Open chandrasekaranpradeep opened 3 weeks ago
@chandrasekaranpradeep do we have a matching issue on TTNN/Metal repo as well?
@nvukobratTT I have not created a issue on TT-Metal repo. Do I need to create a issue for this on TT-metal repo?
@nvukobratTT I have not created a issue on TT-Metal repo. Do I need to create a issue for this on TT-metal repo?
As we already have TTNN repro, let's also create issue on TT-Metal as well :))
Sure I will create a issue on TT-Metal
@nvukobratTT Created a issue in TT-Metal repo - https://github.com/tenstorrent/tt-metal/issues/13667. The concat op with these configuration (i.e input_shape, dim) is present in the Llama 3B rotary embedding. Can we change the priority for the issue?
Summary: In Llama 3B rotary embedding, the ttnn.concat fails when the concat dimension of the input tensors are not TILE aligned.
For more context, here is the exact error message:
Details: The tuple of two input tensor of same shape
(1, 32, 12, 50)
are passed tottnn.concat
op withconcat dim = -1,
while validating the input tensor of the concat op in thettnn::operations::data_movement::ConcatDeviceOperation::validate
function in TTNN, the Tile padding along concatenated dim (3) not supported for concat yet (tensor: 0) error is thrown.The ttnn.concat op expects the concatenated dim of the input tensors should be tile aligned because tile padding along the concat dim of the input tensors are not supported in ttnn.concat op.
Repro:
Concat TTIR:
TTNN test cases:
TT-Forge-fe Concat test: