Dynamic shape support in TOSA backend?

[chadhakunal]

Does the TOSA backend in Torch-MLIR support dynamic shapes (e.g., runtime batch sizes, sequence lengths, or other symbolic dimensions)?

I’m exploring lowering models with variable batch sizes and sequence lengths and want to confirm whether this is currently supported when targeting TOSA.

Thanks!

[justin-ngo-arm]

Hi @chadhakunal, there will be initial support for dynamic shape in TOSA version 1.1.

[chadhakunal]

Thanks for the info, @justin-ngo-arm !
Is the dynamic shape work already in progress? If not, could I contribute to it?
I’m working on a compiler that depends on dynamic shape support, and I’d love to help out if possible.

[justin-ngo-arm]

For now, there is not much that you can contribute to yet. Please keep an eye on https://www.mlplatform.org/tosa/ for any new update regarding dynamic shape support.