← 返回事件
持续讨论AI发版

viable/strict/1788418752: [GPU][AOTI] Support host TMA in C++ wrappers (#193608) (#193608)

图:PyTorch Releases

发生了什么

Summary: Host-side TensorDescriptor values created inside a torch.library.triton_op are captured through a non-Dynamo tracing path. That path previously treated descriptors as opaque constants, so AOTInductor could not reliably preserve the source tensor view or reconstruct the lowered TMA launch ABI in C++ wrappers. This change decomposes supported host descriptors into their base tensors plus stable block-shape me…

摘要按规则整理自下方来源原文

为什么在扩散

来源