Cloud.Sitez - : BABA

AI growth is built upon infrastructure modernization: How businesses must adapt to thrive

admin — Thu, 09 Jan 2025 08:43:11 +0100

It doesn’t take a genius to work out that artificial intelligence (AI) is transforming industries at an unprecedented rate. According to IDC, global spending on AI is expected to reach $632 billion by 2028, with generative AI (GenAI) growing at a remarkable annual rate of 59.2%. Yet, as AI capabilities surge, the infrastructure needed to […]

Alibaba Cloud Enhances Global Partner Ecosystem to Drive AI Adoption

admin — Thu, 09 Jan 2025 08:43:09 +0100

Alibaba Cloud, the cloud computing arm of Alibaba Group, today has unveiled an initiative to revamp its global partnership ecosystem, aiming to accelerate the adoption of AI and cloud computing solutions across industries. The “Alibaba Cloud Partner Rainforest Plan,” unveiled at the Alibaba Cloud Partner Summit 2024, includes several new programs designed to empower partners […]

First phase of Athlete365 Business Accelerator Programme is launched with support from Alibaba.com

admin — Thu, 09 Jan 2025 08:43:06 +0100

To assist athletes in transitioning from sports to business, the International Olympic Committee (IOC) has partnered with Alibaba.com to launch the first phase of the Athlete365 Business Accelerator Programme. The initiative aims to provide current and former elite athletes with the tools and support necessary to pursue entrepreneurial ventures. Alibaba.com is the first worldwide Olympic […]

Strengthening Security in the AI Era: Alibaba Cloud Showcases Security Solutions for Diverse Cloud Environments

admin — Thu, 09 Jan 2025 08:43:02 +0100

In response to the growing trend of organizations adopting multi-cloud and hybrid cloud environments, where data is distributed across various platforms and accessed through a multitude of devices and networks, Alibaba Cloud has recently announced new features for its unified security solution and content delivery service. A recent survey from Statista revealed that as of […]

News roundup: Alibaba Cloud’s Flink 2.0 Breakthroughs, AliExpress’s Black Friday Triumph, and Cainiao’s Logistics Surge

admin — Thu, 09 Jan 2025 08:43:00 +0100

This week, Alibaba Cloud and its ecosystem highlighted groundbreaking advancements and global achievements across multiple sectors. At Flink Forward Asia, Alibaba Cloud unveiled Apache Flink 2.0, introducing unified batch and stream processing, scalable storage solutions, and real-time analytics to empower the GenAI era. A strategic partnership with DreamSmart Group integrates Alibaba’s AI model Qwen into […]

News Roundup: Infinity Nikki’s Debut Hit, Gen Z Spending Trends, and Exclusive 88VIP Perks

admin — Thu, 09 Jan 2025 08:42:58 +0100

This week, we spotlight Alibaba’s latest achievements across gaming partnership, e-commerce, and membership services. Infold Games, powered by Alibaba Cloud, launched the highly successful multi-platform game Infinity Nikki, reaching over 10 million downloads in just four days. On the retail front, Taobao Double 12 captured the attention of young consumers with its innovative “Treasure Hunt” […]

Alibaba Cloud Unveils New Research Model for Enhanced Visual Reasoning

admin — Thu, 09 Jan 2025 08:42:53 +0100

Alibaba Cloud has recently introduced QVQ-72B-Preview (“QVQ”), an open-sourced, experimental research model designed to advance visual reasoning capabilities. QVQ is an open-weight model for multimodal reasoning that has delivered exceptional performance across various benchmarks. Notably, it achieved an impressive score of 70.3% on the Multimodal Massive Multi-task Understanding (MMMU) benchmark, underscoring its strong multidisciplinary understanding […]

News Roundup: Empowering AI Innovation and Celebrating Scientific Excellence

admin — Thu, 09 Jan 2025 08:42:53 +0100

This week, the AEF NextGen Fund, a new $150 million initiative from the Alibaba Entrepreneurs Fund, aims to accelerate AI-powered startups across sectors like healthcare and financial services. With a focus on sustainability and global scaling, this fund strengthens Alibaba’s commitment to fostering innovation in Hong Kong and the Greater Bay Area. In parallel, the […]

2024 Year in Review: A Journey of Rethinking and Refocusing

admin — Thu, 09 Jan 2025 08:42:51 +0100

For many years, capturing Alibaba’s multifaceted impact has been challenging. However, in 2024, the company’s transformative initiatives and resilience took center stage, solidifying its place as one of the powerhouses in sustainability, AI innovation, and e-commerce expansion. With a laser-sharp focus on its core strategic pillars—“E-Commerce + AI”—Alibaba didn’t reshape the landscape of commerce and […]

