Much better memory mgmt for multi-threaded model loading/unloading.

2025-10-08 22:43:49 +00:00 · 2023-05-13 19:05:35 -04:00
parent 2989b74d43
commit ddc24acf33
6 changed files with 243 additions and 74 deletions
--- a/gpt4all-chat/chatllm.cpp
+++ b/gpt4all-chat/chatllm.cpp
@@ -15,6 +15,7 @@
 #include <fstream>

 //#define DEBUG
+//#define DEBUG_MODEL_LOADING

 #define MPT_INTERNAL_STATE_VERSION 0
 #define GPTJ_INTERNAL_STATE_VERSION 0
@@ -37,9 +38,51 @@ static QString modelFilePath(const QString &modelName)
    return QString();
 }

+class LLModelStore {
+public:
+    static LLModelStore *globalInstance();
+
+    LLModelInfo acquireModel(); // will block until llmodel is ready
+    void releaseModel(const LLModelInfo &info); // must be called when you are done
+
+private:
+    LLModelStore()
+    {
+        // seed with empty model
+        m_availableModels.append(LLModelInfo());
+    }
+    ~LLModelStore() {}
+    QVector<LLModelInfo> m_availableModels;
+    QMutex m_mutex;
+    QWaitCondition m_condition;
+    friend class MyLLModelStore;
+};
+
+class MyLLModelStore : public LLModelStore { };
+Q_GLOBAL_STATIC(MyLLModelStore, storeInstance)
+LLModelStore *LLModelStore::globalInstance()
+{
+    return storeInstance();
+}
+
+LLModelInfo LLModelStore::acquireModel()
+{
+    QMutexLocker locker(&m_mutex);
+    while (m_availableModels.isEmpty())
+        m_condition.wait(locker.mutex());
+    return m_availableModels.takeFirst();
+}
+
+void LLModelStore::releaseModel(const LLModelInfo &info)
+{
+    QMutexLocker locker(&m_mutex);
+    m_availableModels.append(info);
+    Q_ASSERT(m_availableModels.count() < 2);
+    m_condition.wakeAll();
+}
+
 ChatLLM::ChatLLM(Chat *parent)
    : QObject{nullptr}
-    , m_llmodel(nullptr)
    , m_promptResponseTokens(0)
    , m_promptTokens(0)
    , m_responseLogits(0)
@@ -49,6 +92,7 @@ ChatLLM::ChatLLM(Chat *parent)
    moveToThread(&m_llmThread);
    connect(this, &ChatLLM::sendStartup, Network::globalInstance(), &Network::sendStartup);
    connect(this, &ChatLLM::sendModelLoaded, Network::globalInstance(), &Network::sendModelLoaded);
+    connect(this, &ChatLLM::shouldBeLoadedChanged, this, &ChatLLM::handleShouldBeLoadedChanged, Qt::QueuedConnection);
    connect(m_chat, &Chat::idChanged, this, &ChatLLM::handleChatIdChanged);
    connect(&m_llmThread, &QThread::started, this, &ChatLLM::threadStarted);
    m_llmThread.setObjectName(m_chat->id());
@@ -59,7 +103,13 @@ ChatLLM::~ChatLLM()
 {
    m_llmThread.quit();
    m_llmThread.wait();
-    delete m_llmodel;
+
+    // The only time we should have a model loaded here is on shutdown
+    // as we explicitly unload the model in all other circumstances
+    if (isModelLoaded()) {
+        delete m_modelInfo.model;
+        m_modelInfo.model = nullptr;
+    }
 }

 bool ChatLLM::loadDefaultModel()
@@ -76,50 +126,103 @@ bool ChatLLM::loadDefaultModel()

 bool ChatLLM::loadModel(const QString &modelName)
 {
+    // This is a complicated method because N different possible threads are interested in the outcome
+    // of this method. Why? Because we have a main/gui thread trying to monitor the state of N different
+    // possible chat threads all vying for a single resource - the currently loaded model - as the user
+    // switches back and forth between chats. It is important for our main/gui thread to never block
+    // but simultaneously always have up2date information with regards to which chat has the model loaded
+    // and what the type and name of that model is. I've tried to comment extensively in this method
+    // to provide an overview of what we're doing here.
+
+    // We're already loaded with this model
    if (isModelLoaded() && m_modelName == modelName)
        return true;

