Failed to lower some tensor manipulation ops from `aten` to `ttnn`

[jdh8]

The following ops refuse to lower to ttnn and stay in aten:

• concat tenstorrent/pytorch2.0_ttnn#188
• expand tenstorrent/pytorch2.0_ttnn#146
• repeat tenstorrent/pytorch2.0_ttnn#189
• reshape tenstorrent/pytorch2.0_ttnn#190

============================================================================== short test summary info ===============================================================================
FAILED tests/lowering/tensor_manipulation/test_concat.py::test_concat[input_shapes0-1] - AssertionError: assert 0 == 1
FAILED tests/lowering/tensor_manipulation/test_expand.py::test_expand[input_shape0-new_shape0] - AssertionError: assert 0 == 1
FAILED tests/lowering/tensor_manipulation/test_expand.py::test_expand_after_op[input_shape0-new_shape0] - AssertionError: assert 0 == 1
FAILED tests/lowering/tensor_manipulation/test_expand.py::test_expand_before_op[input_shape0-new_shape0] - AssertionError: assert 0 == 1
FAILED tests/lowering/tensor_manipulation/test_expand.py::test_expand_between_ops[input_shape0-new_shape0] - AssertionError: assert 0 == 1
FAILED tests/lowering/tensor_manipulation/test_repeat.py::test_repeat[input_shape0-sizes0] - AssertionError: assert 0 == 1
FAILED tests/lowering/tensor_manipulation/test_reshape.py::test_reshape[input_shape0-new_shape0] - AssertionError: assert 0 == 1
============================================================================ 7 failed, 9 passed in 8.42s =============================================================================

[ayerofieiev-tt]

@jdh8 need more details so we can pass the info to the Op team.

Can you please specify individual problems with each op? Ideally we should have single ticket for single issue, with issue categorized

Op Category:

• TM (tensor manipulation/data movement),
• MatMul,
• Conv,
• Eltwise,
• Pool,
• Reduction,
• Misc

Issue Category:

• Arguments order mismatch (lower prioritiy)
• Arguments mismatch (no way to pass configuration to the op)
• Unsupported case (op rejects the input)
• Crash
• Wrong answer
• Low PCC (close to expected value, but smth off)

Specific input: [params]. Link to an xfail test.

[ayerofieiev-tt]

@jdh8 please provide more details so we can fire individual tickets to Op lead

[boris-drazic]

@jdh8 For reshape your test (tests/lowering/tensor_manipulation/test_reshape.py) is not actually dispatching aten.reshape to compiler, but rather aten.view.

Your test module:

def forward(self, x, new_shape):
  return torch.reshape(x, new_shape)

is passed to compiler as

def forward(self, arg0_1):
    view = torch.ops.aten.view.default(arg0_1, [21, 5]);  arg0_1 = None
    return (view,)

which will yield the same result, but without using repeat.

[boris-drazic]

I don't see any code for processing aten OPs reshape, concat, and repeat in torch_ttnn/passes/lowering/to_tt_pass.py. That would explain why they are not being lowered to ttnn during compilation.

[jdh8]

@boris-drazic, I once removed conversion for these ops to keep the size of tenstorrent/pytorch2.0_ttnn#54 reviewable. Now I make separate PR for each op.

[ntarafdar]

@jdh8 for each sub issue can you give us a ttnn unit test thats failing. CC @ayerofieiev-tt

[jdh8]

@tarafdarTT, I've made each sub-issue as a PR, so we can experiment on tests without affecting other ops. In each PR, I 'unlocked' some tests by removing @pytest-mark-xfail, and these are the failing tests.

• concat tenstorrent/pytorch2.0_ttnn#188
• expand tenstorrent/pytorch2.0_ttnn#146
• reshape tenstorrent/pytorch2.0_ttnn#190

[ntarafdar]

@jdh8 oh okay, do you need anything from me (or the TM team ) right now or its that after you lower to ttnn ?