mirror of
https://github.com/pytorch/pytorch.git
synced 2025-10-21 05:34:18 +08:00
Context is in https://www.internalfb.com/excalidraw/EX519691 and https://docs.google.com/document/d/1qnuXLZk_GYt_PksHTwkn7L2ELRDnYlIRPkHAlXTyuhw/edit?tab=t.0. And the description of the previous PR: https://github.com/pytorch/pytorch/pull/164340. The previous PR adds the support on the HOP side for eager execution and AOTAutograd. Dynamo is still passing the HOP a subgraph with wrong shapes. This PR fixes that. This is similar to the HOP implementation, however we additionally need to manually keep the TensorVariable metadata in sync. Pull Request resolved: https://github.com/pytorch/pytorch/pull/163602 Approved by: https://github.com/ydwu4 ghstack dependencies: #164296, #164321, #164419, #164420, #164340