How Janction Efficiently Stores AI/ML Models for Different Users？

Since there are three main categories of AI/ML models, traditional machine learning models (linear regression, decision trees, support vector machines) with storage spaces generally ranging from hundreds of KB to tens of MB; tens of MB to several GB Deep learning models (depending on the depth, width, and number of parameters of the network. For example, a smaller convolutional neural network (CNN) may be tens to hundreds of MB in size, while a large deep neural network (DNN) may will exceed hundreds of MB or even several GB); large language models (LLM) from hundreds of MB to tens of GB, depending on the size of the model and the number of parameters. We use a set of highly reliable and scalable file storage systems and memory caching mechanisms to ensure that models with storage space less than 4GB (specific threshold recommendations and technical confirmations) are used for persistent storage, and can be intelligently preloaded into the memory of GPU instances. middle. For the model itself, we will also use some traditional compression techniques. In addition, based on the region where the GPUs we aggregate are located, we will make corresponding CDN configurations for the model. As for the large language model, due to the huge storage space, we are still examining the technical solutions. We will probably choose a distributed file storage solution, such as Hadoop Distributed File System. At the same time, we will use distributed storage solutions, such as Redis, Memcached, etc.