Inconsistency in ttnn.reshape for single vs multi-device tensors

[avoraTT]

Description

It seems that ttnn.reshape is inconsistent when being performed on single device vs multi-device. Specifically, the inconsistency can be seen when doing a [W, X, Y*Z] -> [W, X, Y, Z] reshape on a tensor that is on a single device, vs. a tensor that is replicated across devices.

For single device tensor, the outputs match torch, but for a multi-device tensor, they do not.

Repro steps

• Machine: any T3K
• Branch: avora/reshape_bug
• Command: pytest models/demos/t3000/llama2_70b/tests/test_rope_reshape.py

[avoraTT]

cc @cglagovichTT

[cglagovichTT]

FYI The models team has been having a few correctness issues with multidevice reshape/transpose. They are tough to repro, but now Ammar has found that unit tests may not be catching them because they exist on multidevice.

[ntarafdar]

Thanks @avoraTT and @cglagovichTT will look at this early next week

[uaydonat]

any updates here?

[ntarafdar]

The team is looking at fixing single device reshape this week still, that has gone overtime afterwards we can look at this.