[Mlir-commits] [mlir] Extending UniformQuantizedType with interface-based support for new storage types in Quant dialect (PR #152966)

[Anurag Singh]

anuragsingh-tt wrote:

> we'd no longer choose to embed the scale information as part of the tensor/datatype information

+1. Hardwares will extract these separately. 

> Define a non-uniform quant type say quant.nonuniform which is a tuple of quant.uniform i.e (quant.uniform, quant.uniform) whose 1st element signify the value and the 2nd element is scale.

What are the issues with instead treating scales/zps as SSA operands on the consuming ops in the case of dynamic quantization? 

https://github.com/llvm/llvm-project/pull/152966