Alibaba Cloud recognized as an Emerging Leader in the Gartner® “Emerging Market Quadrants”

admin — Thu, 09 Jan 2025 08:42:48 +0100

Alibaba Cloud, the cloud computing arm of Alibaba Group, has been named an Emerging Leader in all four quadrants evaluated in the latest Gartner® “Innovation Guide for Generative AI Technologies” report, issued last month. According to Gartner: “The Emerging Market Quadrant is a pilot initiated in June 2024 to provide a visualization of generative AI […]

Data Insights in Cybersecurity Part 1: Intro to ssdeep and XOR.DDoS Case Study

admin — Thu, 09 Jan 2025 08:42:35 +0100

By Yuriy Yuzifovich , Head of Security Innovation Lab (SIL) of Alibaba Cloud,Thanh Nguyen, Principal Data Scientist at SIL, andAnastasia Poliakova, Sr. Security Engineer, SILIntroductionCyber-attacks are becoming more sophisticated, and more automated. With modular code reuse and minor file tampering to generate different binary files to bypass rule-based or signature-based methods, the number of malware samples keeps accelerating and making it very difficult for the cybersecurity community to detect, track, and report these variations. In some extreme cases of highly-variable malware, these variations are purely random simply to generate random hash values, and as such each of these individual variants is not worth of even tracking since it will never repeat in a different machine. The industry has multiple solutions to this problem, from behavioral analysis to tracking and finding anomalies in machine access (zero trust), to name a few.At Security Innovation Labs (SIL), we use multiple behavioral AI methods, including analysis of traffic (DNS, HTTP, IP connections), automated reasoning combining multiple data points, and various types of anomalies. Recently, we had breakthrough results with fuzzy hashing, that at the massive scale of Alibaba Cloud produced insights that were never possible to achieve with other methods. In this series of blog entries, we will describe fuzzy hashing in the context of malware detection and classification in Alibaba Cloud.We also strongly believe in the power of the cybersecurity community to share both direct threat intelligence data, as well as the techniques that makes us, cybersecurity industry, stronger against the bad actors. This is why we presented our findings at BotConf in April 2022 in Nantes, France, and then followed up with publishing our paper “Detecting emerging malware in the cloud before VirusTotal can see it” in The Journal on Cybercrime & Digital InvestigationsFollowing up on the successful presentation, and to respond to a number of questions we received during the conference about our work, we decided to create this series of blog postings to talk about our ssdeep-based validation progress we made. To keep the audience properly entertained, we will supply a number of visualizations, as visual representation of the cybersecurity information is a powerful, yet underappreciated tool of data exploration and decision making.In this article, we will describe the limitations of classical hashes, provide a brief overview of fuzzy hashing and our choice, ssdeep, and show results with XOR.DDoS detection, a popular Linux Trojan malware. In the subsequent entries, we dive deeper into distances between fuzzy hashes, engineering challenges we overcame, how we built our graph, and what results we were able to achieve. The Data Insights in Cybersecurity blog series will not stop there; observing the malware at the cloud scale never leaves at the lack of the insights to share.Problem StatementIn a perfect world, we can easily keep malicious binaries intact, create pairwise similarities between the binaries, and using the perfect distance metric, connect highly similar binaries together, effectively building a graph. We could also easily detect a new variant by comparing a suspect against the whole database of existing binaries to find a close match, thus enabling immediate detection of most popular malware families.But the high volume of the unknown binaries makes this idea impossible to implement. Malware can morph, creating a unique binary for every infected machine, and such a full library of binaries will have a lot of noise, not to mention a lot of resources it will consume, and will carry little value compared to the investment. Can we reduce the binaries to a small digital signature that uniquely represent it but without consuming too much disk space? This is where cryptographic hashes come in, and SHA256 hash is one of the most widely used to keep track of the new malware.Why Fuzzy Hash?The cryptographic hash algorithms are widely used for integrity check and allow to verify if two binary files are identical without directly comparing them. The very nature of a good cryptographic hash, and SHA256 is certainly good, is that the value of the hash does not communicate anything about the hashed information. This wonderful property makes it impossible to compare binaries using hashes, other than declaring if it’s exactly the same binary or not. Even a change of a single byte makes it a completely different SHA value.The graph below is a sample visualization of what the connectivity between binaries, represented as hash values, and process names, looks like in the wild, as observed in the cloud (we replaced actual process names and sha256 values to improve visibility of the graph shapes). As you can see, one SHA256 value (binary) might have different paths of the process referring to it (on different cloud machines). Sometimes the same process path can correspond to d

