At Chainguard, we are constantly aiming to innovate and meet the evolving needs of our customers. Today, we are thrilled to announce an Early Access Program to our CUDA Optimized Images, which aims to enable the safe deployment and management of NVIDIA Accelerated AI applications on containers.
Your safe source for AI
Our mission is to be the safe source for open source, and the delivery of our AI-focused Images is a significant step towards realizing this vision. We are providing a secure alternative for developers who heavily leverage GPUs for their container workloads and are seeking to implement security best practices into their production environments.
The type of risk we are going after is model integrity compromise and more specifically model deployment tampering. As with any software supply chain, the infrastructure that is being used to develop, train, and deploy machine learning models should have a limited attack surface and zero vulnerabilities.
What are Chainguard’s CUDA Optimized Images?
CUDA Optimized images are designed to provide a seamless, simple, and secure solution for deploying and managing NVIDIA Accelerated AI applications. These new images can help you secure your environment while increasing the productivity of your machine learning team. Here’s how:
- Squashing vulnerabilities in your GPU containers and reducing image size
- Image CVEs: The latest available PyTorch runtime image currently contains 145 CVEs including two that are classified as high. Container images are rarely updated representing a risk to organizations using them. By contrast, the Chainguard PyTorch image contains zero CVEs.
- Bulky deployments: Many GPU Optimized images exceed 20 GBs in size. When downloading and deploying these images, machine learning engineering teams and their workflows are severely slowed down.
- Accelerating AI development velocity
- Driver compatibility: Ensuring compatibility between GPU drivers, CUDA versions, and host systems can be challenging. Mismatched versions can lead to runtime errors or performance issues.
- Software dependencies: Installing the CUDA toolkit within the container environment requires careful version matching with GPU drivers and other software dependencies. Version mismatches can lead to compilation errors and runtime failures.
- Debugging challenges: Debugging issues within containerized environments with GPUs can be complex. CUDA Optimized images simplify this process, making it easier to isolate the source of errors. For instance, the Chainguard PyTorch Cuda image has 33 packages vs 268 on the current official PyTorch runtime image.
Help us shape AI security
By participating in our Early Access Program, you will have the opportunity to engage with us where your feedback will shape the future of our product. We value the insights and experiences of our customers, and by involving you in the development process, we aim to gather valuable feedback on usability, configuration requirements, performance metrics, and more.
What does the Early Access Program entail?
- Feedback: As a participant, you can provide iterative feedback on usability, configuration requirements, performance metrics, and any enhancements.
- Limited Availability: To provide personalized interactions we will be limiting access to the program to existing customers and Wolfi Images users.
- Communication: We will maintain transparent communication with participants through regular updates, progress reports, and feedback sessions.
- Timelines: The program will be running until May 15, 2024.
Getting started
Are you running GPU accelerated container images today? If so, you’re in luck — we have two of the most popular applications ready for you!
- NVIDIA NeMo (Neural Modules) is a toolkit developed by NVIDIA for building and training state-of-the-art conversational AI models. It is built on top of PyTorch Lightning and provides high-level abstractions and pre-built components for developing speech and natural language processing (NLP) models. Try Chainguard’s NeMo Image.
- PyTorch is widely used in research and industry for various machine learning tasks, including computer vision, natural language processing, and reinforcement learning. Check out our getting started guide and try Chainguard's PyTorch Image.
If you're interested in being part of our Early Access Program or need an Image not on the list, reach out to us today.
