PyTorch

贾扬清：我所经历的「人工智能已死」到「AI 颠覆世界」的数年巨变丨串台「声东击西」S10E24 adds PyTorch as part of the post-Caffe and [[FacebookAIInfra|Facebook AI Infra]] framework lineage. The source frames frameworks as infrastructure that changes iteration speed and production adoption, not just as developer preference.

PyTorch appears in E228｜谷歌TPU能撼动英伟达吗？前TPU工程师首次揭秘 as the external model-engineering framework that Google must support better if TPU is to expand beyond Google-native teams. The episode says Meta uses PyTorch heavily and that historical PyTorch-TPU compatibility limits made TPU less convenient for many outside teams.

The source frames PyTorch support as a bridge problem. [[XLACompiler|XLA]] can sit below PyTorch, JAX, or TensorFlow, but missing libraries, unsupported operators, weak debugging, or lower utilization can keep a promising specialized chip from becoming practical MaaS Infrastructure.

148. 对游凯超3小时访谈：开源Infra、和模型Co-design 、“如果vLLM失败，我们会后悔一辈子” adds the [[PyTorchFoundation|PyTorch Foundation]] governance context. The episode says [[VLLM|vLLM]] was donated to the foundation so the project name and community ownership would remain protected rather than becoming a closed company asset.

Connections

• TPU, XLA Compiler, and JAX — compiler and framework comparison.
• Meta, Google Cloud, and Anthropic — external customer context.
• CUDA, GPU, and Nvidia — incumbent ecosystem comparison.
• [[PyTorchFoundation|PyTorch Foundation]], [[VLLM|vLLM]], and Open Source AI Infrastructure — foundation governance context added by episode 148.