Ian Wood

Looks like only default inputs are supported, i'm not sure if there is a plan to support this https://github.com/llvm/torch-mlir/blob/ca0e9066755b35c0889c6ab792265b0886325f50/lib/Conversion/TorchToTMTensor/TorchToTMTensor.cpp#L1573-L1575

It appears that unsigned types (for example, `ui8`) are no longer legal for this op see https://github.com/llvm/llvm-project/pull/118075. I'm not sure why there is no verifier to catch this, however.

@lundong where is this IR coming from? The linked IR contains the `ui8` `tosa.rescale` ops.

The IR is old so I'm now getting the verification error: ``` :0: error: 'tosa.const' op requires attribute 'values' :0: note: see current operation: %0 = "tosa.const"() {value = dense_resource...

https://github.com/iree-org/iree/pull/18074

> > It most definitely should be a codegen issue.... If we can narrow it down to the diverging dispatch that would help. > > I can spend some time...

The numerical differences originate from a `linalg.conv_2d_nhwc_hwcf ` op followed by a `linalg.generic` ### Min Repro Instructions ### 1. Checkout and build https://github.com/iree-org/iree/tree/shared/sdxl_quantized in IREE 2. Clone https://github.com/nod-ai/sdxl-scripts and `cd...

Duplicate of https://github.com/iree-org/iree/issues/18200#issuecomment-2286472770? Both might be a problem with `DequantizeLinear`

Looked at it a bit, it is a duplicate of the problem mentioned https://github.com/iree-org/iree/issues/18005#issuecomment-2278885607, but that isn't the initial problem identified in the issue. Maybe worth keeping this one?

~Looks like torch-mlir doesn't support onnx's `start` or `end` arguments for slicing the shape.~ Edit: I was looking at an outdated version