PyTorch 2.15 MPS Update Adds FP8 Support and Memory Optimizations
Current overview (not historical)
PyTorch 2.15 introduces an MPS backend update that, for the first time, supports FP8 low‑precision tensors on macOS. Each FP8 value occupies only one byte, halving the storage needed compared to BF16. The update delivers significant speed gains for Mixture‑of‑Experts (MoE) models, boosting decoding throughput by 5–6.5×. For example, the Qwen3.6‑35B‑A3B model on an M3 Ultra achieves 61.3 tokens per second, up from 9.4 tokens per second. Additionally, the migration of Metal kernels and padding optimizations reduce peak VRAM usage from roughly 270 GiB to just 8 GiB, dramatically improving memory efficiency on Apple silicon.
Written with AI from the reporting · updated 10/20/2026
EventGraph Explorer
One time cutoff for reports, relations and evidence.
New at this step ·PyTorch 2.15 Adds FP8, Accelerates MoE Decoding, and Cuts Wan2.2 Video VAE Memory
- Reports in replay
- 1
- First-party
- 1
- Sources
- 1
Reports first seen at this step · 1
- Hugging Face Blog · First-partyPyTorch 2.15 Adds FP8, Accelerates MoE Decoding, and Cuts Wan2.2 Video VAE MemoryOriginal source ↗
Discovery-time replay; the linked graph and evidence share this cutoff. The full overview and activity remain current.
Event developments
1 developments10/20/2026, 7:18:10 PM
PyTorch 2.15 Adds FP8, Accelerates MoE Decoding, and Cuts Wan2.2 Video VAE MemoryHugging Face Blog
Filtered by first discovery; only representative reports already discovered by this cutoff are shown. Story grouping may have changed later, so this is not a complete historical snapshot.
Relations at this time
Loading time-scoped relations…
Bounded reconstruction, not an exact snapshot: entity names, relationship text and story grouping may include later revisions. Overview and full source archive below remain current.
Story route
A station-by-station route from the first public evidence onward; every station carries its own evidence and uncertainty.
- 1Oct 20, 00:00recorded Oct 20, 19:18single sourcerevised
MPS in PyTorch 2.15
Read original source ↗ Hugging Face BlogReplay this moment in the explorer
Uncertainty marks flag where to check the sources; they are not proven causation.
Full source archive (current)
Current public reports; outside the historical replay.
First-party sources · 1 reports
- Hugging Face Blog · Oct 20, 08:00PyTorch 2.15 Adds FP8, Accelerates MoE Decoding, and Cuts Wan2.2 Video VAE Memory
PyTorch 2.15 introduces FP8 support on MPS, enabling half‑size storage of float8_e4m3fn tensors and scaled matrix multiplication. Mixture‑of‑experts models now decode 5–6.5× faster on an M3 Ultra, with Qwen3.6‑35B‑A3B jumping from 9.4 to 61.3 tokens/s. Wan2.2’s video VAE memory usage drops from 270 GiB to about 8 GiB. The update also speeds up linear layers, softmax, and various audio‑video operations, and requires macOS 15+ and Python 3.11+.
Read original source ↗
Independent coverage · 0 reports
No verifiable reports in this group yet.
Different reports do not necessarily disagree. Check original sources and publication dates.
Report timeline
Explore every side through dated source reports.
- Hugging Face BlogSelectedPyTorch 2.15 Adds FP8, Accelerates MoE Decoding, and Cuts Wan2.2 Video VAE Memory
PyTorch 2.15 introduces FP8 support on MPS, enabling half‑size storage of float8_e4m3fn tensors and scaled matrix multiplication. Mixture‑of‑experts models now decode 5–6.5× faster on an M3 Ultra, with Qwen3.6‑35B‑A3B jumping from 9.4 to 61.3 tokens/s. Wan2.2’s video VAE memory usage drops from 270 GiB to about 8 GiB. The update also speeds up linear layers, softmax, and various audio‑video operations, and requires macOS 15+ and Python 3.11+.