Han-Chung Wang

> This is a perfectly valid program and should be supported correctly in IREE (I think it does today). The semantics is that the initial value of the `matmul_transpose_b` is...

Let's convert it to draft until we understand what's happening and missing in ukernel path? Note that we'll need to revisit ukernel approach that does not use MLIR ukernel path....

I don't know what the actual setup is, but passing input arguments to `outs` is usually problematic because they are readonly. We can fix it like https://github.com/iree-org/iree/pull/21651 but the PR...

You can try using `tensor.empty()` in `outs`, as they will result in buffer allocations on host side, and they should be either writeonly or readwrite tensors. The fft support was...

Sorry that I missed the issue. I'm confused now because we don't have vectorization support for FFT. I suspect that it is because of the copy op that is generated...

Can you make the PR description more descriptive when it is ready for review? Here are some examples: https://google.github.io/eng-practices/review/developer/cl-descriptions.html

Okay, I'm able to compile and run the model. Thanks for the update from @monorimet . [This](https://gist.github.com/hanhanW/3797206187b96304baa99cfb0226524e) is the script that I used to compare the results. I don't see...

We have SDXL VAE integration tests. Closing the issue, feel free to reopen if there are action items. https://github.com/iree-org/iree/blob/main/tests/external/iree-test-suites/sharktank_models/quality_tests/sdxl/vae_cpu.json

I think the first step is making it go down to `TileAndFuse` pipeline. Can you share the IR dump for the `vector.scatter` op issue? It looks like some operations are...

Yes, I'd expect that the map_scatter op gets fused into the forall op. It is wrong if it does not happen. The IR dump looks okay in the first distribution,...