A Short Exploration of Java Class Pre-Initialization

admin — Thu, 09 Jan 2025 08:42:35 +0100

*By Qingfeng*1. BackgroundJava applications are expected to start quickly because of the advent of the cloud-native era. It is becoming increasingly popular to reduce resource costs by dynamically scaling and using serverless computing. Alibaba has done a lot of work and exploration to meet the demands for quick start of applications in Serverless computing scenarios, including AppCDS, Ahead of Time Compilation (AOT), fast class indexing(JarIndex), class pre-initialization, and other technologies to optimize language runtime. Users can obtain up to three times the startup performance improvement without modifying any code. One of the cutting-edge technologies is class pre-initialization.2. Class Pre-Initialization2.1 MotivationProfiling data on Java startup stage shows that the main reason for Javas’ slow startup is that it takes much time to load, link, and initialize classes, which we aim to break one by one.Class Fidning: Fast class indexing can find corresponding Jar file of class in O(1) timeClass loading and linkage: The help of AppCDS technology can reduce time consumption and speed up the startup for load and link operations.Class initialization: For the initialization, we observed that if the initialization of classes has no side effects, it is possible to skip the execution of the initialization of these classes. Consider the following code:public class Foo { private static HashMap cache; static { if (cache == null) { cache = new HashMap(); } for (int i = 0; i < 1024; i++) { cache.put(i, 0); } }}No matter when, no matter how many times, the result of initialization of the Foo class is the same: creating a hash table with a size of 1024 as a cache will not affect the external environment. In such cases, we can apply class pre-initialization, i.e. dump cache objects into CDS images, and directly map cache objects in CDS images to the Java G1 heap when JVM starts (8042668: Provide GC support for shared heap ranges in Class Data Sharing). When the program runs, skip the static code block initialization of the Foo class to speed up the startup:public class Foo { private static HashMap cache; // cache is directly materialized from G1 archive region static { // skip execution }}Another typical example is java.lang.Integer$IntegerCache class. It is an excellent candidate for class pre-initialization, which always creating a range of integer cache in [-128,127].2.2 Class Pre-Initialization DetailsThe native class pre-initialization mechanism requires explicit object materialization calls (jdk.internal.misc.VM.initializeFromArchive(…)) on certain initialization point and only supports a few restricted classes, which are hard-coded in the JVM code and cannot be extended. Alibaba and Google have proposed the Eclipse Adoptium FastStartup Incubator project. They aim to explore the Java quick start technologies, including (but not limited to) class pre-initialization. As aforementioned requirements, only the initialization phase that does not take extra side effect is able to apply class pre-initialization. Alibaba and Google have explored and contributed two approaches for safe class pre-initialization.Provides a Java annotation (jdk.internal.vm.annotation.Preserve), which allows more safe classes to be pre-initialized through manual labeling.Add new JVM option to accept a list of safe classes. This file can be generated by static analysis tools. The static analysis tool is based on GraalVM, it scans all class initialization blocks and construct their call graph, and do further data-flow analysis on that.At the same time, we have added a security check mechanism of class pre-initialization to the JVM to ensure the virtual machine can still work normally in the worst case. The whole workflow is shown in the figure below:2.3 EvaluationClass pre-initialization is based on AppCDS, performances evaluation concerns about AppCDS and itself. We found that about 90% (1800/2000) classes can be pre-initialized safely in a better scenario, average startup performance is improved by 19.2%.In a larger-scale evaluation, we found that class pre-initialization is slightly better than AppCDS, with a 5% performance improvement.This is because some classes with high time consumption on initialization usually have side effects. They cannot be pre-initialized, but the pre-initialization of these classes will still be performed in common paths, and these hot paths will cause a performance penalty.3. ConclusionThere have been three carriages for Java quick start for a long time: OpenJDK CRaC, JVM Runtime optimizations(AppCDS-based optimizations, AOT, JarIndex, etc), and static compilation (OpenJDK Leyden). All of them are committed to optimizing Java application startup performance from different directions. Alibaba has achieved good results in the second direction. We will continue moving forward and continuously optimize the startup performance of Java applications.

