Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/21233
It is possible that OnnxifiOp is created in a thread where weights have been cleaned from the workspace, which is legit use case as we can create the backend once and lower all the weights. So we need to extract the weight shape info the first time we create the backend and save it.
Reviewed By: bertmaher, rdzhabarov
Differential Revision: D15587237
fbshipit-source-id: 1f264dc32c0398c42b618e9c41c119eb13e1c9f1
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17375
Previously we create the onnxGraph first and take it to the onnx manager for registration. It doesn't work well in practice. This diff takes "bring your own constructor" approach to reduce the resource spent doing backend compilation.
Reviewed By: kimishpatel, rdzhabarov
Differential Revision: D14173793
fbshipit-source-id: cbc4fe99fc522f017466b2fce88ffc67ae6757cf
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17198
We come to the point that we need to apply some rules to bind certain ops together to avoid un-inferrable intermediate shapes. We either lower them together to backend or neither. This diff adds a pass for us to add rules like this. The first one is to bind `Gather` with `SparseLengthsWeighted*`.
Reviewed By: ipiszy
Differential Revision: D14118326
fbshipit-source-id: 14bc62e1feddae02a3dd8eae93b8f553d52ac951
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16047
Implements single thead safe map enabling sharing of generated graph between
different ops.
Added model_id to every onnxified op to help create a unique id in the map.
Some formatting fix.
Reviewed By: yinghai
Differential Revision: D13663927
fbshipit-source-id: 27417e8fe752fdd48abb6a87966cd76d592e1206
