Fix test-arm-backend-delegate

[guangy10]

#1336 broke test-arm-backend-delegate job.

[guangy10]

@pytorchbot label "partner: arm" "triaged"

[Jerry-Ge]

Hi @guangy10 the test-arm-backend-delegate seems is already working from this new job? Could you double check that error? Thanks

[guangy10]

Hi @guangy10 the test-arm-backend-delegate seems is already working from this new job? Could you double check that error? Thanks

Hi @Jerry-Ge , from what I can tell here (https://hud.pytorch.org/hud/pytorch/executorch/main, under "Linux" jobs), it's still broken.

[Jerry-Ge]

Hi @guangy10 the test-arm-backend-delegate seems is already working from this new job? Could you double check that error? Thanks

Hi @Jerry-Ge , from what I can tell here (https://hud.pytorch.org/hud/pytorch/executorch/main, under "Linux" jobs), it's still broken.

Hi @guangy10 , thanks for taking a look. Could that due to the 1701 PR hasn't been merged yet? #1701

Since it looks all right from the 1701 PR?

[guangy10]

It looks like the job works in #1701 but linter failed. From main I think the linter issue is already fixed. Let me import your PR to latest stable.

[Jerry-Ge]

It looks like the job works in #1701 but linter failed. From main I think the linter issue is already fixed. Let me import your PR to latest stable.

yes, thanks and the PR is also ready to be merged.

[Jerry-Ge]

Hi @guangy10, question:

How could the jobs in #1701 all work. But they still fail at: https://hud.pytorch.org/hud/pytorch/executorch/main ?

[Jerry-Ge]

For test-arm-reference-delegation, the error is: /bin/sh: 1: tosa_reference_model: not found
For test-arm-backend-delegation, the error seems is from executorch runtime? https://github.com/pytorch/executorch/actions/runs/7688283587/job/20949317262#step:11:521

[guangy10]

Hi @guangy10, question:

How could the jobs in #1701 all work. But they still fail at: https://hud.pytorch.org/hud/pytorch/executorch/main ?

Interesting. Before I imported #1701, I saw the job test-arm-backend-delegate passed on the PR and the only failure is the linter which has been fixed in main. Let me dig into it.

[Jerry-Ge]

Hi @guangy10, question:
How could the jobs in #1701 all work. But they still fail at: https://hud.pytorch.org/hud/pytorch/executorch/main ?

Interesting. Before I imported #1701, I saw the job test-arm-backend-delegate passed on the PR and the only failure is the linter which has been fixed in main. Let me dig into it.

And they work well on our internal CI. There should be something tiny missing from the CI setup here?

[guangy10]

Hi @guangy10, question:
How could the jobs in #1701 all work. But they still fail at: https://hud.pytorch.org/hud/pytorch/executorch/main ?

Interesting. Before I imported #1701, I saw the job test-arm-backend-delegate passed on the PR and the only failure is the linter which has been fixed in main. Let me dig into it.

And they work well on our internal CI. There should be something tiny missing from the CI setup here?

@Jerry-Ge I think it's because #1693 introduced new types numeric_limits<> in half.h. So when you build your runtime, you got the errors like:

In file included from /pytorch/executorch/backends/arm/../../../executorch/runtime/core/portable_type/complex.h:11,
                 from /pytorch/executorch/backends/arm/../../../executorch/runtime/core/exec_aten/exec_aten.h:35,
                 from /pytorch/executorch/backends/arm/../../../executorch/runtime/core/evalue.h:10,
                 from /pytorch/executorch/backends/arm/../../../executorch/runtime/core/event_tracer.h:9,
                 from /pytorch/executorch/backends/arm/../../../executorch/runtime/backend/backend_execution_context.h:11,
                 from /pytorch/executorch/backends/arm/../../../executorch/runtime/backend/interface.h:13,
                 from /pytorch/executorch/backends/arm/runtime/ArmBackendEthosU.cpp:15:
/pytorch/executorch/backends/arm/../../../executorch/runtime/core/portable_type/half.h:521:7: error: 'numeric_limits' is not a class template
  521 | class numeric_limits<torch::executor::Half> {
      |       ^~~~~~~~~~~~~~
/pytorch/executorch/backends/arm/../../../executorch/runtime/core/portable_type/half.h:[521](https://github.com/pytorch/executorch/actions/runs/7707006124/job/21003481476#step:11:522):45: error: explicit specialization of non-template 'std::numeric_limits'
  521 | class numeric_limits<torch::executor::Half> {
      |                                             ^
/pytorch/executorch/backends/arm/../../../executorch/runtime/core/portable_type/half.h:530:38: error: 'std::numeric_limits' is not a template
  530 |   static constexpr auto has_denorm = numeric_limits<float>::has_denorm;
      |                                      ^~~~~~~~~~~~~~

See the full job here: https://github.com/pytorch/executorch/actions/runs/7707006124/job/21003481476

@manuelcandales Could you take a look?