arXiv:2610.24161 · v4
TorchTPU - A Hybrid Eager/Compiled Architecture for Native PyTorch Execution on TPU
Claudio BasileMichael VoznesenskyGunhyun ParkBen WilsonParker HollowayZhanyong WanVlad BelousManish KhannaPaul GanssleKevin GleasonAndrew DameBrendan FolieBhavya BahlHaifeng JinPedro Goncalves MokarzelArian ArfaianLuke BoyerChi ZhangYiling YuanAdam MainzAditya PatelHuangze TangJay GanLizzie LiangWill FroomTushar AgarwalKarlo BasioliTheotime CombesEmilio G. CotaMinos PapadopoulosTudor PiperneaYarong MuQi ZhouYaoshiang HoAleksey VlasenkoAlexander BrauckmannZiyin HuangJack OhJerry HsuTanvi BajpaiJialei ChenWenjun GuChintan ShahChris ChanSayce FalkChris JonesRob TorresKat KoLee HowesFreya AzadAndi GavrilescuJoe PamerBill Jia
首次提交 2026-10-17 03:38
版本更新 2026-10-29 23:44
雷达观测 2026-10-30 13:28
摘要
TorchTPU is an open-source collaboration between Google and Meta that introduces a native and principled PyTorch backend for Google TPUs. It enables seamless out-of-the-box model execution and TPU migration while delivering industry-competitive inference and training performance through eager execution optimizations, torch.compile integration, and native hardware acceleration.
作者备注：23 pages, 5 figures