[ROCm] Enable gemm fusion autotuner.

[zoranjovanovic-ns]

No description provided.

[hsharsha]

@zoranjovanovic-ns Is there a PR against 0.4.35 to fix common test failures between 0.4.31 and 0.4.35?

[zoranjovanovic-ns]

@hsharsha Do you have log file or list of failing test cases?

[i-chaochen]

I still can see there are 7 failed unit tests @hsharsha

//xla/service/gpu:dot_dimension_sorter_test_gpu_amd_any
//xla/service/gpu:triton_fusion_numerics_verifier_test_gpu_amd_any
//xla/service/gpu/fusions/triton:triton_fusion_emitter_parametrized_test_gpu_amd_any
//xla/service/gpu/fusions/triton:triton_support_legacy_test_gpu_amd_any
//xla/service/gpu/fusions/triton:triton_support_test
//xla/service/gpu/tests:gpu_triton_custom_call_test_gpu_amd_any
//xla/tests:dot_operation_test_gpu_amd_any

Executed 276 out of 276 tests: 269 tests pass and 7 fail locally

wondering is it expectable?

[zoranjovanovic-ns]

Will go through the list and investigate.

[hsharsha]

@zoranjovanovic-ns These tests are failing on 31-qa https://github.com/ROCm/frameworks-internal/issues/10326#issuecomment-2614607763 Please rebase on top of https://github.com/ROCm/xla/tree/r0.4.31-qa-test-fix

[zoranjovanovic-ns]

@hsharsha Should I create PR for https://github.com/ROCm/xla/tree/r0.4.31-qa-test-fix then?

[hsharsha]

@hsharsha Should I create PR for https://github.com/ROCm/xla/tree/r0.4.31-qa-test-fix then?

I need to merge it to 0.4.31-qa. i cleaned up few cuda specific tests in there. With that these tests are failing with triton autotuner

//xla/service/gpu:triton_fusion_numerics_verifier_test_gpu_amd_any
//xla/service/gpu/fusions/triton:triton_fusion_emitter_parametrized_test_gpu_amd_any
//xla/service/gpu/fusions/triton:triton_support_legacy_test_gpu_amd_any
//xla/service/gpu/fusions/triton:triton_support_test

//xla/tools/multihost_hlo_runner:functional_hlo_runner_test_gpu_amd_any