Closed Sudharsan-V closed 4 months ago
We have support for RM reshape, but it is fairly limited. What are your original and target reshape sizes?
Original shape: [1, 1, 60, 768] Target shape: [1, 60, 12, 64]
Same issue encountered with Mistral model bringup where most of the reshapes are done on CPU for various reasons:
Extended for a row-major layout, and also for values of HW that are not multiples of 32
Tested for the following shapes
- [4, 4, 32, 32] to [-1, 2, 32, 32],
- [4, 4, 32, 32] to [2, -1, 32, 32],
- [4, 4, 32, 32] to [4, 1, -1, 32],
- [4, 4, 32, 32] to [4, 2, 32, -1],
- [2, 3, 4, 4] to [2, 12, 1, 4],
- [1, 3, 6, 4] to [3, 1, 3, 8],
- [1, 1, 60, 768] to [1, 60, 12, 64],
- [2, 4, 128, 4] to [1, 2, 64, 32],
- [1, 11, 64, 2] to [1, 11, 1, 128],
- [1, 2, 64, 32] to [2, 4, 128, 4],
- [3, 2, 2, 32] to [1, 3, 1, -1]
Closing this issue as stale since many updates were made since then for reshape op. Also both Mistral and Mixtral bringup has been integrated on main without this issue.
The current state of the TT Metal reshape operation allows for the reshaping of tensors only when the tensor is in the TILE Layout.
The request is to extend this capability to enable the TT Metal reshape operation to function even when the tensor is in a Row major layout.
We have a use case where the tensor needs to be reshaped. But,
tt_lib.tensor.reshape
can function only when the tensor is in the TILE layout.