Rob Suderman

You need to execute it on an x86 machine, then just `iree-compiler --iree-hal-target-backends=llvm-cpu ` should yield the following error: ``` suderman@build[]: iree-compile /tmp//test.mlir --iree-hal-target-backends=llvm-cpu lld: error: undefined symbol: fminf >>>...

You are right, apparently just with `--iree-llvmcpu-link-embedded=false` it builds correctly. Happy to close but I would be curious why the flag changes the success.

We should change this at the reduce op variants level instead. https://github.com/llvm/torch-mlir/blob/0c29ccf1439c91c7a2175a167d4bdb2c01a03e63/lib/Dialect/Torch/Transforms/ReduceOpVariants.cpp#L240 Instead of matching to `torch.aten._scaled_dot_product_flash_attention_for_cpu` to `aten._scaled_dot_product_flash_attention_for_cpu`. Or preferably match with both. I am guessing that the ingestion...

Please add a test to the torch test suite

> calar_dense op. Do you know what torch op would generate this when we export it? When I ran into the op it was generated by Having trouble generating the...

Fix is uploaded to: https://github.com/rsuderman/torch-mlir/tree/torch_scf We will need to check if `while` or `if` have similar issues. The tests only roundtripped tensors, they never included any computation in the body....

I can see benefit to this optimization however this is more avoiding the compilation issue we have been encountering rather than preventing the crash.