Resnet 18 iree path about torch-mlir HOT 5 CLOSED

Can you try using the latest IREE? The first error is a known error that was fixed recently. Not sure about the second.

from torch-mlir.

I just ran into the hal.constant_pool.load issue myself. It was pointing at an opaque constant dumped by the reproducer. Do you have large_elements_limit on your get_asm call? If you do, then removing it should fix the issue.

I will try to make IREE give a better diagnostic for this.

from torch-mlir.

Hello Sean, Thanks for the answer!

I upgraded to the latest iree-runtime-snapshot and iree-compiler-snapshot from https://github.com/google/iree/releases/tag/snapshot-20210816.466 and the first error is gone. Could you point me to the patch which fixed this problem so that I can merge it into my custom iree project?

As for the second error, I also changed to the lastest iree snapshot and I dropped the large_elements_limit option in my get_asm call:

but the error still exists with a slightly different error message:

from torch-mlir.

The first error was fixed by iree-org/iree#6629

For debugging the second, can you file an IREE bug with appropriate starting IR and iree-translate command to reproduce? I don't think this is npcomp-specific. (feel free to add me on that bug, I work actively with IREE as well)

btw, I'll be on vacation for the next week, but I'll respond right as I get back

from torch-mlir.

Sure, Thanks!

from torch-mlir.