-
I don't think I'll get it done right now, so putting this here to track progress.
Currently, the shape of a tensor is controlled by `_shape`, a vector of dimensions of each rank, and `_mapper`, whi…
-
**REPRODUCER**
```mlir
func.func @vectorization_test(%extracted_slice : tensor, %arg0: index, %arg2: index, %3: tensor, %4: tensor) -> tensor{
%c0 = arith.constant 0 :index
%8 = linalg.generic {…
-
Using this issue to note down internal discussion about conv performance.
### Problem
1x1 filter convolutions get converted to `linalg.generic` ops during the global optimization pipeline. For exam…
-
`ttnn.permute` fails with inner-most dim = 1 when doing transpose (tested with 2D tensors, e.g. `(2, 1)` with permutation `(1, 0)`).
This is blocking the `aten.t.default` to `ttnn.permute` conversi…
-
DINO v2 finds that high values of stochastic depth is very helpful for larger models in terms of performance and they also gave an efficient implementation that only operates on the un-masked samples …
-
The shape of an input/output at::Tensor matches logical, even when the corresponding TensorView has an non-trivial allocation domain. See https://github.com/NVIDIA/Fuser/blob/2e4ed54ebd09946adef8d38b0…
-
### Describe the issue
when axis != rank -1, cuda softmax can tranpose input , then call softmax kernel, then transpose back;
```c++
auto temp_input = Tensor::Create(X->DataType(), TensorShape…
-
### 🚀 The feature, motivation and pitch
I'm working with COO sparse tensors and would like to get a permutation of any COO sparse tensor.
In the current version, `torch.permute` throws the following…
-
Hi all,
Thanks for making the course public.
I struggled to find any forums for this course online so I'm resorting to creating an issue here as a last ditch effort instead of getting in touch …
-
TL;DR:
The ONNX documentation for operator `Transpose` does not remove an ambiguity. The current example is
> For example, when perm=(1, 0, 2), given an input tensor of shape (1, 2, 3), the o…