prepare for selectable backends

exllama.cpp

@@ -0,0 +1,67 @@
+#include "exllama.h"
+
+#include <QJsonObject>
+#include <QJsonDocument>
+
+ExLlama::ExLlama()
+{
+	connect(&m_webSocket, &QWebSocket::textMessageReceived, this, &ExLlama::socketMessage);
+}
+
+bool ExLlama::ready()
+{
+	return m_webSocket.isValid();
+}
+
+void ExLlama::socketMessage(const QString& message)
+{
+	QJsonDocument jsonDocument = QJsonDocument::fromJson(message.toUtf8());
+	QJsonValue idVal = jsonDocument[QStringLiteral("request_id")];
+	if(!idVal.isDouble())
+	{
+		qDebug()<<"Got invalid response on socket";
+		return;
+	}
+	int id = idVal.toInt();
+
+	if(!isValidId(id))
+	{
+		qDebug()<<"Got unkown response id on socket";
+		return;
+	}
+
+	QJsonValue responseValue = jsonDocument[QStringLiteral("response")];
+	if(!responseValue.isString())
+	{
+		qDebug()<<"Got invalid response on socket";
+		return;
+	}
+
+	feedResponse(Response(responseValue.toString(), id, true));
+}
+
+void ExLlama::generateImpl(const Request& request)
+{
+	QJsonObject json;
+	json[QStringLiteral("action")] = QStringLiteral("infer");
+	json[QStringLiteral("request_id")] = static_cast<double>(request.getId());
+	json[QStringLiteral("text")] = request.getText();
+	json[QStringLiteral("max_new_tokens")] = 50;
+	json[QStringLiteral("stream")] = false;
+
+	QJsonDocument jsonDocument(json);
+	QString requestText = QString::fromUtf8(jsonDocument.toJson(QJsonDocument::JsonFormat::Compact));
+	qDebug()<<__func__<<' '<<requestText;
+	m_webSocket.sendTextMessage(requestText);
+}
+
+void ExLlama::open(const QUrl& url)
+{
+	m_webSocket.close();
+	m_webSocket.open(url);
+}
+
+ExLlama::~ExLlama()
+{
+	m_webSocket.disconnect();
+}