Igniting the AI Revolution — A Journey with Qwen, RAG, and LangChain

admin — Thu, 09 Jan 2025 08:42:34 +0100

Igniting the AI Revolution — A Journey with Qwen, RAG, and LangChainBy FarruhIn the era of Artificial Intelligence (AI), extracting meaningful knowledge from vast datasets has become critical for both businesses and individuals. Enter Retrieval-Augmented Generation (RAG), a breakthrough that has turbocharged the capabilities of AI, empowering systems to not only generate human-like text but also pull in relevant information in real-time. This fusion produces responses that are both rich in context and precise in detail.As we set sail on the exciting voyage through the vast ocean of Artificial Intelligence (AI), it’s essential to understand the three pillars that will be our guiding stars: Generative AI, Large Language Models (LLMs), LangChain, Hugging Face, and the useful application on this RAG (Retrieval-Augmented Generation).Large Language Models and Generative AI: The Engines of InnovationAt the core of our journey lie Large Language Models (LLMs) and Generative AI — two potent engines driving the innovation vessel forward.Large Language Models (LLMs)LLMs, such as Qwen, GPT, and others, are the titans of text, capable of understanding and generating human-like language on a massive scale. These models have been trained on extensive corpora of text data, allowing them to predict and produce coherent and contextually relevant strings of text. They are the backbone of many natural language processing tasks, from translation to content creation.Generative AI (GenAI)Generative AI is the artful wizard of creation within the AI realm. It encompasses technologies that generate new data instances that resemble the training data, such as images, music, and, most importantly for our voyage, text. In our context, Generative AI refers to the ability of AI to craft novel and informative responses, stories, or ideas that have never been seen before. It enables AI to not just mimic the past but to invent, innovate, and inspire.LangChain: Orchestrating Your AI SymphonyLangChain serves as the architect of our AI workflow, meticulously designing the structure that allows for seamless integration and interaction between various AI components. This framework simplifies the complex process of chaining together data flow from intelligent subsystems, including LLMs and retrieval systems, making tasks such as information extraction and natural language understanding more accessible than ever before.Hugging Face: The AI Model MetropolisHugging Face stands as a bustling metropolis where AI models thrive. This central hub offers a vast array of pre-trained models, serving as a fertile ground for machine learning exploration and application. To gain entry to this hub and its resources, you must create a Hugging Face account. Once you take this step, the doors to an expansive world of AI await you — just visit Hugging Face and sign up to begin your adventure.RAG: Harnessing Vector Databases for Accelerated IntelligenceRetrieval-Augmented Generation (RAG) is a sophisticated AI technique that marries the inventive power of Generative AI with the precision of knowledge retrieval, creating a system that’s not only articulate but also deeply informed. To unlock the full potential and efficiency of RAG, it integrates vector databases — a powerful tool for speedily sifting through vast information repositories. Here’s an enhanced breakdown of how RAG operates with vector databases:Retrieval with Vector Databases: RAG begins its process by querying a vector database, which houses embedded representations of a large corpus of information. These embeddings are high-dimensional vectors that encapsulate the semantic essence of documents or data snippets. Vector databases enable RAG to perform lightning-fast searches across these embeddings to pinpoint content that is most relevant to a given query, much like an AI swiftly navigating a digital library to find just the right book.Augmentation with Context: The relevant information retrieved from the vector database is then provided to a generative model as contextual augmentation. This step equips the AI with a concentrated dose of knowledge, enhancing its ability to craft responses that are not only creative but also contextually rich and precise.Generation of Informed Responses: Armed with this context, the generative model proceeds to produce text. Unlike standard generative models that rely solely on learned patterns, RAG weaves in the specifics from the retrieved data, resulting in outputs that are both imaginative and substantiated by the retrieved knowledge. The generation is thus elevated, yielding responses that are more accurate, informative, and reflective of true context.The integration of vector databases is key to RAG’s efficiency. Traditional metadata search methods can be slower and less precise, but vector databases facilitate near-instantaneous retrieval of contextually relevant information, even from extremely large datasets. This approach not only saves valuable time but also

Deploy Your Own AI Chat Buddy — The Qwen Chat Model Deployment with Hugging Face Guide