-    if (isModelLoaded()) {
+    QString filePath = modelFilePath(modelName);
+    QFileInfo fileInfo(filePath);
+
+    // We have a live model, but it isn't the one we want
+    bool alreadyAcquired = isModelLoaded();
+    if (alreadyAcquired) {
        resetContextProtected();
-        delete m_llmodel;
-        m_llmodel = nullptr;
+#if defined(DEBUG_MODEL_LOADING)
+        qDebug() << "already acquired model deleted" << m_chat->id() << m_modelInfo.model;
+#endif
+        delete m_modelInfo.model;
+        m_modelInfo.model = nullptr;
        emit isModelLoadedChanged();
+    } else {
+        // This is a blocking call that tries to retrieve the model we need from the model store.
+        // If it succeeds, then we just have to restore state. If the store has never had a model
+        // returned to it, then the modelInfo.model pointer should be null which will happen on startup
+        m_modelInfo = LLModelStore::globalInstance()->acquireModel();
+#if defined(DEBUG_MODEL_LOADING)
+        qDebug() << "acquired model from store" << m_chat->id() << m_modelInfo.model;
+#endif
+        // At this point it is possible that while we were blocked waiting to acquire the model from the
+        // store, that our state was changed to not be loaded. If this is the case, release the model
+        // back into the store and quit loading
+        if (!m_shouldBeLoaded) {
+            qDebug() << "no longer need model" << m_chat->id() << m_modelInfo.model;
+            LLModelStore::globalInstance()->releaseModel(m_modelInfo);
+            m_modelInfo = LLModelInfo();
+            emit isModelLoadedChanged();
+            return false;
+        }
+
+        // Check if the store just gave us exactly the model we were looking for
+        if (m_modelInfo.model && m_modelInfo.fileInfo == fileInfo) {
+#if defined(DEBUG_MODEL_LOADING)
+            qDebug() << "store had our model" << m_chat->id() << m_modelInfo.model;
+#endif
+            restoreState();
+            emit isModelLoadedChanged();
+            return true;
+        } else {
+            // Release the memory since we have to switch to a different model.
+#if defined(DEBUG_MODEL_LOADING)
+            qDebug() << "deleting model" << m_chat->id() << m_modelInfo.model;
+#endif
+            delete m_modelInfo.model;
+            m_modelInfo.model = nullptr;
+        }
    }

-    bool isGPTJ = false;
-    bool isMPT = false;
-    QString filePath = modelFilePath(modelName);
-    QFileInfo info(filePath);
-    if (info.exists()) {
+    // Guarantee we've released the previous models memory
+    Q_ASSERT(!m_modelInfo.model);

+    // Store the file info in the modelInfo in case we have an error loading
+    m_modelInfo.fileInfo = fileInfo;
+
+    if (fileInfo.exists()) {
        auto fin = std::ifstream(filePath.toStdString(), std::ios::binary);
        uint32_t magic;
        fin.read((char *) &magic, sizeof(magic));
        fin.seekg(0);
        fin.close();
-        isGPTJ = magic == 0x67676d6c;
-        isMPT = magic == 0x67676d6d;
+        const bool isGPTJ = magic == 0x67676d6c;
+        const bool isMPT = magic == 0x67676d6d;
        if (isGPTJ) {
-            m_modelType = ModelType::GPTJ_;
-            m_llmodel = new GPTJ;
-            m_llmodel->loadModel(filePath.toStdString());
+            m_modelType = LLModelType::GPTJ_;
+            m_modelInfo.model = new GPTJ;
+            m_modelInfo.model->loadModel(filePath.toStdString());
        } else if (isMPT) {
-            m_modelType = ModelType::MPT_;
-            m_llmodel = new MPT;
-            m_llmodel->loadModel(filePath.toStdString());
+            m_modelType = LLModelType::MPT_;
+            m_modelInfo.model = new MPT;
+            m_modelInfo.model->loadModel(filePath.toStdString());
        } else {
-            m_modelType = ModelType::LLAMA_;
-            m_llmodel = new LLamaModel;
-            m_llmodel->loadModel(filePath.toStdString());
+            m_modelType = LLModelType::LLAMA_;
+            m_modelInfo.model = new LLamaModel;
+            m_modelInfo.model->loadModel(filePath.toStdString());
        }
-
-        restoreState();
-
-#if defined(DEBUG)
-    qDebug() << "chatllm modelLoadedChanged" << m_chat->id();
-    fflush(stdout);
+#if defined(DEBUG_MODEL_LOADING)
+        qDebug() << "new model" << m_chat->id() << m_modelInfo.model;
+#endif
+        restoreState();
+#if defined(DEBUG)
+        qDebug() << "modelLoadedChanged" << m_chat->id();
+        fflush(stdout);
 #endif
-
        emit isModelLoadedChanged();

        static bool isFirstLoad = true;
@@ -129,19 +232,20 @@ bool ChatLLM::loadModel(const QString &modelName)
        } else
            emit sendModelLoaded();
    } else {
+        LLModelStore::globalInstance()->releaseModel(m_modelInfo); // release back into the store
        const QString error = QString("Could not find model %1").arg(modelName);
        emit modelLoadingError(error);
    }

-    if (m_llmodel)
-        setModelName(info.completeBaseName().remove(0, 5)); // remove the ggml- prefix
+    if (m_modelInfo.model)
+        setModelName(fileInfo.completeBaseName().remove(0, 5)); // remove the ggml- prefix

-    return m_llmodel;
+    return m_modelInfo.model;
 }

 bool ChatLLM::isModelLoaded() const
 {
-    return m_llmodel && m_llmodel->isModelLoaded();
+    return m_modelInfo.model && m_modelInfo.model->isModelLoaded();
 }

 void ChatLLM::regenerateResponse()
@@ -226,7 +330,7 @@ bool ChatLLM::handlePrompt(int32_t token)
    // m_promptResponseTokens and m_responseLogits are related to last prompt/response not
    // the entire context window which we can reset on regenerate prompt
 #if defined(DEBUG)
-    qDebug() << "chatllm prompt process" << m_chat->id() << token;
+    qDebug() << "prompt process" << m_chat->id() << token;
 #endif
    ++m_promptTokens;
    ++m_promptResponseTokens;
@@ -287,12 +391,12 @@ bool ChatLLM::prompt(const QString &prompt, const QString &prompt_template, int3
    m_ctx.n_batch = n_batch;
    m_ctx.repeat_penalty = repeat_penalty;
    m_ctx.repeat_last_n = repeat_penalty_tokens;
-    m_llmodel->setThreadCount(n_threads);
+    m_modelInfo.model->setThreadCount(n_threads);
 #if defined(DEBUG)
    printf("%s", qPrintable(instructPrompt));
    fflush(stdout);
 #endif
-    m_llmodel->prompt(instructPrompt.toStdString(), promptFunc, responseFunc, recalcFunc, m_ctx);
+    m_modelInfo.model->prompt(instructPrompt.toStdString(), promptFunc, responseFunc, recalcFunc, m_ctx);
 #if defined(DEBUG)
    printf("\n");
    fflush(stdout);
@@ -307,26 +411,55 @@ bool ChatLLM::prompt(const QString &prompt, const QString &prompt_template, int3
    return true;
 }

