Unsupported by backend lowering: `torch.operator` op

[PeiMu]

I’m new to the torch-mlir, and I’m trying it with a simple op, but met such error:

error: unsupported by backend lowering: torch.operator op
note: see current operation: %41 = “torch.operator”(%38, %40) {name = “aten.maximum”} : (!torch.tensor<[?,?],f32>, !torch.tensor<[?,?],f32>) → !torch.tensor
note: this is likely due to a missing op that needs to be generated by torch_ods_gen.py
Traceback (most recent call last):
File “/torch-mlir/examples/iou_of.py”, line 119, in
compiled = compile_module(mlp_module)
File “/torch-mlir/examples/iou_of.py”, line 99, in compile_module
pm.run(mb.module)
RuntimeError: Failure while executing pass pipeline.

As far as I’m concerned, such an error means there’s something wrong in the pass manager of “torch-backend-to-linalg-on-tensors-backend-pipeline”. But I have no idea how to fix that or is there an error with my code?

Thanks!

---

I proposed such question in the community too:
https://llvm.discourse.group/t/unsupported-by-backend-lowering-torch-operator-op/4590

[cathyzhyi]

@ds1231h Thanks for your interest in the project! The error msg is saying this is likely due to a missing op that needs to be generated by torch_ods_gen.py so this op is not supported yet. You could refer to https://github.com/llvm/torch-mlir/wiki/Torch-ops-E2E-implementation If you are interested in adding the support.

[PeiMu]

@ds1231h Thanks for your interest in the project! The error msg is saying this is likely due to a missing op that needs to be generated by torch_ods_gen.py so this op is not supported yet. You could refer to https://github.com/llvm/torch-mlir/wiki/Torch-ops-E2E-implementation If you are interested in adding the support.

Thanks for the reply! Good tutorial!

[silvasean]

I added a few ops in #383

The remaining ones are

    %2 = torch.aten.slice.Tensor %0, %int0, %none, %none, %int1 : !torch.tensor<[1024,4],f32>, !torch.int, !torch.none, !torch.none, !torch.int -> !torch.tensor<[?,4],f32>
    %3 = torch.aten.select.int %2, %int1, %int2 : !torch.tensor<[?,4],f32>, !torch.int, !torch.int -> !torch.tensor<[?,1],f32>

which @dan-garvey is already working on as part of #342

:) Your code should be working soon!

[PeiMu]

I added a few ops in #383

The remaining ones are

    %2 = torch.aten.slice.Tensor %0, %int0, %none, %none, %int1 : !torch.tensor<[1024,4],f32>, !torch.int, !torch.none, !torch.none, !torch.int -> !torch.tensor<[?,4],f32>
    %3 = torch.aten.select.int %2, %int1, %int2 : !torch.tensor<[?,4],f32>, !torch.int, !torch.int -> !torch.tensor<[?,1],f32>

which @dan-garvey is already working on as part of #342

:) Your code should be working soon!

Thanks a lot!

[PeiMu]

@silvasean Thanks for your MR with #383 !

I've tried the test case but got an error with "IouOfModule":

python -m e2e_testing.torchscript.main --filter IouOfModule --verbose

Exception: 
Lower Torch Backend IR -> Linalg-on-Tensors Backend IR failed with the following diagnostics:
error: failed to legalize operation 'torch.copy.to_tensor'
note: see current operation: %13 = "torch.copy.to_tensor"(%0) : (!torch.vtensor<[?,?],f32>) -> !torch.tensor<[?,?],f32>

My code is already up to date.

Is there anything wrong with my command line?

From the error message, it seems like missing the "torch.copy.to_tensor" in Linalg, is that right? So is there's any tutorial for this condition?

Thanks!

[PeiMu]

@silvasean Thanks for your MR with #383 !

I've tried the test case but got an error with "IouOfModule":

python -m e2e_testing.torchscript.main --filter IouOfModule --verbose

Exception: 
Lower Torch Backend IR -> Linalg-on-Tensors Backend IR failed with the following diagnostics:
error: failed to legalize operation 'torch.copy.to_tensor'
note: see current operation: %13 = "torch.copy.to_tensor"(%0) : (!torch.vtensor<[?,?],f32>) -> !torch.tensor<[?,?],f32>

My code is already up to date.

Is there anything wrong with my command line?

From the error message, it seems like missing the "torch.copy.to_tensor" in Linalg, is that right? So is there's any tutorial for this condition?

Thanks!

I'm not sure if it needs the third step in this tutorial - https://github.com/llvm/torch-mlir/wiki/Torch-ops-E2E-implementation? Thanks @cathyzhyi

[silvasean]

@dan-garvey is working on the missing ops already -- you just have to wait for him to land his changes :)

torch.copy.to_tensor ops don't need to be lowered to linalg -- they usually are a symptom of other ops that need to be lowered (in this case torch.aten.slice.Tensor and torch.aten.select.int)

[silvasean]

@ds1231h - while you are waiting, do you have another module that you can try? It is useful for us to know what we are missing! Please file bugs!

[PeiMu]

@dan-garvey is working on the missing ops already -- you just have to wait for him to land his changes :)

torch.copy.to_tensor ops don't need to be lowered to linalg -- they usually are a symptom of other ops that need to be lowered (in this case torch.aten.slice.Tensor and torch.aten.select.int)

Ok...I understand that, and I will try to add these two ops.

@ds1231h - while you are waiting, do you have another module that you can try? It is useful for us to know what we are missing! Please file bugs!

Yes, there are about seven modules in all, for each has some error message. I would like to fix them by myself first. :) I will also record the relevant process in the community, and push code in github (I wish I could)

Thanks a lot for your help!