TorchServe seems to share code with MXNET. I think this is more AWS/Facebook responding to Google's power in ML with TensorFlow/GCP. AWS doesn't have a framework, and Facebook doesn't have a cloud. This gives PyTorch better feature parity with TensorFlow (TorchServe == TF Serving) and gives AWS a point of differentiation from GCP.

> AWS doesn't have a framework

They do - it's MXNet. It's just that no one really cares about it.

Yeah, but as you say, MXNet doesn't have the market of PyTorch or TF, and while AWS has deep ties within MXNet, they don't outright control it to the same extent that Google owns TF or Facebook has PyTorch.

It's an Apache project sure, but AWS employs the vast majority of contributors. I'd say in practice it controls MXNet like G/FB control TF/PT. Although MXNet is a much more community friendly project than TF.