[torchlib] Inefficient implementations

[justinchuby]

Existing inefficient implementations that need to be improved:

• Embedding bag: [torchlib] Implement aten.embedding_bag.padding_idx #2219, [torchlib] torch.ops.aten.embedding_bag.padding_idx may see padding_idx=None #2145, Embedding_bag mismatch #1056
• as_strided
• Dynamo-based ONNX exporter adds seemingly unnecessary operations to the graph pytorch/pytorch#162061

[crypto-a]

Hi @justinchuby, is this issue still available? I’d love to work on it if possible

[justinchuby]

Yes, it is. Thank you

[crypto-a]

Assign it to me, and I will start working on it.

[crypto-a]

Here is how I am gonna resolve the outstanding issues with embedding_bag:

1. Fix Numerical Mismatch: I will correct the implementation logic in nn.py to ensure the ONNX Script output exactly matches the results from PyTorch's native function, fixing the Tensor-likes are not close! errors.

2. Handle None Padding Index: I will modify aten_embedding_bag_padding_idx to properly handle cases where padding_idx=None, which will resolve the DLRM model export crash.

3. Ensure CI Stability: I will validate my fix by running the full test suite to ensure no new bugs or side effects are introduced. I'll also fix any linting errors and add specific test coverage for the None case.

[crypto-a]

@justinchuby would you prefer One PR per issue or all in one?

[justinchuby]

Please create separate PRs, thanks.

[crypto-a]

@justinchuby I've added the pull request with the full set of changes needed to support padding_idx=None in aten_embedding_bag