cutile-python
cutile-python copied to clipboard
NVIDIA
cuTile is a programming model for writing parallel kernels for NVIDIA GPUs
Does cuTile support a custom backward pass? It wasn't clear from the examples. This is obviously needed to training and something that triton supports. thanks!
### Version 1.0.0 ### Version 13.1 sm_120 ### Which installation method(s) does this occur on? Pip ### Describe the bug. I'm writing a softmax kernel using cutile: ```python @ct.kernel def...
### Is this a new feature, an improvement, or a change to existing functionality? New Feature ### How would you describe the priority of this feature request? Critical (currently preventing...
Hi, Since both **CuTe DSL** and **cuTile** are NVIDIA-maintained libraries with partially overlapping (but not identical) functionality, I’m wondering about their longer-term relationship. Is there any plan for **CuTe DSL**...
### Is this a new feature, an improvement, or a change to existing functionality? New Feature ### How would you describe the priority of this feature request? Critical (currently preventing...
### Version 1.0.0 ### Version 13.1 ### Which installation method(s) does this occur on? _No response_ ### Describe the bug. I almost forget the issue: when trying to use cuTile...
I'm trying to implement flash attention 4 in CuTile and got stuck on the polynomial exponent. Essentially flash attention 4 uses a polynomial approximation for exp2 in order to reduce...
## Description The block following TILE_X = max(1, 1024 // TILE_Y) seems unnecessary. Since integer division ensures (1024 // Y) * Y
## Description The link in the Pull Request template seems to be incorrect, update to the correct one. ## Checklist - [x] I am familiar with the [Contributing Guidelines](https://github.com/cutile-python/blob/HEAD/CONTRIBUTING.md). -...