Bill Jia (@googlecloud) and Zongwei Zhou (@Meta) introduce TorchTPU, a native PyTorch backend for Google TPUs, and agentic developer workflows at #PyTorchCon 2026.
TorchTPU is an open source collaboration between Google and Meta, developed in close partnership with PyTorch core tes with Hugging Face Transformers for modeling, TorchTitan for training, and vLLM and SGLang for serving.
They demonstrate how long-horizon agentic workflows can migrate complex model workloads from GPUs to TPUs. They explore using agents for performance optimization, including quantization, custom kernel generation, and advanced sharding strategies.
Watch the keynote livestream: events.linuxfoundation.org/pytorch-confer…
TorchTPU is an open source collaboration between Google and Meta, developed in close partnership with PyTorch core tes with Hugging Face Transformers for modeling, TorchTitan for training, and vLLM and SGLang for serving.
They demonstrate how long-horizon agentic workflows can migrate complex model workloads from GPUs to TPUs. They explore using agents for performance optimization, including quantization, custom kernel generation, and advanced sharding strategies.
Watch the keynote livestream: events.linuxfoundation.org/pytorch-confer…
Oct 21, 2026 · 4:30 PM UTC
4
6
69
7,793
@googlecloud @Meta Read our latest blog introducing TorchTPU to learn more: pytorch.org/blog/introduci…
0
3
11
1,811
@PyTorch @googlecloud @Meta The agent-assisted migration is what really caught my attention. Getting a model to run on different hardware is one thing; making it run well, with all the kernel and sharding work involved, is another. That's where TorchTPU could make a real difference.
0
0
0
55
