Add an (opset) version to Python objects

[Jegp]

Currently, there's no way to know which version a NIR graph is being read as. Every graph should contain an opset code so implementors can verify exactly what version they're working on.

We are currently writing version numbers into all the NIR files. That uses the NIR version number instead of an opset code. I reckon there are three ways of handling this:

1. Move the versioning into the Python object (which will be serialized anyway)
2. Use version numbers in the files but opset codes in the nodes
3. Use both version numbers and opset codes in files and objects

I would argue for setting 1 to avoid redundancy. We'll have to represent versioning in Python, so we might as well centralize it. And by making it a property of the objects, we can retain the read_version function in the I/O code and proxy that to the object inside the graph.

[bvogginger]

Thanks for raising this topic @Jegp.

I had a look at how ONNX does the operator versioning and I found the following page helpful: https://onnx.ai/onnx/repo-docs/Versioning.html#operator-sets

I think it is a great idea to add operator set versions to NIR.

First, all operators should get the property opset = 1.

Then, when an operator is modified, it gets an increased opset number, following the approach in ONNX.

For the NIRNodes, the question is whether the opset version is an actual variable of the Node, or just part of the documentation. In ONNX this is part of the schema.

As intermediary step, I suggest to store the opset version along with the NIR version into the NIR file.

Note that ONNX offers a "VersionConverter", which allows to translate graphs from one operator version to the other one. For each operator one register converters...

For now, I suggest to just start with adding the opset version to keep track of it. We might do this even prior to merging #129.

Later, we can think about converters etc..

[Xiangyu-Li97]

Hi @Jegp @bvogginger,

I'd like to work on this issue.

Based on the discussion above, I'm planning to implement the following:

1. Add opset_version as a class variable to NIRNode base class (default = 1)
2. Include opset_version in the serialized output (to_dict)
3. Store opset versions in the HDF5 file alongside the NIR version
4. Add read_opset_versions() function to read opset info from files
5. Add get_opset_versions() method to NIRGraph to get all node type versions

This follows the approach suggested by @bvogginger - starting simple with version tracking, leaving VersionConverter for future work.

I have a working implementation with tests ready. Would you like me to submit a PR?

[Jegp]

Wonderful! I think the plan sounds solid. It would be nice to also add tests for the writing and reading to make sure that the opcodes are stored and parsed correctly.

Please go ahead with a PR if you have one. That makes it easier to discuss code specific problems.