Llama 3.2 1B - Add specific op tests #622
Conversation
mstojkovicTT
requested review from
nvukobratTT,
pilkicTT and
dgolubovicTT
as code owners
| def test_add(shapes): | ||
|
|
||
| if shapes[0] != shapes[1]: | ||
| pytest.xfail("eltwise_add broadcast not supported") |
There was a problem hiding this comment.
For all xfail tests, can you create issues on MLIR within this milestone, and attach generated TTIRs as well? For easier tracking, let's also create reminder issues to remove those xfails once MLIR issues are fixed. Ofc, we can use the "Blocker" filed to track which issues from FFE is waiting some other issues on MLIR side :))
Let's create separate issues for each op type (e.g. add, cosine, etc).
Note: This is not a blocker for merging this PR :))
| ) | ||
| def test_concat(inputs_and_dim): | ||
| in_shape1, in_shape2, dim = inputs_and_dim | ||
| print(in_shape1, in_shape2, dim) |
There was a problem hiding this comment.
Let's remove these prints, seems redundant as we can get shape info from test logs as well :))
There was a problem hiding this comment.
Oops, i left those from testing, removed now
|
|
||
|
|
||
| @pytest.mark.parametrize("shapes", [(1, 11, 64)]) | ||
| @pytest.mark.xfail(reason="Cosine not supported in TTIR") |
There was a problem hiding this comment.
As cos and sin are now supported, can we do a rebase on top of the latest main and check if they work?
| ((1, 11), 128256, 2048), | ||
| ], | ||
| ) | ||
| @pytest.mark.xfail(reason="ttnn.embedding op fails while reshaping the input_tensor in TILE_LAYOUT") |
There was a problem hiding this comment.
Is this still the case? The latest MLIR uplift should have solved this one. Let's double-check before proceeding.
There was a problem hiding this comment.
As far as i know, it still has the same issue. Maybe i am missing something, but the test still fails
|
|
||
| if source_shape != target_shape or source_and_target_shape == ((1, 32, 11, 64), (1, 32, 11, 64)): | ||
| pytest.xfail( | ||
| "Unable to reshape a tensor in TILE_LAYOUT to non-tile height and width! Please convert the tensor to ROW_MAJOR_LAYOUT first." |
There was a problem hiding this comment.
Let's check on latest main if this one is fixed as well
There was a problem hiding this comment.
Same as embedding it seems
| source_shape, target_shape = source_and_target_shape | ||
|
|
||
| if len(source_shape) != 4 or len(target_shape) != 4: | ||
| pytest.xfail("Reshape for dim != 4: Unhandled attribute type") |
There was a problem hiding this comment.
Is this issue on FFE side?
| ((1, 8, 11, 64), 2), | ||
| ], | ||
| ) | ||
| @pytest.mark.xfail(reason="TTNN: Tensor layout issues with non tile dim aligned shapes") |
There was a problem hiding this comment.
Let's also check latest main, should be solved
In preparation for the rollout of Llama 3.2 1B, I have implemented a series of tests to assess our current support infrastructure's readiness. These tests are designed to ensure we meet the operational requirements necessary for a successful model bring-up.
fixes #533