eos_ignores -> emits_eos

chat.cpp

@@ -1,6 +1,7 @@
 #include "chat.hpp"
 
 #include <iostream>
+#include <limits>
 #include <commoncpp/utils.hpp>
 #include <justlm.hpp>
 
@@ -9,12 +10,11 @@ namespace LM {
 namespace Chat {
 LM::Inference::Params Inference::get_params() const {
     LM::Inference::Params fres;
+    fres.n_eos_ignores = config.emits_eos?0:std::numeric_limits<unsigned>::max();
     if (config.max_context_size.has_value())
         fres.n_ctx = config.max_context_size.value();
     if (config.repeat_last.has_value())
         fres.n_repeat_last = config.repeat_last.value();
-    if (config.eos_ignores.has_value())
-        fres.n_eos_ignores = config.eos_ignores.value();
     if (config.top_k.has_value())
         fres.top_k = config.top_k.value();
     if (config.top_p.has_value())
@@ -52,12 +52,14 @@ LM_SCHEDULABLE(bool) Inference::reset() {
 LM_SCHEDULABLE(bool) Inference::append(const ModelConfig::Prompt &promptConfig, const std::string &message, const EvaluateCallback &on_evaluate) {
     auto inference = LM_COAWAIT get_underlaying();
     bool non_cancelled = true;
-    // Append prompt prefix
-    LM_ERROR_CATCH(LM_COAWAIT inference->append(promptConfig.prefix, [on_evaluate] (float progress) -> bool {
-        if (on_evaluate)
-            on_evaluate(progress/3.0f, true);
-        return true; // Can't be cancelled here
-    }), LM_BOOL_ERROR, {LM_CORETURN "";});
+    // Append prompt prefix if needed
+    if (!common::utils::ends_with(inference->get_prompt(), promptConfig.prefix)) {
+        LM_ERROR_CATCH(LM_COAWAIT inference->append(promptConfig.prefix, [on_evaluate] (float progress) -> bool {
+            if (on_evaluate)
+                on_evaluate(progress/3.0f, true);
+            return true; // Can't be cancelled here
+        }), LM_BOOL_ERROR, {LM_CORETURN "";});
+    }
     // Append prompt
     LM_ERROR_CATCH(LM_COAWAIT inference->append(message, [on_evaluate, &non_cancelled] (float progress) -> bool {
         return non_cancelled = on_evaluate?on_evaluate(progress/3.0f+100.0f/3.0f, false):true;

commoncpp

@@ -1 +1 @@
-Subproject commit ec148c3c3447b26b3213b0277566445456693bdf
+Subproject commit 54f9a8e586d6c710999cb7f033f6915cff92e35f

examples/CMakeLists.txt

@@ -1,6 +1,6 @@
 cmake_minimum_required(VERSION 3.5)
 
-project(anyproc LANGUAGES CXX)
+project(justchat LANGUAGES CXX)
 
 set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)

examples/cli.cpp

@@ -14,7 +14,10 @@ int main(int argc, char **argv) {
     const auto model_config = argv[1];
     // Construct chat model
     LM::Chat::Inference model(model_config);
-    model.start();
+    if (!model.start()) {
+        std::cerr << "Failed to load model." << std::endl;
+        return EXIT_FAILURE;
+    }
     std::string instruction;
     for (;;) {
         std::cout << "> ";

include/justchat/model_config.hpp

@@ -30,11 +30,11 @@ public:
                 language = "en";
     Prompt mutable prompt;
     bool allow_system_prompt = true,
-         strict_prompt = false;
+         strict_prompt = false,
+         emits_eos = true; // Weather the model emits an eos at the end of each response
 
     std::optional<unsigned> max_context_size,
                             repeat_last, // How many tokens to repeat-penalize
-                            eos_ignores, // How many times to ignore EOS
                             top_k,
                             mirostat_version; // Version of mirostat to use; 0 for none
     std::optional<float> top_p,

model_config.cpp

@@ -25,8 +25,8 @@ void ModelConfig::fill(KeyValueMap &&map, bool ignore_extra) {
             max_context_size = std::stoul(value);
         else if (key == "repeat_last")
             repeat_last = std::stoul(value);
-        else if (key == "eos_ignores")
-            eos_ignores = std::stoi(value); // -1 for "infinite"
+        else if (key == "emits_eos")
+            emits_eos = parse_bool(value);
         else if (key == "top_k")
             top_k = std::stoi(value);
         else if (key == "mirostat_version")