Explicitly clear the kv cache each time we eval tokens to match n_past. (#1808)

Inspired by Microsoft docs for LoadLibraryExA (https://learn.microsoft.com/en-us/windows/win32/api/libloaderapi/nf-libloaderapi-loadlibraryexa).
When using LOAD_LIBRARY_SEARCH_DLL_LOAD_DIR, the lpFileName parameter must specify a fully qualified path, also it needs to be backslashes (\), not forward slashes (/).

gpt4all-backend/dlhandle.h

@@ -53,6 +53,8 @@ public:
     }
 };
 #else
+#include <algorithm>
+#include <filesystem>
 #include <string>
 #include <exception>
 #include <stdexcept>
@@ -75,7 +77,9 @@ public:
 
     Dlhandle() : chandle(nullptr) {}
     Dlhandle(const std::string& fpath) {
-        chandle = LoadLibraryExA(fpath.c_str(), NULL, LOAD_LIBRARY_SEARCH_DEFAULT_DIRS | LOAD_LIBRARY_SEARCH_DLL_LOAD_DIR);
+        std::string afpath = std::filesystem::absolute(fpath).string();
+        std::replace(afpath.begin(), afpath.end(), '/', '\\');
+        chandle = LoadLibraryExA(afpath.c_str(), NULL, LOAD_LIBRARY_SEARCH_DEFAULT_DIRS | LOAD_LIBRARY_SEARCH_DLL_LOAD_DIR);
         if (!chandle) {
             throw Exception("dlopen(\""+fpath+"\"): Error");
         }

gpt4all-backend/llamamodel.cpp

@@ -298,6 +298,8 @@ LLModel::Token LLamaModel::sampleToken(PromptContext &promptCtx) const
 
 bool LLamaModel::evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const
 {
+    llama_kv_cache_seq_rm(d_ptr->ctx, 0, ctx.n_past, -1);
+
     llama_batch batch = llama_batch_init(tokens.size(), 0, 1);
 
     batch.n_tokens = tokens.size();

gpt4all-chat/chat.cpp

@@ -10,14 +10,9 @@ Chat::Chat(QObject *parent)
     , m_id(Network::globalInstance()->generateUniqueId())
     , m_name(tr("New Chat"))
     , m_chatModel(new ChatModel(this))
-    , m_responseInProgress(false)
     , m_responseState(Chat::ResponseStopped)
     , m_creationDate(QDateTime::currentSecsSinceEpoch())
     , m_llmodel(new ChatLLM(this))
-    , m_isServer(false)
-    , m_shouldDeleteLater(false)
-    , m_isModelLoaded(false)
-    , m_shouldLoadModelWhenInstalled(false)
     , m_collectionModel(new LocalDocsCollectionsModel(this))
 {
     connectLLM();

gpt4all-chat/chat.h

@@ -155,15 +155,15 @@ private:
     QString m_response;
     QList<QString> m_collections;
     ChatModel *m_chatModel;
-    bool m_responseInProgress;
+    bool m_responseInProgress = false;
     ResponseState m_responseState;
     qint64 m_creationDate;
     ChatLLM *m_llmodel;
     QList<ResultInfo> m_databaseResults;
-    bool m_isServer;
-    bool m_shouldDeleteLater;
-    bool m_isModelLoaded;
-    bool m_shouldLoadModelWhenInstalled;
+    bool m_isServer = false;
+    bool m_shouldDeleteLater = false;
+    bool m_isModelLoaded = false;
+    bool m_shouldLoadModelWhenInstalled = false;
     LocalDocsCollectionsModel *m_collectionModel;
 };
 

gpt4all-chat/chatlistmodel.cpp

@@ -16,9 +16,6 @@ ChatListModel *ChatListModel::globalInstance()
 
 ChatListModel::ChatListModel()
     : QAbstractListModel(nullptr)
-    , m_newChat(nullptr)
-    , m_serverChat(nullptr)
-    , m_currentChat(nullptr)
 {
     addChat();
 

gpt4all-chat/chatlistmodel.h

@@ -239,9 +239,9 @@ private Q_SLOTS:
     }
 
 private:
-    Chat* m_newChat;
-    Chat* m_serverChat;
-    Chat* m_currentChat;
+    Chat* m_newChat = nullptr;
+    Chat* m_serverChat = nullptr;
+    Chat* m_currentChat = nullptr;
     QList<Chat*> m_chats;
 
 private: