chat: report reason for fallback to CPU

gpt4all-chat/chat.cpp

@@ -57,6 +57,7 @@ void Chat::connectLLM()
     connect(m_llmodel, &ChatLLM::generatedNameChanged, this, &Chat::generatedNameChanged, Qt::QueuedConnection);
     connect(m_llmodel, &ChatLLM::reportSpeed, this, &Chat::handleTokenSpeedChanged, Qt::QueuedConnection);
     connect(m_llmodel, &ChatLLM::reportDevice, this, &Chat::handleDeviceChanged, Qt::QueuedConnection);
+    connect(m_llmodel, &ChatLLM::reportFallbackReason, this, &Chat::handleFallbackReasonChanged, Qt::QueuedConnection);
     connect(m_llmodel, &ChatLLM::databaseResultsChanged, this, &Chat::handleDatabaseResultsChanged, Qt::QueuedConnection);
     connect(m_llmodel, &ChatLLM::modelInfoChanged, this, &Chat::handleModelInfoChanged, Qt::QueuedConnection);
 
@@ -352,6 +353,12 @@ void Chat::handleDeviceChanged(const QString &device)
     emit deviceChanged();
 }
 
+void Chat::handleFallbackReasonChanged(const QString &fallbackReason)
+{
+    m_fallbackReason = fallbackReason;
+    emit fallbackReasonChanged();
+}
+
 void Chat::handleDatabaseResultsChanged(const QList<ResultInfo> &results)
 {
     m_databaseResults = results;

gpt4all-chat/chat.h

@@ -26,6 +26,7 @@ class Chat : public QObject
     Q_PROPERTY(QString modelLoadingError READ modelLoadingError NOTIFY modelLoadingErrorChanged)
     Q_PROPERTY(QString tokenSpeed READ tokenSpeed NOTIFY tokenSpeedChanged);
     Q_PROPERTY(QString device READ device NOTIFY deviceChanged);
+    Q_PROPERTY(QString fallbackReason READ fallbackReason NOTIFY fallbackReasonChanged);
     QML_ELEMENT
     QML_UNCREATABLE("Only creatable from c++!")
 
@@ -90,6 +91,7 @@ public:
 
     QString tokenSpeed() const { return m_tokenSpeed; }
     QString device() const { return m_device; }
+    QString fallbackReason() const { return m_fallbackReason; }
 
 public Q_SLOTS:
     void serverNewPromptResponsePair(const QString &prompt);
@@ -118,6 +120,7 @@ Q_SIGNALS:
     void collectionListChanged(const QList<QString> &collectionList);
     void tokenSpeedChanged();
     void deviceChanged();
+    void fallbackReasonChanged();
 
 private Q_SLOTS:
     void handleResponseChanged(const QString &response);
@@ -129,6 +132,7 @@ private Q_SLOTS:
     void handleModelLoadingError(const QString &error);
     void handleTokenSpeedChanged(const QString &tokenSpeed);
     void handleDeviceChanged(const QString &device);
+    void handleFallbackReasonChanged(const QString &device);
     void handleDatabaseResultsChanged(const QList<ResultInfo> &results);
     void handleModelInfoChanged(const ModelInfo &modelInfo);
     void handleModelInstalled();
@@ -142,6 +146,7 @@ private:
     QString m_modelLoadingError;
     QString m_tokenSpeed;
     QString m_device;
+    QString m_fallbackReason;
     QString m_response;
     QList<QString> m_collections;
     ChatModel *m_chatModel;

gpt4all-chat/chatllm.cpp

@@ -267,27 +267,46 @@ bool ChatLLM::loadModel(const ModelInfo &modelInfo)
                 if (requestedDevice != "CPU") {
                     const size_t requiredMemory = m_llModelInfo.model->requiredMem(filePath.toStdString());
                     std::vector<LLModel::GPUDevice> availableDevices = m_llModelInfo.model->availableGPUDevices(requiredMemory);
+                    LLModel::GPUDevice *device = nullptr;
+
                     if (!availableDevices.empty() && requestedDevice == "Auto" && availableDevices.front().type == 2 /*a discrete gpu*/) {
-                        m_llModelInfo.model->initializeGPUDevice(availableDevices.front());
-                        actualDevice = QString::fromStdString(availableDevices.front().name);
+                        device = &availableDevices.front();
                     } else {
                         for (LLModel::GPUDevice &d : availableDevices) {
                             if (QString::fromStdString(d.name) == requestedDevice) {
-                                m_llModelInfo.model->initializeGPUDevice(d);
-                                actualDevice = QString::fromStdString(d.name);
+                                device = &d;
                                 break;
                             }
                         }
                     }
+
+                    if (!device) {
+                        emit reportFallbackReason("<br>Using CPU: device not found");
+                    } else if (!m_llModelInfo.model->initializeGPUDevice(*device)) {
+                        emit reportFallbackReason("<br>Using CPU: failed to init device");
+                    } else {
+                        actualDevice = QString::fromStdString(device->name);
+                    }
                 }
 
                 // Report which device we're actually using
                 emit reportDevice(actualDevice);
 
                 bool success = m_llModelInfo.model->loadModel(filePath.toStdString());
-                if (!success && actualDevice != "CPU") {
+                if (actualDevice == "CPU") {
+                    // we asked llama.cpp to use the CPU
+                } else if (!success) {
+                    // llama_init_from_file returned nullptr
+                    // this may happen because ggml_metal_add_buffer failed
                     emit reportDevice("CPU");
+                    emit reportFallbackReason("<br>Using CPU: llama_init_from_file failed");
                     success = m_llModelInfo.model->loadModel(filePath.toStdString());
+                } else if (!m_llModelInfo.model->usingGPUDevice()) {
+                    // ggml_vk_init was not called in llama.cpp
+                    // We might have had to fallback to CPU after load if the model is not possible to accelerate
+                    // for instance if the quantization method is not supported on Vulkan yet
+                    emit reportDevice("CPU");
+                    emit reportFallbackReason("<br>Using CPU: unsupported quantization type");
                 }
 
                 MySettings::globalInstance()->setAttemptModelLoad(QString());
@@ -299,11 +318,6 @@ bool ChatLLM::loadModel(const ModelInfo &modelInfo)
                     m_llModelInfo = LLModelInfo();
                     emit modelLoadingError(QString("Could not load model due to invalid model file for %1").arg(modelInfo.filename()));
                 } else {
-                    // We might have had to fallback to CPU after load if the model is not possible to accelerate
-                    // for instance if the quantization method is not supported on Vulkan yet
-                    if (actualDevice != "CPU" && !m_llModelInfo.model->usingGPUDevice())
-                        emit reportDevice("CPU");
-
                     switch (m_llModelInfo.model->implementation().modelType()[0]) {
                     case 'L': m_llModelType = LLModelType::LLAMA_; break;
                     case 'G': m_llModelType = LLModelType::GPTJ_; break;

gpt4all-chat/chatllm.h

@@ -127,6 +127,7 @@ Q_SIGNALS:
     void requestRetrieveFromDB(const QList<QString> &collections, const QString &text, int retrievalSize, QList<ResultInfo> *results);
     void reportSpeed(const QString &speed);
     void reportDevice(const QString &device);
+    void reportFallbackReason(const QString &fallbackReason);
     void databaseResultsChanged(const QList<ResultInfo>&);
     void modelInfoChanged(const ModelInfo &modelInfo);
 

gpt4all-chat/main.qml

@@ -1013,7 +1013,7 @@ Window {
             anchors.rightMargin: 30
             color: theme.mutedTextColor
             visible: currentChat.tokenSpeed !== ""
-            text: qsTr("Speed: ") + currentChat.tokenSpeed + "<br>" + qsTr("Device: ") + currentChat.device
+            text: qsTr("Speed: ") + currentChat.tokenSpeed + "<br>" + qsTr("Device: ") + currentChat.device + currentChat.fallbackReason
             font.pixelSize: theme.fontSizeLarge
         }