Convert aten.cumsum to ttnn.moreh_cumsum

[jerrysky3]

Ticket

#367, #372

Problem description

Add conversion from aten.cumsum to ttnn.moreh_cumsun

Currently moreh_cumsum only supports n and c dim of (n, c h, w) 4D tensor. This conversion covers the existing model cases by unsqueeze (h, w) to (h, w, 1, 1)

What's changed

• Add conversion from aten.cumsum to ttnn.moreh_cumsum
• Add tests for conversion
• Fix incorrect conversion in add_data_move_pass.py (see details in add_data_move_pass.py might choose incorrect tensor layout for ttnn.reshape #372)

[jerrysky3]

Same as the comment above regarding the check

[jerrysky3]

During testing, I found that these check might be incorrect for some cases. For example: ttnn.reshape can reshape (1, 32) into (1, 32, 1, 1). But we currently don't maintain the output shape of of converted ttnn ops, can_be_tilized will still check based on the stalled shape (1, 32) and returns incorrect decision.

Here is an example if we print the dst_node.meta["val"] of ttnn.reshape from (1, 32) to (1, 32, 1, 1)

arg0_1: {'val': FakeTensor(..., size=(1, 32), dtype=torch.bfloat16), 'tensor_meta': TensorMetadata(shape=torch.Size([1, 32]), dtype=torch.bfloat16, requires_grad=False, stride=(32, 1), memory_format=torch.contiguous_format, is_quantized=False, qparams={})}

We can see the dst_node.meta["val"].size() will return (1, 32) instead of the expected output shape (1, 32, 1, 1)

I filed #372 to track this issue and found that #322 seems to do lots of improvements for these checks. So I'd propose remove these checks for now (be more conservative) and let #322 make overall improvements

[kevinwuTT]

Yeah, there has been some new changes in TTNN that makes some of these Layout workarounds obsolete. However, they're still not perfect and I find that some layout changes are still needed for some models. @jerrysky3 Have you checked if this PR passes all the models? If it does then we can merge this first, and I'll continue to work my PR #322 and merge afterwards.

[jerrysky3]

After rebase all model tests are passed https://github.com/tenstorrent/pytorch2.0_ttnn/actions/runs/11772057825