Skip to content

GitLab

Switch branch/tag
  1. 13 Aug, 2023 1 commit
  3. 10 Aug, 2023 1 commit
  5. 01 Aug, 2023 3 commits
  7. 29 Jul, 2023 1 commit
  9. 28 Jul, 2023 3 commits
  11. 27 Jul, 2023 1 commit
  13. 26 Jul, 2023 2 commits
  15. 23 Jul, 2023 6 commits
  17. 22 Jul, 2023 1 commit
  19. 21 Jul, 2023 1 commit
  21. 18 Jul, 2023 1 commit
  23. 17 Jul, 2023 2 commits
  25. 08 Jul, 2023 1 commit
    • Volodymyr Kyrylov's avatar
      rotary: update cos/sin cache when switching from inference mode · 70ab266a
      Volodymyr Kyrylov authored 
      This resolves RuntimeErrors after running evaluation in inference mode:

```
  File "/home/proger/.local/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
    return forward_call(*args, **kwargs)
  File "/home/proger/.local/lib/python3.10/site-packages/flash_attn/modules/mha.py", line 492, in forward
    qkv = self.rotary_emb(qkv)
  File "/home/proger/.local/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
    return forward_call(*args, **kwargs)
  File "/home/proger/.local/lib/python3.10/site-packages/flash_attn/layers/rotary.py", line 229, in forward
    return apply_rotary_emb_qkv_(
  File "/home/proger/.local/lib/python3.10/site-packages/torch/autograd/function.py", line 506, in apply
    return super().apply(*args, **kwargs)  # type: ignore[misc]
RuntimeError: Inference tensors cannot be saved for backward. To work around you can make a clone to get a normal tensor and use it in autograd.
```
      70ab266a
  27. 04 Jul, 2023 1 commit
  29. 03 Jul, 2023 2 commits
  31. 02 Jul, 2023 1 commit
  33. 02 Jun, 2023 1 commit
  35. 30 May, 2023 2 commits
  37. 27 May, 2023 1 commit
  39. 19 May, 2023 1 commit
  41. 06 May, 2023 2 commits
  43. 05 May, 2023 1 commit
  45. 21 Apr, 2023 4 commits