+void ChatLLM::setShouldBeLoaded(bool b)
+{
+#if defined(DEBUG_MODEL_LOADING)
+    qDebug() << "setShouldBeLoaded" << m_chat->id() << b << m_modelInfo.model;
+#endif
+    m_shouldBeLoaded = b; // atomic
+    emit shouldBeLoadedChanged();
+}
+
+void ChatLLM::handleShouldBeLoadedChanged()
+{
+    if (m_shouldBeLoaded)
+        reloadModel();
+    else
+        unloadModel();
+}
+
+void ChatLLM::forceUnloadModel()
+{
+    m_shouldBeLoaded = false; // atomic
+    unloadModel();
+}
+
 void ChatLLM::unloadModel()
 {
-#if defined(DEBUG)
-    qDebug() << "chatllm unloadModel" << m_chat->id();
-#endif
+    if (!isModelLoaded())
+        return;
+
    saveState();
-    delete m_llmodel;
-    m_llmodel = nullptr;
+#if defined(DEBUG_MODEL_LOADING)
+    qDebug() << "unloadModel" << m_chat->id() << m_modelInfo.model;
+#endif
+    LLModelStore::globalInstance()->releaseModel(m_modelInfo);
+    m_modelInfo = LLModelInfo();
    emit isModelLoadedChanged();
 }

-void ChatLLM::reloadModel(const QString &modelName)
+void ChatLLM::reloadModel()
 {
-#if defined(DEBUG)
-    qDebug() << "chatllm reloadModel" << m_chat->id();
+    if (isModelLoaded())
+        return;
+
+#if defined(DEBUG_MODEL_LOADING)
+    qDebug() << "reloadModel" << m_chat->id() << m_modelInfo.model;
 #endif
-    if (modelName.isEmpty()) {
+    if (m_modelName.isEmpty()) {
        loadDefaultModel();
    } else {
-        loadModel(modelName);
+        loadModel(m_modelName);
    }
 }

@@ -348,7 +481,7 @@ void ChatLLM::generateName()
    printf("%s", qPrintable(instructPrompt));
    fflush(stdout);
 #endif
-    m_llmodel->prompt(instructPrompt.toStdString(), promptFunc, responseFunc, recalcFunc, ctx);
+    m_modelInfo.model->prompt(instructPrompt.toStdString(), promptFunc, responseFunc, recalcFunc, ctx);
 #if defined(DEBUG)
    printf("\n");
    fflush(stdout);
@@ -415,7 +548,7 @@ bool ChatLLM::serialize(QDataStream &stream, int version)
    QByteArray compressed = qCompress(m_state);
    stream << compressed;
 #if defined(DEBUG)
-    qDebug() << "chatllm serialize" << m_chat->id() << m_state.size();
+    qDebug() << "serialize" << m_chat->id() << m_state.size();
 #endif
    return stream.status() == QDataStream::Ok;
 }
@@ -452,7 +585,7 @@ bool ChatLLM::deserialize(QDataStream &stream, int version)
        stream >> m_state;
    }
 #if defined(DEBUG)
-    qDebug() << "chatllm deserialize" << m_chat->id();
+    qDebug() << "deserialize" << m_chat->id();
 #endif
    return stream.status() == QDataStream::Ok;
 }
@@ -462,12 +595,12 @@ void ChatLLM::saveState()
    if (!isModelLoaded())
        return;

-    const size_t stateSize = m_llmodel->stateSize();
+    const size_t stateSize = m_modelInfo.model->stateSize();
    m_state.resize(stateSize);
 #if defined(DEBUG)
-    qDebug() << "chatllm saveState" << m_chat->id() << "size:" << m_state.size();
+    qDebug() << "saveState" << m_chat->id() << "size:" << m_state.size();
 #endif
-    m_llmodel->saveState(static_cast<uint8_t*>(reinterpret_cast<void*>(m_state.data())));
+    m_modelInfo.model->saveState(static_cast<uint8_t*>(reinterpret_cast<void*>(m_state.data())));
 }

 void ChatLLM::restoreState()
@@ -476,9 +609,9 @@ void ChatLLM::restoreState()
        return;

 #if defined(DEBUG)
-    qDebug() << "chatllm restoreState" << m_chat->id() << "size:" << m_state.size();
+    qDebug() << "restoreState" << m_chat->id() << "size:" << m_state.size();
 #endif
-    m_llmodel->restoreState(static_cast<const uint8_t*>(reinterpret_cast<void*>(m_state.data())));
+    m_modelInfo.model->restoreState(static_cast<const uint8_t*>(reinterpret_cast<void*>(m_state.data())));
    m_state.clear();
    m_state.resize(0);
 }