-1, ' ', false, json::error_handler_t::replace), "application/json");
+      } else {
+          const auto chunked_content_provider = [&](size_t, DataSink &sink) {
+              while (llama.has_next_token) {
+                  std::string token_text = llama.doCompletion();
+
+                  json data;
+                  if (llama.has_next_token) {
+                      data = {{"content", token_text}, {"stop", false}};
+                  } else {
+                      // Generation is done, send extra information.
+                      data = {
+                          {"content", token_text},
+                          {"stop", true},
+                          {"model", llama.params.model_alias},
+                          {"tokens_predicted", llama.num_tokens_predicted},
+                          {"generation_settings", format_generation_settings(llama)},
+                          {"prompt", llama.params.prompt},
+                          {"stopping_word", llama.stopping_word},
+                          {"generated_text", llama.generated_text}};
                   }
-                catch (const json::exception &e)
+
-                {
+                  std::string str =
-                  // Some tokens have bad UTF-8 strings, the json parser is very sensitive
+                      "data: " +
-                  json data = {
+                      data.dump(-1, ' ', false, json::error_handler_t::replace) +
-                      {"content", "Bad encoding token"},
+                      "\n\n";
-                      {"tokens_predicted", 0}};
+                  sink.write(str.data(), str.size());
-                  return res.set_content(data.dump(), "application/json");
               }
-              } });
+
+              sink.done();
+              return true;
+          };
+          res.set_chunked_content_provider("text/event-stream", chunked_content_provider);
+      }
+  });
 
   svr.Post("/tokenize", [&llama](const Request &req, Response &res)
             {
@@ -774,72 +787,6 @@ int main(int argc, char **argv)
               return res.set_content(data.dump(), "application/json");
             });
 
-  svr.Get("/next-token", [&llama, &final_text](const Request &req, Response &res)
-          {
-            if(llama.params.embedding) {
-                res.set_content("{}", "application/json");
-                return;
-            }
-            std::string result = "";
-            if (req.has_param("stop")) {
-                llama.has_next_token = false;
-            } else {
-              result = llama.doCompletion(); // inference next token
-            }
-            try {
-              json data;
-              if (llama.has_next_token)
-              {
-                  //fprintf(stdout, "Result: %s\n", result);
-                  final_text += result;
-                  data = {
-                        {"content", result },
-                        {"stop", false }
-                  };
-              }
-              else
-              {
-                  // Generation is done, send extra information.
-                  data = {
-                        {"content", result },
-                        {"stop", true },
-                        {"tokens_predicted", llama.num_tokens_predicted},
-                        {"generation_settings", format_generation_settings(llama)},
-                        {"prompt", llama.params.prompt},
-                        {"stopping_word", llama.stopping_word},
-                        {"generated_text", final_text}
-                  };
-              }
-
-              return res.set_content(data.dump(), "application/json");
-            } catch (const json::exception &e) {
-              // Some tokens have bad UTF-8 strings, the json parser is very sensitive
-              json data;
-              if (llama.has_next_token)
-              {
-                  final_text += u8"\uFFFD";
-                  data = {
-                        {"content", result },
-                        {"stop", false }
-                  };
-              }
-              else
-              {
-                  // Generation is done, send extra information.
-                  data = {
-                        {"content", u8"\uFFFD" },
-                        {"stop", true },
-                        {"tokens_predicted", llama.num_tokens_predicted},
-                        {"generation_settings", format_generation_settings(llama)},
-                        {"prompt", llama.params.prompt},
-                        {"stopping_word", llama.stopping_word},
-                        {"generated_text", final_text}
-                  };
-              }
-              return res.set_content(data.dump(), "application/json");
-            }
-          });
-
   fprintf(stderr, "%s: http server Listening at http://%s:%i\n", __func__, sparams.hostname.c_str(), sparams.port);
 
   if(params.embedding) {