admin — Thu, 09 Jan 2025 08:42:34 +0100

Deploy Your Own AI Chat Buddy — The Qwen Chat Model Deployment with Hugging Face GuideBy FarruhAlright, you tech-savvy human, brace yourself for a thrilling adventure into the land of artificial intelligence! We’re not just dipping our toes here; we’re diving headfirst into the deep end with the Qwen Chat Model. What’s on the agenda? Setting up a cleverer chatbot than a fox and respecting privacy like a top-notch secret agent. Intrigued? You should be! Let’s start our journey by understanding Generative AI and LLM (Large Language Model).Generative AIGenerative AI refers to the branch of artificial intelligence focused on creating new content, whether text, images, music, or other forms of media. This type of AI leverages machine learning models, particularly generative models, to understand patterns, features, and relationships in large datasets and generate outputs that are new and often indistinguishable from human-created content.Types of Generative ModelsGenerative Adversarial Networks (GANs): A type of neural network architecture where two models (the generator and discriminator) are trained simultaneously. The generator creates new data instances while the discriminator evaluates them. The process results in increasingly more convincing outputs.Variational Autoencoders (VAEs): These models generate new instances similar to the input data. They’re often used in image generation.Transformers: Originally designed for NLP tasks, transformer models like GPT (Generative Pretrained Transformer) can generate coherent and contextually relevant text. They are also being adapted for generative tasks for other types of data.ApplicationsContent Creation: Generative AI can produce original artwork, write stories or articles, compose music, and create virtual environments for games and simulations.Data Augmentation: It can generate additional training data for machine learning models, helping to improve their accuracy and robustness.Personalization: Algorithms can tailor content to individual preferences, improving user engagement.Drug Discovery: Generative models can propose new molecular structures for drugs that could be effective against specific diseases.ChallengesQuality Control: Ensuring that the generated content meets quality standards and is free of biases present in the training data.Computational Requirements: Training generative models often requires significant computational power and large datasets.Interpretability: Understanding how these models make decisions and generate outputs can be challenging, which impacts trust and reliability.Generative AI continues to evolve rapidly, and its capabilities are expanding the boundaries of what machines can create, offering both exciting opportunities and challenges that need to be managed responsibly.LLMWhat are Large Language Models (LLMs)? They are a type of artificial intelligence based on deep learning techniques that are designed to understand, generate, and work with human language. They are called “large” because they consist of many millions, or even billions, of parameters, which allow them to capture a wide array of language nuances and contexts.LLMs are trained on vast amounts of text data and use architectures such as Transformer neural networks, which can process sequences of data (like sentences) and pay attention to different parts of the sequence when making predictions. This makes them particularly effective for a range of natural language processing (NLP) tasks, such as:Text generation: LLMs can write essays, create poetry, or generate code based on prompts given to them.Translation: They are capable of translating text between various languages with a high degree of accuracy.Question answering: LLMs can provide answers to questions by understanding context and extracting information.Summarization: They can condense long documents into concise summaries.Sentiment analysis: LLMs can determine the sentiment behind the text, such as identifying if a review is positive or negative.Why Qwen? A Quick RundownAre you on the lookout for an AI that can chat, create content, summarize, code, and much more, all while respecting your right to privacy? Look no further, the Qwen Chat Model is here to transform your data center into a bastion of secure AI-powered interactions.Qwen isn’t your average chatbot. It’s built on a massive language model and has been trained on a staggering 3 trillion tokens of multilingual data. This AI marvel understands both English and Chinese intricately and has been fine-tuned for human-like interaction.Why Go Local with Qwen?Deploying Qwen locally on your server is about taking control. It’s about ensuring that the conversations you have, the data processed, and the privacy promised remain under your purview. Whether you’re a business looking to integrate an intelligent chat system, a developer keen on AI research, or simply an enthusiast eager to explore the bounds of conversational AI, Qwen is your go-to choice.Now, why wo

Developing Multimodal Services using Qwen and Model Studio

admin — Thu, 09 Jan 2025 08:42:33 +0100

Author: FarruhIntroductionWe are on the cusp of a new era in artificial intelligence. With multimodal AI, the synergy between audio, visual, and textual data is not just an idea but an actionable reality, in which the Qwen Family of Large Language Models (LLMs) plays a pivotal role. This blog will serve as your gateway to understanding and implementing multimodal AI using Alibaba Cloud’s Model Studio, Qwen-Audio, Qwen-VL, Qwen-Agent, and OpenSearch (LLM-Based Conversational Search Edition).Here is the demo video linkHigh-Level Architecture OverviewAt its core, the multimodal AI we discuss today hinges on the following technological pillars:Qwen-Audio: Processes a wide array of audio inputs, converting them into actionable text.Qwen-VL: Analyzes images with unprecedented precision, revealing nuanced details and text within visuals.OpenSearch (LLM-Based Conversational Search Edition): Tailors Q&A systems to specific enterprise needs, leveraging vector retrieval and large-scale models.Qwen-Agent: Orchestrates intelligent agents that follow instructions and execute complex tasks.Model Studio: The one-stop AI development platform that brings our multimodal ecosystem to life.We used a planner agent that controls all solutions and the logic between them. The Planner Agent on Model Studio integrates all solutions into one Generative AI pipeline. Above this, with Python, an API will be created, ready for deployment on Alibaba Cloud’s Elastic Computing Service (ECS), and connected to DingTalk IM or any other IM platform you choose.Deep Dive into Qwen-Audio: A Symphony of Sound and LanguageQwen-Audio is not just an audio processing tool — it’s an auditory intelligence that speaks the language of sound with unparalleled fluency. It deals with everything from human speech to the subtleties of music, transforming audio to text with remarkable acuity, redefining how we interact with machines using sound as a medium.The Visual Frontier: Qwen-VL’s Pioneering VisionIn the realm of vision, Qwen-VL stands tall with models like Qwen-VL-Plus and Qwen-VL-Max that set new benchmarks in image processing. These models not only match but exceed the capabilities of industry giants, offering an extraordinary level of visual understanding. Whether it’s recognizing minute details in a million-pixel image or comprehending complex visual scenes, Qwen-VL is your lens to clarity.OpenSearch (LLM-Based Conversational Search Edition): One-Stop Multimodal SAAS RAGOpenSearch (LLM-Based Conversational Search Edition) embodies the quest for precision in a sea of data. It’s the beacon that enterprises need to navigate the complexities of industry-specific Q&A systems. The solution is elegant — vectorize your business data, index it, and let OpenSearch find the answers that are as accurate as they are relevant to your enterprise.Qwen-Agent: The Architect of Intelligent InteractionThe Qwen-Agent framework is where the building blocks of intelligence are assembled to create something truly special. With it, developers can construct agents that not only understand instructions but can use tools, plan, and remember. It’s not just an AI — it’s a digital being that can learn and evolve to meet your application’s needs.Model Studio: The GenAI PowerhouseAt the heart of this ecosystem lies Model Studio, Alibaba Cloud’s generative AI playground. This is where models are not just trained but born, tailored to the unique requirements of each application. It’s where the full spectrum of AI — from data management to deployment — comes together in a secure, responsible, and efficient manner.The API: Your Multimodal MaestroThe final act in our symphony is the creation of a unified API. Using Python and FlaskAPI, we will encapsulate the intelligence of our multimodal models into an accessible, scalable, and robust service. Deployed on ECS, this API will become the bridge that connects your applications to the intelligent orchestration of Qwen LLMs, ready to be engaged via DingTalk IM or any IM service of your preference.Integrating Qwen Family LLMs with Model Studio overall steps can be described below:Initial setup and configuration of Model Studio.Detailed instructions for integrating Qwen-Audio and Qwen-VL with your applications.Strategies for leveraging OpenSearch for creating intelligent enterprise solutions, link.Best practices for developing and deploying Qwen-Agent for enhanced AI interactions.Tips for orchestrating all these components into a single, cohesive API.Deployment guidelines on Alibaba Cloud ECS and connectivity with DingTalk IM.Detail step-by-step tutorials where by following you will become adept at creating AI applications that can see, hear, and understand the world in ways that were previously unimaginable.Use Cases: Bringing Multimodal AI to LifeMultimodal AI isn’t a distant dream — it’s already unlocking new opportunities across various industries. Here are some real-world applications where the Qwen Family LLMs and Model Studio integration

