Limitations in TT-Scatter Op

[Sudharsan-V]

The current version of the TT scatter operation falls short of providing the full functionality available in the PyTorch scatter operation.
Specifically, the scatter operation necessitates the index tensor (parameter) to be of the int64 data type, which is currently unsupported. Additionally, the constraint on accessing tensor elements through indexing represents a limitation in our use case.

As a result of these limitations, the full realization of the TT scatter operation's utility is hindered, thereby adversely impacting the results of the gs-demo.

Result when using TT-scatter op

Input prompt: A man is sitting on a roof
Output: A man is sitting on a roof.\nis a man a mantrafficacy. A. A+W, aka.1000.1 and #12000.1\n dévelop, #1, #10.10.10.'

[ajakovljevicTT]

Hi all, we've hit the problem of scatter not supporting full torch functionality relating to starting indexes when converting the StableHLO dialect and lowering to MLIR. Do we have perhaps have a roadmap for full support of the scatter op?