HomeNews

PyTorch 2.15 MPS Backend Launches Enhanced Mac GPU Support

Kamran
Kamran
•
Oct 31, 2026 3:55 AM
•
0
 min read
Select Emergent as your Preferred news source
PyTorch 2.15 MPS Backend Launches Enhanced Mac GPU Support

💡 TL;DR

  • PyTorch 2.15 brings significant MPS backend improvements for Apple Silicon Macs with expanded operator coverage and performance gains.
  • The update enables faster model training and inference on Mac GPUs through optimized Metal Performance Shaders integration.
  • Developers can now leverage more PyTorch operations natively on Apple hardware without CPU fallback penalties.

PyTorch 2.15 delivers substantial improvements to its Metal Performance Shaders (MPS) backend, bringing enhanced GPU acceleration to Apple Silicon Macs. The update expands operator support and performance optimizations for developers training and deploying machine learning models on macOS hardware. Hugging Face highlighted the release as a major step forward for Mac-based AI development workflows.

Officially launched on October 20, 2026

PyTorch 2.15 became generally available on October 20, 2026, with the MPS backend enhancements as a flagship feature. The release represents months of collaboration between the PyTorch team and Apple to optimize deep learning frameworks for the M-series chip architecture. Developers can now access the updated framework through standard PyTorch installation channels.

Expanded operator coverage for Mac GPUs

The MPS backend in PyTorch 2.15 introduces support for dozens of additional operators that previously required CPU fallback. This expansion means more model architectures can run entirely on Apple Silicon GPUs without performance degradation from switching between accelerators. Key additions include advanced tensor operations, specialized loss functions, and optimized attention mechanisms commonly used in transformer models.

The update particularly benefits researchers and engineers working with large language models and computer vision architectures on Mac hardware. Operations that formerly triggered CPU fallback now execute natively on the GPU, reducing training time and improving inference throughput. Benchmark tests show performance gains ranging from 15% to 40% depending on model architecture and batch size.

Performance optimizations and memory efficiency

Beyond operator coverage, PyTorch 2.15 implements memory management improvements specific to the unified memory architecture in Apple Silicon. The MPS backend now handles memory allocation more efficiently, reducing overhead and enabling larger batch sizes during training. These optimizations make better use of the shared memory pool between CPU and GPU in M-series chips.

Key performance enhancements include:

  • Faster kernel launches for frequently used operations
  • Improved memory pooling to reduce allocation overhead
  • Optimized data transfer between Python and Metal layers
  • Better cache utilization for repeated tensor operations

Integration with Hugging Face ecosystem

Hugging Face confirmed that PyTorch 2.15 works seamlessly with its Transformers library and other ecosystem tools. Developers using popular pre-trained models can immediately benefit from the MPS improvements without code changes. The compatibility extends to training pipelines, inference endpoints, and model fine-tuning workflows on Mac hardware.

The integration allows data scientists to prototype and develop models locally on Apple Silicon before scaling to cloud GPU clusters. This workflow reduces development costs and iteration time while maintaining full compatibility with production environments. Models trained with the enhanced MPS backend show no accuracy differences compared to CUDA or CPU training.

What this means for Mac-based AI development

PyTorch 2.15's MPS enhancements solidify Apple Silicon as a viable platform for serious machine learning work beyond basic prototyping. Developers gain access to hardware-accelerated training and inference without requiring external GPU resources or cloud instances. The update particularly benefits researchers, students, and small teams working on Mac hardware who need local development capabilities. As the PyTorch team continues optimizing for Apple's architecture, the performance gap between Mac GPUs and dedicated NVIDIA hardware narrows for many common workloads.

About the writer

Kamran Alam is an Engineer at Emergent, where he builds the core platform powering AI-driven software creation. He previously co-founded Jeevam Health (YC S20) as CTO and holds a B.Tech. from IIT Roorkee.

Start Building
on Emergent today
Try Emergent