Mechanism: Native grouped matrix multiplication, specialized linear-layer and softmax kernels, Metal padding kernels, and FP8 tensor support change where MPS executes workload operations and how much memory some tasks require. These changes reduce work that previously passed through MPSGraph or required CPU interaction. FP8 tensors add a lower-storage representation, while new padding kernels address a separate memory problem in tiled video decoding.
Why now: The release addresses practical constraints the authors identify in running language, image, and video models on Apple silicon: unsupported operations, performance cliffs, and memory growth. Their reported MoE example, Qwen3.6-35B-A3B on an M3 Ultra with torch.compile, moves from 9.4 to 61.3 tokens per second versus PyTorch 2.14. The Wan2.2 tiled VAE example reports peak total driver memory falling from about 270 GiB to 8 GiB. These figures make the backend work legible as an operational change, not merely a list of new operators.
First implication: Teams using Macs for local inference or media generation can benchmark workloads on PyTorch 2.15 to determine whether the specific backend changes help their use case. The release also specifies macOS 15 or later and Python 3.11 or later, so environment compatibility is part of the upgrade decision.
Evidence boundary: The source is a Hugging Face team overview with selected workload measurements; it does not establish representative gains across all Macs, models, or software stacks. The published throughput and memory figures should be treated as examples tied to stated configurations, not universal performance guarantees.
First watch condition: Watch for independent reproduction of these gains across different Apple silicon generations and commonly used local inference stacks.