Could OpenXLA support input dynamic shape?

[jinjidejinmuyan]

I want to know that Could OpenXLA support input dynamic shape?
For example, I have an input tensor with a dynamic dimension. TensorRT’s approach is to specify the minimum, optimal, and maximum values for this dynamic dimension of the input. Does OpenXLA have a similar feature?
Thank you for your reply.

[aniruthraj]

Hi @jinjidejinmuyan,

Yes, OpenXLA does support input dynamic shape. Dynamic dimensions in OpenXLA are represented by ?(e.g., tensor<16x?xf32>). A key feature is bounded dynamism, whose value has a known upper bound(similar to TensorRT's max value). We can represent bounded dynamism using we represent bounded dynamism using #stablehlo.bounds. OpenXLA also supports unbounded dynamism, which refers to a dynamic dimension with no known bound. However, 'min' and 'optimal' values are not explicitly handled by OpenXLA. Please refer to this documentation for more information on dynamism in OpenXLA.

Thank you!