MPs in PyTorch 2.15
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
PyTorch 2.15 brings FP8 tensors and scaled matrix multiplication to Apple's MPS backend, plus native grouped matmul and a dedicated decoding path that lift mixture-of-experts LLM decoding about 5–6.5× on an M3 Ultra.