Masaki Kozuki

> given that we previously had issues with adam + cudagraphs, can you please describe here how you are dealing with the changing cpu state. It was straightforward except having...

I think I checked parity of states while implementing this but let me double check. and excuse me for delaying the fix for msvc build. (related to #81894)

From my side, this PR is ready for another round of review. I don't know what to do for the failing jobs as they don't look directly related to this...

linux-xenial-cuda11_3-py3_7-gcc7-deloy /build failed against for 1aff1ed: https://github.com/pytorch/pytorch/runs/7849107006?check_suite_focus=true#step:10:28829 ```console sccache: error: failed to execute compile sccache: caused by: error reading compile response from server sccache: caused by: Failed to read response...

Thank you for taking a look @malfet > One can update torch-deploy job to run against 11.6 (because why not), but still having files that needs more than 20 min...

By splitting files, the previously failing CUDA 11.3 job looks fine but another CUDA 11.3 of [linux-xenial-cuda11.3-py3.7-gcc7-bazel-test / build-and-test](https://github.com/pytorch/pytorch/runs/7985450807?check_suite_focus=true#logs) seems failing. [the raw log](https://pipelines.actions.githubusercontent.com/serviceHosts/7d146c05-69c3-4c20-a0e7-818111670117/_apis/pipelines/1/runs/2196050/signedlogcontent/18?urlExpires=2022-08-24T05%3A14%3A52.1486028Z&urlSigningMethod=HMACV1&urlSignature=hiUmfaptGgdNzoNDTxRxe4yMRiY3tZKDwfa5Yjf7IM4%3D) howeve doesn't seem informative

Please excuse my lack of explanation. I meant classification task with VGG19. That's why I ask adding caffemodel URL to `examples` and add (pickable) VGG19 to `links`.

This message is expected when either you install APEX without CUDA extensions or your CUDA version is

> Same here with torch LTS (torch-1.8.2+cu111) and CUDA 11.1 on a fresh installation and Nvidia Driver Version: 510.47.03 More precisely when importing `fused_weight_gradient_mlp_cuda` and without triggering the exception message:...

cuda 10.2 cannot compile `fused_weight_gradient_mlp_cuda`. so I expect (a) apex to be `import`-able and (b) the message to be displayed. Without any description of environment and steps to reproduce, there...