Hardware-grounded model architecture design principles, MoE co-design patterns, low-precision numerics, and parallelism guides for training and serving on Google Cloud TPUs.
-
Updated
Nov 13, 2026
Hardware-grounded model architecture design principles, MoE co-design patterns, low-precision numerics, and parallelism guides for training and serving on Google Cloud TPUs.
Native Google Cloud TPU (v6e Trillium / TorchTPU) inference & Pallas kernel acceleration for MOSS-Image 5B — 38.4% MFU (2.96x speedup), 16.7s generation, and 4.86x faster zero-copy drop_tpu offload.
Captum explainability on Cloud TPUs via TorchTPU: compatibility matrix, device-generic test suite, upstream fixes and a tutorial. Built for the TorchTPU Sprint.
To associate your repository with the torchtpu topic, visit your repo's landing page and select "manage topics."