GenAI Model Optimization: Guide to Fine-Tuning and Quantization

admin — Thu, 09 Jan 2025 08:42:33 +0100

By FarruhArtificial Intelligence has transcended from a buzzword to a vital tool in both business and personal applications. As the AI field grows, so does the need for more efficient and task-specific models. This is where fine-tuning and quantization come into play, allowing us to refine pre-built models to better suit our needs and to do so more efficiently. Below is a guide designed to take beginners through the process of fine-tuning and quantizing a language model using Python and the Hugging Face Transformers library.The Importance of Fine-Tuning and Quantization in AIFine-tuning is akin to honing a broad skill set into a specialized one. A pre-trained language model might know a lot about many topics, but through fine-tuning, it can become an expert in a specific domain, such as legal jargon or medical terminology.Quantization compliments this by making these large models more resource-efficient, reducing the memory footprint and speeding up computation, which is especially beneficial when deploying models on edge devices or in environments with limited computational power.The Value for Businesses and IndividualsBusinesses can leverage fine-tuned and quantized models to create advanced AI applications that didn’t seem feasible due to resource constraints. For individuals, these techniques make it possible to run sophisticated AI on standard hardware, making personal projects or research more accessible.Setting Up Your Hugging Face AccountBefore tackling the code, you’ll need access to AI models and datasets. Hugging Face is the place to start:Visit Hugging Face.Click Sign Up to make a new account.Complete the registration process.Verify your email, and you’re all set!Preparing the EnvironmentFirst, the necessary libraries are imported. You’ll need the torch library for PyTorch functionality, and the transformers library from Hugging Face for model architectures and pre-trained weights. Other imports include datasets for loading and handling datasets, and peft and trl for efficient training routines and quantization support.import torchfrom datasets import load_datasetfrom transformers import ( AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, TrainingArguments, pipeline, logging,)from peft import LoraConfig, PeftModelfrom trl import SFTTrainerSelecting the Model and DatasetNext, the code specifies the model and dataset to use, which are crucial for fine-tuning. The model_name variable holds the identifier of the pre-trained model you wish to fine-tune, and dataset_name is the identifier of the dataset you'll use for training.model_name = "Qwen/Qwen-7B-Chat"dataset_name = "mlabonne/guanaco-llama2-1k"new_model = "Qwen-7B-Chat-SFT"Fine-Tuning ParametersParameters for fine-tuning are set using TrainingArguments. This includes the number of epochs, batch size, learning rate, and more, which determine how the model will learn during the fine-tuning process.training_arguments = TrainingArguments( output_dir="./results", num_train_epochs=1, per_device_train_batch_size=1, gradient_accumulation_steps=1, learning_rate=2e-4, weight_decay=0.001, # ... other arguments)Quantization with BitsAndBytesThe BitsAndBytesConfig configures the model for quantization. By setting load_in_4bit to True, you're enabling the model to use a 4-bit quantized version, reducing its size and potentially increasing speed.bnb_config = BitsAndBytesConfig( load_in_4bit=use_4bit, bnb_4bit_quant_type=bnb_4bit_quant_type, bnb_4bit_compute_dtype=compute_dtype, bnb_4bit_use_double_quant=use_nested_quant,)Fine-Tuning and Training the ModelThe model is loaded with the specified configuration, and the tokenizer is prepared. The SFTTrainer is then used to fine-tune the model on the loaded dataset. After training, the model is saved for future use.model = AutoModelForCausalLM.from_pretrained( model_name, quantization_config=bnb_config, # ... other configurations)trainer = SFTTrainer( model=model, train_dataset=dataset, # ... other configurations)trainer.train()trainer.model.save_pretrained(new_model)Evaluating Your ModelWith the model fine-tuned and quantized, you can now generate text based on prompts to see how well it performs. This is done using the pipeline function from transformers.pipe = pipeline(task="text-generation", model=model, tokenizer=tokenizer, max_length=200)result = pipe(f"[INST] {prompt} [/INST]")print(result[0]['generated_text'])Engaging Tutorial ReadersThis guide should walk the readers step by step, from setting up their environment to running their first fine-tuned and quantized model. Each step should be illustrated with a snippet from the code provided, explaining its purpose and guiding the reader on how to modify it for their needs.ConclusionBy the end of this tutorial, readers will have a solid understanding of how to fine-tune and quantize a pre-trained language model. This knowledge opens up a new world of possibilities for AI applications, m

Apsara Conference 2024 Is Here

admin — Thu, 09 Jan 2025 08:42:32 +0100

Get ready to revolutionize your vision at the leading annual cloud computing conference. Join us in Yunqi Town, Hangzhou, China, to discover AI on Cloud: Recharging · Innovating · Transforming.Join us for an action-packed conference where AI meets innovation and stay at the forefront of the cloud and AI revolution. Don’t miss our massive 40,000-square-meter intelligent technology expo, featuring three main forums and over 400+ interactive sessions.Mark your calendars for this epic event from September 19–21, 2024!Learn more: https://www.alibabacloud.com/en/apsara-conference/index

Alibaba Cloud 2024 Autumn MVP Application Now Open

admin — Thu, 09 Jan 2025 08:42:32 +0100

https://mvp.alibabacloud.com/Are you interested in collaborating with Alibaba Cloud scientists and architects to develop cloud computing solutions? Would you like to join a team of skilled professionals revolutionizing cloud technologies and Alibaba Cloud products?If you answered yes to the questions above or know somebody who fits this profile, Alibaba Cloud MVP is an excellent opportunity! If you are driven and passionate, have a community spirit, and have extraordinary technical abilities, nominate yourself or a friend as an Alibaba Cloud MVP (Autumn 2024 batch) by 10 September 2024.Click Here to Apply >>What Is an Alibaba Cloud MVP?Alibaba Cloud MVP (Most Valuable Professional) is an award for tech leaders and experts passionate about helping others understand and use Alibaba Cloud technologies.The MVP program is open to all professional developers worldwide. It awards candidates aiming to improve products and documents, build awareness, engage users and developers, and have excellent knowledge and experience in cloud computing and related areas. MVPs are a vital part of the Alibaba Cloud community.Who Is Qualified to Become an Alibaba Cloud MVP?An expert on cloud computing or related areasA person familiar with (and actively using) multiple Alibaba Cloud productsA contributor to the Alibaba Cloud community (in at least one of the following):Writing blog posts about Alibaba CloudSpeaking at tech events or webinars about Alibaba CloudInteracting on tech forums about Alibaba CloudHelping improve Alibaba Cloud ProductsWhat Are the Benefits of Becoming an Alibaba Cloud MVP?If you enroll in the program:You will receive an MVP certification and a welcome package.You can join our group and participate in exclusive internal training with Alibaba Cloud.You can attend and/or speak at Alibaba Cloud events (such as the Apsara Conference, Alibaba Cloud Day, and Developer Summit). You can visit Alibaba HQ and meet Alibaba Cloud experts.Let’s say you upgrade to a higher tier by contributing more to the community. In that case, you will get additional benefits, including up to USD 1500 in credits, free ACP, and exposure on Alibaba Cloud’s official channel.Click the following link for a list of comprehensive benefits of becoming an Alibaba Cloud MVP: https://mvp.alibabacloud.com/mvp/benefitshttps://mvp.alibabacloud.com/

