Google Cloud recently launched the OpenXLA project, a community-based open source ecosystem of ML compilers and infrastructure projects aimed at making ML frameworks easy to use with a variety of hardware backends for faster, more flexible and impactful development.
Proprietary software is a drag on AI and ML innovation, Google advocates an open source approach. Sachin Gupta, Google vice president and general manager of infrastructure, says in a blog post about the project:
"At Google, we believe that open source software is essential to overcoming the challenges associated with inflexible strategies. And as a major contributor to the Cloud Native Computing Foundation (CNCF), we have more than two decades of experience working with the community to turn OSS projects into accessible and transparent catalysts for technological progress. We are committed to open ecosystems of all kinds, and that commitment extends to AI / ML - we strongly believe that no one company should own AI / ML innovation."
The OpenXLA Project
Developers often run into incompatibilities between frameworks and hardware when creating ML solutions.
The OpenXLA project is a modular, community-driven open source compiler ecosystem coded by AI / ML leaders including AMD, Arm, Google, Intel, Meta, NVIDIA...It will enable efficient reduction, optimization, and deployment of ML models from most major frameworks (TensorFlow, PyTorch, and JAX) to any hardware backend, including CPUs, GPUs, and ML ASICs.
The first objectives
The community will begin by collaboratively evolving the XLA, (Accelerated Linear Algebra) compiler, a linear algebra compiler decoupled from TensorFlow, which allows for the acceleration of TensorFlow models without the need to necessarily modify the source code.
For example, in the case of the BERT model, it allowed a 7-fold increase in performance and a 5-fold increase in batch size for an MLPerf submission using 8 Volta V100 GPUs.
It will also evolve StableHLO, a portable ML computational operation set that facilitates the deployment of frameworks on different hardware options, inspired by the MHLO dialect to which it has brought new features, including serialization and versioning.
During the seed phase of the project in 2022, Google engineers will assume responsibility for the technical leadership of the project. Collaboration principles, code review processes, and a community infrastructure for OpenXLA will be established in 2023 when the project leaves the TensorFlow organization.
All people involved in development or integration with XLA are invited to participate in the discussions. To participate, members can request an invitation to join the GitHub and SIG Discord organization, which will be announced later.
For more information: https:
//github.com/openxla/xla
Translated from
Google Cloud lance OpenXLA, un projet open source visant à optimiser le développement de modèles ML
