Dmytro Dzhulgakov

Caffe2 nets usually have stubs for the inputs in a form of ConstantFill in the init_net. It causes caffe2_net_to_onnx_model to complain as it doesn't know how to handle this op....

Enable Kineto in OSS builds by fixing build condition (resubmit)

1

Resubmit of https://github.com/pytorch/pytorch/pull/89174 . I think I fixed underlying issues back then, but only CI would tell. Context: This PR enables Kineto on OSS builds because of how the flags...

Make models immutable for inference and pass state explicitly

1

Summary: Usual prediction setup assumes that there's a single instance of nn.Module (scripted) and it gets invoked from multiple threads concurrently to serve prediction requests. It means that the execution...

Simple docker build for self-contained dynolog daemon image. Useful for containerized environments. Improvement suggestions are welcome. Also update gitignore.

torch.nn.LayerNorm support for arbitrary axis in order to allow NCHW application

24

### 🚀 The feature, motivation and pitch LayerNorm starts to be applied to image data on per-channel basis (e.g. in ConvNeXt model). `torch.nn.LayerNorm` support normalization only on the last several...