Building RAG Services on Compute Nest with Alibaba Cloud Model Studio and AnalyticDB for…

admin — Thu, 09 Jan 2025 08:42:32 +0100

Building a RAG Service on Compute Nest with Alibaba Cloud Model Studio and AnalyticDB for PostgreSQLBy FarruhThis tutorial provides a step-by-step guide to setting up a Retrieval-Augmented Generation (RAG) service using Alibaba Cloud Model Studio, Compute Nest, and AnalyticDB for PostgreSQL. With Model Studio, you can leverage top-tier generative AI models like Qwen to develop, deploy, and manage AI applications effortlessly. This setup ensures secure and efficient data handling within your enterprise, enhancing AI capabilities and enabling seamless natural language queries.IntroductionAlibaba Cloud Model Studio provides a comprehensive platform for developing generative AI applications. Using Compute Nest and AnalyticDB for PostgreSQL, you can create a secure, efficient Retrieval-Augmented Generation (RAG) service to enhance AI capabilities within your enterprise.Overview of Alibaba Cloud Model StudioFeatures shown in this diagram will be launched graduallyWhat is Model Studio?Alibaba Cloud Model Studio is an end-to-end platform aimed at simplifying the development, deployment, and management of generative AI models. With access to industry-leading foundation models like Qwen-Max, Qwen-Plus, Qwen-Turbo, and Qwen 2 series, Model Studio provides tools for model fine-tuning, evaluation, deployment, and integration with enterprise systems.Key Capabilities of Model Studio1. Easy Access to Leading Foundation Models (FM):Models like Qwen-Max, Qwen-Plus, Qwen-Turbo, and the Qwen 2 series power your applications with enhanced AI capabilities.2. Built-In Model Inference and Evaluation Workflows:Support for Supervised Fine-Tuning (SFT) and Low-Rank Adaptation (LoRA).Model compression, inference acceleration, and multi-dimensional evaluation tools.One-click model deployment.3. Simplified Generative AI Application Development:Visual workflows for developing applications.Template-based prompt engineering.Extensive APIs for integration with business systems.4. Comprehensive Security Measures:Isolated VPC networks for securing data.tools for content governance and human-in-the-loop interventions to ensure responsible AI practices.5. Third-Party Models:Support for third-party models like Tongyi, showcased in Q&A, writing, and NL2SQL (Natural Language to SQL) functionalities.6. Data Management:Dataset cleansing and management.Retrieval-Augmented Generation (RAG) for enhanced search and data access.7. Industry-Specific Models:Custom models for sectors like healthcare, finance, and legal services.8. API and SDK:Assistant API and a suite of SDKs for quick integration and agent development.PrerequisitesBefore starting, ensure you have:An active Alibaba Cloud account.Familiarity with cloud services and AI models.Step 1: Alibaba Cloud Account SetupIf you haven’t already, sign up for an Alibaba Cloud account: Sign up.Step 2: Access Compute NestNavigate to Compute Nest and locate the service for Generative AI: Compute NestStep 3: Set Up an Instance and Its ParametersConfigure the necessary parameters for the instance:Service Instance Name: Provide a meaningful name for the instance.Elastic Computing Services (ECS) Parameters: Recommended to choose ecs.c6.2xlarge for faster document processing.Instance Password: Create a secure password for the instance.Step 4: Setup AnalyticDB for PostgreSQLConfigure an AnalyticDB for PostgreSQL instance:Instance Specification: Select the suitable specification based on your data volume.Segment Storage Size: Adjust according to your needs.DB Username: By default kbsuser, or choose your own username.DB Password: Create a strong password (avoid using symbols like “@”).Step 5: Configure WebUI CredentialsConfigure the web UI credentials to manage and interact with your RAG service:Username: Default is admin, or choose another username.Password: Create a strong, secure password.Step 6: Add Model Studio API KeyAdd your Model Studio API key to authenticate and facilitate communication between services:API Key: Enter the API key you obtained from your Model Studio setup.Here is a guide on how to obtain your Model Studio API key.Step 7: Network ConfigurationChoose the appropriate network settings to ensure secure and reliable connectivity:Choose Existing Infrastructure Configuration1. Select whether to create a new VPC (Virtual Private Cloud) or use an existing one.WhetherCreateVpc: Choose Create if you need a new VPC.2. VPC ID: Enter the ID of an existing VPC or create a new one.Create VPC: If creating a new VPC, follow the Alibaba Cloud VPC Creation Guide.3. VSwitch ID: Select the ID of an existing VSwitch or create a new one.Create VSwitch: Instructions are available in the VSwitch Creation Guide.4. Tags and Resource Groups:Tag: Specify a tag that is attached to the created resource.Tag Key: Choose the tag key.Tag Value: Choose the tag value.Resource Group: Select the resource group to which the created service instance belongs.Create Resource Group: Follow the instructions to Create a Resource Group.Aft

Tomorrow’s Canvas | Writing Competition

admin — Thu, 09 Jan 2025 08:42:31 +0100

Are you excited about the future? We are. The Alibaba Cloud Blog team are inviting you to share your insights into how AI will transform our futures.With “Tomorrow’s Canvas”, we want you to paint a picture of our futures using the technology available at your fingertips. Join us in exploring the limitless possibilities of AI, with a chance to be featured on our blog.How to enter:● Create a piece of artwork using AI diffusion tools.● Explain your vision in words. Talk about the ideas and concepts you have explored in your work.● Also in your article, you can talk more about your skills with AI diffusion and image-editing tools, to explain how you created your final image.Find out more on how to participateCheck out our first post