Justin Chu

@liwuhen Please feel free to create a pull request!

Yes - but the ability to name values with any string as been useful for e.g. matching initializers from PyTorch / safetensors to ONNX models. Practically this was never enforced....

The way ExportedProgram in PyTorch handles it by having a mapping: ``` inputs_to_parameters={ 'arg0_1': 'branch1.0.weight', 'arg1_1': 'branch1.0.bias', 'arg2_1': 'branch2.0.weight', 'arg3_1': 'branch2.0.bias', }, ``` This way the names in the graph...

The issue with ONNX is that the initializers need to take the same name as values in the graph, so I can't think of a way to implement something that's...

Related https://github.com/onnx/onnx/issues/5348

I wonder if we should relax this and only reserve a few characters like quotation marks

This type of stateful operations (updating best_num_matches) can be a bit trickly.

https://github.com/microsoft/onefuzz potentially related

I just realized this too. Also for E5M2 +/- inf, should it be +/- max/min as well? (I mean the specification should be corrected)

~@yuanyao-nv Does saturate not apply to float4?~ float4 doesn't have nan