Summary:
Placement is leaked to too many classes!
In this diff, we consolidate all placement lookup into one place: Graph::ApplyDevicePlacement.
After applying placement, the in-memory graph, tensorMeta, weightMeta would already have the re-mapped device.
The subsequence weight loading, sample input loading, target device inference would look up the re-mapped device from graph's tensorMeta.
graph's tensorMeta becomes the only ground truth!
Test Plan:
Need to add some tests before landing.
This is a big change.
Rollback Plan:
Differential Revision: D78841818
Pull Request resolved: https://github.com/pytorch/pytorch/pull/158996
Approved by: https://github.com/henryoier
Summary:
In general, device_ is not very useful in OpKernel. Remove it to avoid misuse.
Also, the meaning of `device_` is also ambiguous in the OpKernel.
For StaticDispatch kernels, we always call cpu kernel.
For C10Kernel, we rely on input tensor's device and dispatcher to determine which device to run on.
For ops involves multiple device, e.g. aten._to_copy(device), the meaning of device is ill-defined.
Test Plan:
CI
Rollback Plan:
Reviewed By: henryoier, dolpm, kqfu, zhxchen17
Differential Revision: D78704840
Pull Request resolved: https://github.com/pytorch/pytorch/pull/158944
Approved by: https://github.com/dolpm
