OpenAI Standardizes on PyTorch

OpenAI has standardized its deep learning framework on PyTorch. This move is designed to increase research productivity at scale on GPUs, enabling the team to create and share optimized model implementations more efficiently.

Research Productivity and Iteration Speed

PyTorch was selected as the primary framework to accelerate the research cycle. OpenAI reports that switching to PyTorch decreased iteration time on research ideas in generative modeling from weeks to days. The framework's flexibility makes it easier for researchers to execute and test new ideas quickly.

Open Source Contributions and Tooling

As part of the transition to PyTorch, OpenAI is updating its open-source resources and specialized kernels:

• Spinning Up in Deep RL: OpenAI has released a PyTorch-enabled version of this educational resource for deep reinforcement learning.
• Blocksparse Kernels: OpenAI is developing PyTorch bindings for its highly-optimized blocksparse GPU kernels, with plans to open-source these bindings in the following months.

Framework Strategy and Community Engagement

While PyTorch is now the primary framework, OpenAI will continue to use other frameworks when a specific technical reason necessitates it. By standardizing, OpenAI joins a developer community including organizations like Facebook and Microsoft to push the performance and scale of GPU-based deep learning.