Michael Suo

You can add a config file to your project that flake8 should pick up automatically

btw: `atom-keyboard-macros` doesn't fully support `vim-mode-plus`. Seems like a bunch of motions/commands are not supported. For example, `df` will not be recorded in a macro.

One further note: this change is incompatible with Python 3.5 or lower, since it relies on the following behavior from [astimezone](https://docs.python.org/3/library/datetime.html#datetime.datetime.astimezone): > Changed in version 3.6: The [astimezone()](https://docs.python.org/3/library/datetime.html#datetime.datetime.astimezone) method can...

I think it's a really good start! I left some comments inline. The most important thing that needs to be reworked atm is the special status of Tensors as operands...

Some notes from porting this spec to thrift and using it to serialize an internal model: - `KeywordArgument` is unnecessary, `Node` can just take a `Map`. @SherlockNoMad this is another...

Citation for the claim that the equivalent python code does not do this: https://github.com/pytorch/pytorch/blob/master/torch/nn/functional.py#L2213

@pytorchbot merge

I think what's happening is that the linker is throwing out the static registration of the operators when linking in the torchvision lib, since the `RegisterOperator` object is never referenced...

> pre_autograd_aten_dialect = capture_pre_autograd_graph(m, (config, config.total_loc_num, device)) Slightly tangential: curious why you are using this API to capture the graph, as it is a private API (in `torch._export`).