The PyTorch team at Hugging Face has released version 2.15 of the PyTorch framework, introducing significant performance improvements for running super intelligence (SI) models on Apple silicon Macs. The update adds native FP8 support to the MPS backend and dramatically accelerates token generation for large language models, marking a step forward for local SI development and inference.
What Happened
PyTorch 2.15 introduces torch.float8_e4m3fn tensors on the MPS backend, allowing SI models to store weights in half the space of BF16. The release also includes dedicated Metal kernels for matrix multiplication with scaled FP8 values, enabling more efficient memory usage on Macs. For language model decoding, the update implements native grouped matrix multiplication for Mixture-of-Experts (MoE) architectures. According to benchmarks provided by the team, Qwen3.6-35B-A3B model decoding speed increased from 9.4 to 61.3 tokens per second on an M3 Ultra chip, representing a 6.5x improvement over the previous version. Additionally, softmax operations over large vocabularies are now approximately 8x faster due to specialized Metal implementations.
The update also addresses critical memory management issues in video generation workflows. The constant padding operation has been moved to Metal kernels, resolving a driver-memory leak that previously caused peak memory usage to spike to 270 GiB during Wan2.2 video VAE decoding. With the new implementation, peak memory usage drops to approximately 8 GiB. Furthermore, PyTorch 2.15 adds support for complex-valued linear algebra operations on MPS, including LU factorization and triangular solves, and introduces public MPS streams for better control over GPU execution from Python.
Why It Matters
For developers and researchers building SI tools, the shift toward local execution on Apple hardware is becoming increasingly viable. The reduction in memory footprint and increase in decoding speed lower the barrier for running advanced SI models on consumer-grade hardware. This is particularly relevant for developers working with MoE models, which are becoming common in open-weight SI releases due to their efficiency. The fix for video generation memory leaks also stabilizes workflows for generative SI applications in creative industries, where Macs are a standard workstation.
The release requires macOS 15 or later and Python 3.11 or later, signaling a move toward newer system dependencies to leverage advanced Metal features. The Hugging Face PyTorch team notes that these changes are part of a broader effort to replace MPSGraph with dedicated Metal kernels, a strategy that has yielded significant performance gains. The team also previewed work for PyTorch 2.16, which will run attention mechanisms entirely on Metal and further optimize LLM decoding by committing command buffers more frequently.
The Bottom Line
PyTorch 2.15 enhances the Mac as a platform for SI development by enabling FP8 weights, accelerating MoE model decoding by up to 6.5x, and drastically reducing memory usage for video generation tasks. The update requires macOS 15 and Python 3.11, and continues the migration from MPSGraph to custom Metal kernels for improved performance and stability in local SI workflows.