[Ops] Support for reduce average op (ttnn.mean)

[nvukobratTT]

No description provided.

[nvukobratTT]

One group of ops fail because of the required squeeze (test setup and FFE op processing), while other group fails with tt-nn specific issues.

[nvukobratTT]

@chandrasekaranpradeep this one is unblocked. Can you confirm if it works? :))

[chandrasekaranpradeep]

@nvukobratTT There are two test cases in the forge/test/mlir/test_ops.py

1. test_reduce_mean
   In this test cases, TTNN: Issues with ttnn implicit reshape is resolved
   The below test cases alone is failing with pcc issue (PCC: 0.7203957195745748)
   pytest forge/test/mlir/test_ops.py::test_reduce_mean[-1-input_shape2] -vss
2. test_mean
   In this test cases, the FFE: Unsupported squeeze operation issue is resolved, now all the test cases here are failing with Unable to reshape a tensor in TILE_LAYOUT to non-tile height and width! Please convert the tensor to ROW_MAJOR_LAYOUT first. issue

Logs:
test_mean.log
test_reduce_mean.log

[chandrasekaranpradeep]

@nvukobratTT Created a PR for reduce average op test cases - #349

[chandrasekaranpradeep]

Created a issue for the reduce average (i.e ttnn.mean) data mismatch
TT-MLIR - tenstorrent/tt-mlir#869
TT-Metal - tenstorrent/tt-metal#13621

Data Mismatch Failing Test cases:

Reduce_avg:
    Input_shape = (1, 12, 3200)
    dim = -1
    keep_dim = True
pytest forge/test/mlir/test_ops.py::test_reduce_mean[-1-input_shape2] -vss

[chandrasekaranpradeep]

Created a issue for the reduce average (i.e ttnn.mean) data mismatch
TT-MLIR - tenstorrent/tt-mlir#869
TT-Metal - tenstorrent/tt-metal#13621

@nvukobratTT Reduce average op configuration (i.e Input_shape = (1, 12, 3200), dim = -1, keep_dim = True) is present in the Llama 3b model. Can we change the priority of the issue from P1 to P0?

[chandrasekaranpradeep]

In tt-metal latest main, Unable to reshape a tensor in TILE_LAYOUT to non-tile height and width! Please convert the tensor to ROW_MAJOR_LAYOUT first. issue is resolved for test_mean pytest.
Once the tt-mlir is uplifted with latest tt-metal main, need to verify the issue is resolved in forge or not and update the test_mean

[nvukobratTT]

Created a issue for the reduce average (i.e ttnn.mean) data mismatch
TT-MLIR - tenstorrent/tt-mlir#869
TT-Metal - tenstorrent/tt-metal#13621

@nvukobratTT Reduce average op configuration (i.e Input_shape = (1, 12, 3200), dim = -1, keep_dim = True) is present in the Llama 3b model. Can we change the priority of the issue from P1 to P0?

Yup, I'll convert it right away :))

[nvukobratTT]

In tt-metal latest main, Unable to reshape a tensor in TILE_LAYOUT to non-tile height and width! Please convert the tensor to ROW_MAJOR_LAYOUT first. issue is resolved for test_mean pytest. Once the tt-mlir is uplifted with latest tt-metal main, need to verify the issue is resolved in forge or not and update the test_mean

Yup! I think uplift of this one is already WIP :)) We expect it to land soon.

Please keep track of it and test out this op accordingly :))

[chandrasekaranpradeep]

Yup! I think uplift of this one is already WIP :)) We expect it to land soon.
Please keep track of it and test out this op accordingly :))

Sure Nikola