Speed often determines success or failure in business.
The acceleration of model training means that enterprises can develop new AI products or services faster, thereby bringing them to market faster and meeting customer needs. This rapid response capability helps companies seize opportunities in the competition and gain market advantages. Not only that, acceleration also means that enterprises can reduce the time required for training, thereby reducing the occupation of hardware resources and energy consumption, resulting in higher ROI.
As the world's leading high-performance AI data access platform, Alluxio has been widely used in the core stages of AI model training and inference. This time, we have teamed up with Zhongguancun’s well-known incubator brand Taili Station and DataFun, a vertical community of data intelligence developers , to invite AI Infra experts from the Internet, automobile, and large model industries to bring industry partners an accelerating AI model training. A wonderful offline salon with themes.
We sincerely invite AI Infra-related IT/technical leaders, architects, developers and researchers, and ecological partners to sign up and participate.
Recognize QR code
Sign up now to participate
√
Event time: April 26, 2024 (Friday) 14:00-17:10
√
Co-organizer: Alluxio x Inno.EcoS Lab x DataFun
√
Venue: Room 401, 4th Floor, Tower A, Dongsheng Building, No. 8 Zhongguancun East Road, Haidian District, Beijing
Topic 1: Application and deployment of Alluxio in autonomous driving model training
Huixi Intelligence will switch the file cache for autonomous driving algorithm training from NAS to Alluxio in 2023. The caching system based on Alluxio solves the problems that have plagued research and development for a long time: serious lags in concurrent data access, repeated data downloads, waste of space due to repeated data, inefficiency and high operational risks caused by manual management of storage capacity , improving the availability of the data system by 10 times. The cost is halved and the ease of use is greatly improved, thus helping the team greatly improve the efficiency of algorithm development.
√
Why did Huixi choose Alluxio?
√
How to use it in autonomous driving across clusters?
√
How can its functionality and performance be adequately tested and verified?
√
How to improve Alluxio’s operation and maintenance capabilities?
Topic 2: How Alluxio accelerates AI storage under hybrid cloud
In 2023, Zhihu adopted Alluxio for the first time in the model distribution scenario, which not only solved the problem of cross-cloud dedicated line bandwidth, but also brought about a 2-3 times improvement in reading performance.
With the development of large language model training within Zhihu, the algorithm team has also put forward higher requirements for storage. The Fuse interface provided by Alluxio has well met the needs of the business side, and has established a firm foothold within Zhihu, and quickly Development, from the initial cluster, developed into multiple clusters.
Zhihu has a hybrid cloud architecture. In order to reduce the delay in data access during model training, a set of Alluxio clusters are deployed in each public cloud. Alluxio's transparent caching capabilities are used to quickly distribute training data on offline HDFS clusters to each public cloud. On a GPU machine, and without any data migration or copying, it greatly improves the GPU utilization of training tasks.
Topic 3: Alluxio AI-a new generation of AI/ML training platform data I/O solution
In the era of data-driven AI, efficient access to large amounts of data in storage is critical for model training and serving. However, I/O challenges often hinder performance and limit GPU utilization.
本次分享，我们将介绍基于 Alluxio 构建的高性能数据访问层，如何克服 I/O 挑战并显著提高 GPU 利用率。通过丰富的用户案例和实验数据，您将了解在Alluxio中缓存数据集和模型的方法以及在性能方面的提升。
Theme 4: The technical accumulation and advantages of Moda community in the direction of large model training and inference
The MoTa community affiliated to Tongyi Lab has made a lot of explorations in training in the direction of LLM/MLLM/SD. In order to facilitate community developers to train and apply LLM, and truly achieve inclusive AI, the MoTa community has developed open source training The inference framework SWIFT has the training capabilities of 200+ LLM and 100+ open source data sets, and can be easily extended to new models. In the direction of SD AIGC, Moda Community and the basic vision team of Tongyi Laboratory developed the training framework Scepter. This framework uses LoRA and self-developed SCEdit technology to achieve convenient fine-tuning and inference of Vincentian graph models, and can support controllable generate.
...
and there are more exciting contents waiting for you to discover on site
......
14:00-14:05 Host opens the show
14:05-14:45 Alluxio AI new generation AI/ML training platform data I/O solution
14:45-15:25 How Alluxio accelerates AI storage under hybrid cloud
15:25-15:40 Tea break
15:40-16:20 Application and deployment of Alluxio in autonomous driving model training
16:20-17:00 The technical accumulation and advantages of Moda community in the direction of large model training and inference
17:00-17:10 Technical exchange & end
Recognize QR code
Sign up now to participate
Participate in this event
Everyone who attends the event will receive a souvenir
At the same time, there is an on-site activity for filling out questionnaires and drawing gifts. Exquisite gifts are waiting for you.
If you have any questions, please scan the QR code of the assistant at the end of the article and contact us ~
Alluxio is the world's leading provider of high-performance data platforms for analytics and AI, accelerating the value realization of enterprise AI products and maximizing infrastructure return on investment. The Alluxio data platform sits between computing and storage systems, providing a unified view of workloads on the data platform at every stage of the data workflow. The platform provides high-performance data access no matter where the data resides, simplifies data engineering, improves GPU utilization, and reduces cloud computing and storage costs. Enterprises can significantly accelerate model training and model serving and build AI infrastructure on existing data lakes without using dedicated storage.
With the support of leading investors, Alluxio provides services to global technology, Internet, financial and telecommunications companies. Currently, 9 of the top 10 Internet companies in the world are using Alluxio. For more information, please visit www.alluxio.com.cn.
Inno.EcoS Lab Taili Station is the incubator brand of Zhongguancun Dongsheng Science and Technology Park. It is an industrial innovation incubation acceleration network and innovation platform established with the Inno.EcoS high-tech enterprise growth ecosystem as the core. Taili Station focuses on the three major industrial fields of life sciences, digital economy, and new energy/new materials. It has been deeply engaged in industrial services for more than 10 years. It gathers innovation and entrepreneurial resources from around the world to provide multiple choices for high-tech enterprises in the pre-incubation, acceleration, and growth stages. Multiple types and locations of office space and supporting innovative enterprise technology services.
Founded at the end of 2017, DataFun is a vertical community focused on serving data intelligence developers. Driven by the mission of “creating millions of data-intelligent developers and helping tens of thousands of enterprises become digitally intelligent”, through nearly 6 years of continuous operation, more than 4,000 experts in the field have been invited to share their experiences, and more than 100,000 experts have accumulated experience in the form of videos, pictures and texts. 2,000 application cases, and influenced 500,000 precise developers across the entire network. At DataFun, you can connect to authoritative experts, cutting-edge technologies, best practices and outstanding developer groups in the field of data intelligence. I hope that DataFun can accompany developers, enterprises and industries to rush into the era of data intelligence.
✦
[Add assistant to learn more event details]
✦
✦
【Recent Popularity】
✦
✦
【Baodian Market】
✦
This article is shared from the WeChat public account - Alluxio (Alluxio_China).
If there is any infringement, please contact [email protected] for deletion.
This article participates in the " OSC Source Creation Plan ". You who are reading are welcome to join and share together.