vAccel v0.6 is out!

Nubificus LTD proudly presents the latest release of the vAccel framework. vAccel v0.6.0 is available, as well as a hot fix followup for x86_64, aarch64, and armv7l architectures, with the essential binaries for users to get started.

A major addition to our previous release is Torch support. Users are now able to run inference on Torch models seamlessly, on local and remote targets (CPU/GPU etc.). A short screen cast is available to check torch with vAccel in action!

vAccel v0.6 includes updated helper functions for easier argument definition (exec), as well as enhanced CI and testing support.

A fun addition was native Golang bindings! Users can now natively interact with Go programs and enjoy hardware acceleration from simple web services! Golang bindings helped a lot with our Knative integration!

vAccel v0.6 offers updated API remoting functionality over generic sockets, supporting AF_VSOCK and AF_INET, enabling local and remote execution over the network. AF_VSOCK support is also updated with a streaming optimization, to reduce the amount of memory allocated by the gRPC transport layer.

This iteration’s update also contains important bug fixes and performance optimizations. For the list of changes, see the RELEASE notes.

The individual components are packaged as binary artifacts or deb packages for users to install them directly. For a list of the binary components please visit vAccel’s documentation page. The core vAccel library is open-source, available on github.

The roadmap for v0.7 contains enhanced Torch support, OpenCV-native bindings, a C++-based transport layer, the move to the meson build system, and more!

vAccel

vAccel enables workloads to enjoy hardware acceleration while running on environments that do not have direct (physical) access to acceleration devices. With a slim design and precise abstractions, vAccel semantically exposes hardware acceleration features to users with little to no knowledge of software acceleration framework internals.

vAccel integrates with container runtimes such as kata-containers. v0.6 brings updated support for kata-containers v3.X, both for the Go and rust runtimes, so deploying an application that requires hardware acceleration in a sandboxed container in k8s is now possible without complicated hardware setups!

Serverless computing workflows are now able to enjoy compute-offload mechanisms to provide AI/ML services as functions, triggered thousands or millions of times by events, on-demand, auto-scaling to multiple physical and virtual nodes, without the need to directly attach a hardware device to the instance. This functionality is enabled through vAccel’s virtualization backends, enabling the execution of hardware-accelerated functions on Virtual Machines. Support for numerous hypervisors is available, including AWS Firecracker, QEMU/KVM, Cloud Hypervisor, and Dragonball, the stock hypervisor of kata-containers! See the relevant documentation on how to run an example on a VM.

End-users can get a sneak peek at what vAccel has to offer on the project’s website, on github, or by browsing through the available documentation. To get started, follow the Quick start guide.