MPS Support in PyTorch 2.15

Summary: PyTorch 2.15 introduces MPS (Metal Performance Shaders) support. This enables more efficient execution of AI workloads, especially on devices with Apple Silicon chips.

PyTorch 2.15’s introduction of MPS support significantly improves the machine learning development and execution experience on Apple Silicon-based hardware. MPS is a framework developed by Apple for its own GPUs, providing direct hardware acceleration. This integration allows developers to perform faster model training and inference on their Mac devices.

Technically, MPS offers low-level access to GPU resources via the Metal API. With PyTorch integrating this support, tensor operations and model computations can be executed directly on the GPU cores within Apple Silicon. This provides a significant performance boost and energy efficiency compared to CPU-based or emulation-layer solutions.

From a live systems and infrastructure perspective, this development is particularly impactful in development environments and small-scale deployments. Developers can iterate and test their models faster on powerful Mac workstations. However, since large-scale production deployments typically focus on NVIDIA GPUs and the CUDA ecosystem, MPS support primarily benefits the development and prototyping phases.

Engineers, especially in teams using Apple hardware, should consider evaluating and integrating this new MPS support into their workflows. It’s important to perform performance benchmarks to determine which workloads run more efficiently with MPS and to leverage this feature to accelerate development cycles.

🔗 Source: Hugging Face Blog
📡 Authoritative feeds & live tracking: Radar