Open
Description
The following test fails:
def test_5d_concat(device):
torch_input_tensors = [torch.rand(1, 32, 56, 56, 1, dtype=torch.bfloat16) for _ in range(3)]
torch_result = torch.cat(torch_input_tensors, dim=-1)
ttnn_input_tensors = [ttnn.from_torch(x, layout=ttnn.TILE_LAYOUT, device=device) for x in torch_input_tensors]
ttnn_result = ttnn.concat(ttnn_input_tensors, dim=-1) # throws here
ttnn_result = ttnn.to_torch(ttnn_result)
assert_with_pcc(torch_result, ttnn_result, 0.9999)
With error: Tensor rank is greater than 4
Performing the test with a ROW_MAJOR_LAYOUT inputs instead, and then attempting to convert the result to TILE _LAYOUT fails on the subsequent ttnn.to_layout call. Test:
def test_5d_concat(device):
torch_input_tensors = [torch.rand(1, 32, 56, 56, 1, dtype=torch.bfloat16) for _ in range(3)]
torch_result = torch.cat(torch_input_tensors, dim=-1)
ttnn_input_tensors = [ttnn.from_torch(x, layout=ttnn.ROW_MAJOR_LAYOUT, device=device) for x in torch_input_tensors]
ttnn_result = ttnn.concat(ttnn_input_tensors, dim=-1)
ttnn_result = ttnn.to_layout(ttnn_result, ttnn.TILE_LAYOUT) # throws here
ttnn_result = ttnn.to_torch(ttnn_result)
assert_with_pcc(torch_result, ttnn_result, 0.9999)
With error:
E RuntimeError: TT_FATAL @ /localdev/lpanos/tt-metal/ttnn/cpp/ttnn/operations/data_movement/reshape_view/reshape.cpp:383: false
E info:
E Attempting to reshape between two shapes with different volumes
If we are to perform eltwise operations after the concatenation, its result must be in TILE_LAYOUT.
Activity