Ean Garvey

It seems to be an issue with a transpose on the norm input -- the abort also goes away when I increase --iree-opt-level to O2 or higher. Perhaps we are...

@AWoloszyn if you have cycles for it, your context on HIP runtime may help here the most. PTAL

Output with AMD_LOG_LEVEL=3: [Azure](https://sharkpublic.blob.core.windows.net/sharkpublic/ean/wan/semabrt_log_level_3.txt)

Thank you @AWoloszyn, reducing the sequence length (in this model that means restricting height and width of output) to below 65535 does circumvent the error, though it is obviously not...

Exports of the original post's MLIR can be reproduced by following the instructions on the shark-ai feature branch @ wan_exports: https://github.com/nod-ai/shark-ai/blob/wan_exports/sharktank/sharktank/torch_exports/wan/README.md This is the nn.module we are exporting through iree-turbine...

The next step here, if I understand correctly, is to improve code generation for the attention shape with head dim 384 used in this VAE decode. I can help with...

@Groverkss I am reimplementing this model in sharktank to get a better export, but I suspect we will still need to support these cases. Has there been any change since...

@gpetters-amd FWIW the existing api file in apps/shark_studio/api/sd.py can be converted to apps/shark_studio/api/turbine-models-sd.py (that's sort of what you have going at the moment, I think) and refactored to match an...

By all means, I'd take a look at any PRs submitted for this. It is, however, not an entry-level task and I don't recommend it for a first-time contributor. @gmalasan

It may help to enumerate the cases where as_strided is used in pytorch lowerings and decompositions. Chunk (example in OP) is just one of many sources and is likely one...