Open esmalTT opened 3 days ago
TTNN as_tensor fails when creating multi-device tensors in tile layout that are in bfloat8 format. The following test demonstrates this:
as_tensor
@pytest.mark.parametrize("width", [16032 * 2, 16032]) @pytest.mark.parametrize("dtype", [ttnn.bfloat8_b, ttnn.bfloat16]) def test_multi_device_tensor(mesh_device, width, dtype, use_program_cache): a = torch.rand([1, 1, 32, width]) a = ttnn.as_tensor(a, mesh_mapper=ttnn.ShardTensorToMesh(mesh_device, -1), layout=ttnn.TILE_LAYOUT, dtype=dtype)
Running will show a failure on the following test case: test_multi_device_tensor[wormhole_b0-True-dtype0-16032] with the following error:
test_multi_device_tensor[wormhole_b0-True-dtype0-16032]
E RuntimeError: TT_THROW @ ../ttnn/cpp/ttnn/tensor/host_buffer/functions.hpp:148: tt::exception E info: E Tensor must have OwnedStorage
Close this issue when the failing test case passes with the correct output.
cc @cglagovich
Summary
TTNN
as_tensor
fails when creating multi-device tensors in tile layout that are in bfloat8 format. The following test demonstrates this:Running will show a failure on the following test case:
test_multi_device_tensor[wormhole_b0-True-dtype0-16032]
with the following error:Close this issue when the failing test case passes with the correct output.