-
### Checklist
- [X] 1. I have searched related issues but cannot get the expected help.
- [X] 2. The bug has not been fixed in the latest version.
- [X] 3. Please note that if the bug-related iss…
-
2. Ratio of `Triton`/ `XeTLA` keep same except for attention caused by `XeTLA` attention absolute number degraded
3. Both `Triton` and `XeTLA` softmax cases degraded, so `Triton`/ `XeTLA` not changed…
-
Would it be possible to add functionality for **Grad-CAM** or **attention map** similar to those used in DINO?
Thank you!
-
"We do not have so many requests, actually.
We also have some internal discussions, but there are a lot of alternatives for the faster (lightweight) encoder and Squeezeformer does not come to a hig…
-
When I run to the file “referring-expression-counting-main/GroundingDINO/groundingdino/models/GroundingDINO/ms_deform_attn.py”, I get the error
![image](https://github.com/user-attachments/assets/c82…
-
### 🐛 Describe the bug
I struggled a bit to get a repro, but I think this is in the realm of reasonable and identifies the behavior that causes my runs to diverge.
```python
import torch
impor…
-
### 🐛 Describe the bug
I am running a FlexAttention operation and it returns different output shapes with and without compile. The correct output shapes are those returned without compile.
```P…
-
From reading this thread:
https://github.com/pytorch/pytorch/issues/96099#issuecomment-1480430583
It seems to me that the relative positional embedding can be integrated with `scaled_dot_product_att…
-
### Checklist
- [x] 1. I have searched related issues but cannot get the expected help.
- [ ] 2. The bug has not been fixed in the latest version.
- [ ] 3. Please note that if the bug-related issue y…
-
Having some generation issues with NMT models trained with OpenNMT-py, which include OpenNMT-py versions before flash attention existed and one I'm currently training with the most recent which includ…